JP4751230B2 - Prosodic segment dictionary creation method, speech synthesizer, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method and a program for generating dictionary of prosodic element sequences, a speech synthesizer, and a speech synthesizing program and a speech synthesizing method, wherein a phrase, such as a fixed form phrase which is repeatedly used with high possibility, and another phrase are suitably concatenated with natural intonation. <P>SOLUTION: The speech synthesizer 100 includes a standard prosody dictionary 11 consisting of pitch frequency information by speech units and a prosodic elementary unit dictionary 12 consisting of pitch frequency information of a predetermined phrase unit and prosodic feature information showing prosodic features of predetermined prosodic units of source speech data before and after the phrase in an original sound data. A phrase part, matching the predetermined phrase, in a prosody information series generated from the standard prosody dictionary 11 is replaced with the pitch frequency information of the predetermined phrase unit in the prosodic elementary unit dictionary 12, and based upon the prosodic feature information, standard pitch frequencies corresponding to predetermined prosodic units before and after it are adjusted. <P>COPYRIGHT: (C)2008,JPO&amp;INPIT

Description

  The present invention relates to a prosodic segment dictionary creation method, a prosodic segment dictionary creation program, a speech synthesizer, and a speech synthesis program used in speech synthesis for generating synthesized speech waveform data from sentence text generated by analyzing a text sentence And a speech synthesis method.

Conventionally, as a technique for synthesizing speech from sentence text generated from text data such as sentences, there are speech synthesis apparatuses described in Patent Document 1 and Patent Document 2.
The speech synthesizer of Patent Document 1 registers a phrase (text) to be uttered naturally in a language dictionary, and extracts a basic frequency pattern (hereinafter referred to as F0 pattern) extracted when the entire registered text is uttered naturally. Are stored in the prosodic dictionary separately from the F0 pattern for normal text. When the input text is registered in the language dictionary, the F0 pattern is selected by referring to the F0 pattern group for the registered text, and when the input text is not registered in the language dictionary, The prosodic information is generated by selecting the F0 pattern with reference to the F0 pattern group for normal text. As a result, for the registered phrase, synthesized speech close to natural utterance can be output.

Further, the speech synthesizer of Patent Document 2 can change only a part of a sentence by rule synthesis, and the other part is synthesized at the time of rule synthesis when synthesizing using synthesis parameters or speech waveform data created by analysis. Analyzes with the rule synthesis unit by extracting and using only the part of the content that is used as the rule synthesis part, taking into account the surrounding sentence environment such as the sentence of the fixed part near the sentence Improve the connectivity of the prosody of the division. This makes it possible to synthesize with good naturalness.
JP 2004-198917 A JP-A-9-188515

  However, in the speech synthesizer described in Patent Document 1, since the part where the natural utterance F0 pattern is reflected is only the matched phrase part in the input text, the naturalness of the matched part can be reproduced. There is a possibility that the prosody of the continuous portion of the F0 pattern for normal text and the F0 pattern for natural speech and the prosody of the whole sentence are not well-balanced and naturalness cannot be reproduced.

  Further, in the speech synthesizer described in Patent Document 2, one phrase is extracted from synthesized speech waveform data including two phrases generated by rule synthesis, and this is synthesized with a phrase of analysis sound. When the speech segment at the connection side end of each of these two phrases is, for example, when the phoneme environment on the analysis sound side and the phoneme environment on the extracted rule synthesis sound side do not satisfy a predetermined condition, 2 There is a risk that the connectivity after the connection of two phrases is lost and the naturalness is impaired.

  Therefore, the present invention has been made paying attention to such an unsolved problem of the conventional technology, and a phrase that is highly likely to be repeatedly used such as a fixed phrase and the connection of other phrases. To provide a prosody segment dictionary creation method and prosody segment dictionary creation program, a speech synthesizer, a speech synthesis program, and a speech synthesis method, which are suitable for performing in a state with more natural inflection. It is said.

In order to achieve the above object, the prosody segment dictionary creating method according to claim 1 according to the present invention comprises:
A prosodic segment dictionary creation method for creating a prosodic dictionary used for speech synthesis,
A first prosodic information extraction step of extracting first prosodic information that is prosodic information of a speech waveform data portion of a predetermined phrase included in the speech waveform data from speech waveform data corresponding to an utterance sentence uttered by a predetermined speaker; ,
A second prosodic information extracting step for extracting second prosodic information as prosody information from at least one of the speech waveform data parts corresponding to a predetermined prosodic unit preceding and following the speech waveform data part of the predetermined phrase;
A third prosodic information generation step of generating third prosodic information that is prosodic information corresponding to the predetermined prosodic unit with reference to a standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit;
Based on the second prosodic information extracted in the second prosodic information extracting step and the third prosodic information generated in the third prosodic information generating step, characteristic features of the second prosodic information and the third prosodic information are characterized. Prosodic feature information generation step for generating prosodic feature information indicating the difference,
A prosodic segment dictionary creating step of creating a prosodic segment dictionary based on the first prosodic information for each predetermined phrase extracted in the first prosodic information extracting step and the prosodic feature information generated in the prosodic feature information generating step; It is characterized by including.

  That is, for example, if the prosodic information is pitch frequency information, the pitch frequency information of a predetermined phrase in the speech waveform data corresponding to the utterance sentence uttered by the predetermined speaker and the predetermined phrase in the speech waveform data The prosodic segment dictionary is created based on the prosodic feature information indicating the characteristic difference from the third prosodic information of the pitch frequency information corresponding to at least one of the preceding and succeeding predetermined prosodic units. .

  Therefore, for example, for a prosodic information sequence generated only with standard prosodic information (pitch frequency information in units of speech), the prosodic information sequence portion of the predetermined phrase is replaced with the prosody segment dictionary created above for the same phrase. It is possible to replace the pitch frequency information having, and for the prosodic information sequence portion of a predetermined prosodic unit generated from the standard prosodic information preceding and following the prosodic information sequence portion of the replaced phrase, Based on the prosodic feature information corresponding to the replaced phrase, it is possible to make an appropriate adjustment in accordance with the prosodic information sequence portion of the replaced phrase. , The sentence including the specified phrase and the phrases before and after it are uttered with prosody (inflection, speech speed, volume, etc.) that is closer to natural speech. Effect that it is possible to generate (playback) by the synthesized speech waveform data is obtained. Here, the prosodic unit is a unit for generating a prosodic pattern, and uses various units such as a phoneme, a mora, an accent phrase, a unit obtained by dividing one accent phrase into a plurality of sections, a plurality of accent phrases, and an exhalation paragraph. be able to.

  Note that the range of the sentence portion uttered (reproduced) with prosody (inflection, speech speed, volume, etc.) closer to natural speech varies depending on the range of prosodic feature information and the range of adjustment. For example, if one exhalation paragraph preceding or following the phrase of the first prosodic information is within the above range, the prosody of the preceding or succeeding exhalation paragraph and the phrase of the first prosodic information that is continuous therewith can be made more natural speech. It is possible to make it a close prosody. In addition, the prosodic feature information is not limited to the prosodic unit that is continuous with the first prosodic information phrase, but the prosodic feature information of the prosodic unit having a great prosodic influence on the first prosodic information phrase, and the standard prosodic corresponding to the prosodic unit is generated. The prosodic information sequence portion generated from the information is adjusted.

  Here, the first prosodic information, the second prosodic information, and the third prosodic information are composed of information related to prosody such as pitch frequency, volume (power), and phoneme duration (speech rate). The pitch frequency is the reciprocal of the period having a periodicity in the voice, and represents the pitch of the voice. The volume indicates the strength of utterance in units of speech. For example, in the case of an isolated word utterance, generally the beginning of the word is uttered strongly, and the volume decreases toward the ending. The phonological continuation length indicates, for example, the length (time) of the syllable utterance. For example, the ending syllable is uttered relatively longer than the word. Short words tend to be uttered later (that is, each syllable is longer) and longer words are uttered earlier (each syllable is shorter). The same applies to the prosodic segment dictionary creation program according to claim 5, the speech synthesizer according to claims 6 to 9, the speech synthesis program according to claims 11 to 14, and the speech synthesis method according to claims 15 to 18. It is.

  The first prosodic information is information that is composed of prosodic information for each predetermined phrase constituting the speech waveform data and is not modeled by an HMM (Hidden Markov Model) or the like. The standard prosodic information includes prosodic information for each speech unit data constituting speech waveform data, and includes information modeled by an HMM or the like. The same applies to the prosodic segment dictionary creation program according to claim 5, the speech synthesizer according to claims 6 to 9, the speech synthesis program according to claims 11 to 14, and the speech synthesis method according to claims 15 to 18. It is.

  The standard prosodic dictionary is composed of standard prosodic information that is pitch frequency information for each voice unit described above, and generates a prosodic information sequence for every sentence by combining standard prosodic information of a plurality of voice units. It is possible. Here, the standard prosody information includes, for example, a phoneme label and an average value and a variance value of one or more representative points in a speech unit having a pitch frequency corresponding to the phoneme label. The same applies to the prosodic segment dictionary creation program according to claim 5, the speech synthesizer according to claims 6 to 9, the speech synthesis program according to claims 11 to 14, and the speech synthesis method according to claims 15 to 18. It is.

  The second prosodic information is composed of prosodic information for each predetermined prosodic unit constituting the speech waveform data, and is composed of prosodic information for each single or plural speech unit data. The same applies to the prosodic segment dictionary creation program according to claim 5, the speech synthesizer according to claims 6 to 9, the speech synthesis program according to claims 11 to 14, and the speech synthesis method according to claims 15 to 18. It is.

  Also, the voice unit can be configured in a predetermined unit using, for example, phonemes or mora. In other words, phonemes and mora themselves are used as speech units, or phonemes and mora are used to compose the words “I” into speech units, or “I am” and “I am” compose words. It can be a unit. Here, a phoneme is a unit of sound used in one language, and is the smallest unit that makes a difference in meaning. When a certain sound is discriminating from other sounds in the language, one phoneme is used. It is recognized as a phoneme. A mora is a phoneme combination having a length equal to the sum of one consonant phoneme and one short vowel phoneme. Further, in the case of Japanese, the mora corresponds to a kana character unit and is slightly different from the syllable. A unit of sound that is counted as 5, 7, 5, 7, 7, etc. in haiku or waka, and a sound that stretches out “-” (long vowel), a clogging sound “tsu” (promotion sound), a bouncing sound “n” (repellent sound) ) Etc. are also considered 1 mora. The same applies to the prosodic segment dictionary creation program according to claim 5, the speech synthesizer according to claims 6 to 9, the speech synthesis program according to claims 11 to 14, and the speech synthesis method according to claims 15 to 18. It is.

  The predetermined prosodic unit includes an accent phrase, an exhalation paragraph, and the like. For example, when a predetermined phrase is included in the exhalation paragraph, the preceding part and the subsequent part in the exhalation paragraph can be a predetermined prosodic unit. In addition, an exhalation paragraph preceding a predetermined phrase, an accent phrase preceding a predetermined phrase, and the like are also a predetermined prosodic unit. The exhalation paragraph is, for example, a sentence before and after being punctuated by a Japanese punctuation mark. For example, in the case of “Tomorrow's weather will be sunny”, “Tomorrow's weather” and “It will be sunny” are the exhalation paragraphs. The accent phrase will be described later. The same applies to the prosodic segment dictionary creation program according to claim 5, the speech synthesizer according to claims 6 to 9, the speech synthesis program according to claims 11 to 14, and the speech synthesis method according to claims 15 to 18. It is.

A prosodic segment dictionary creating method according to claim 2 of the present invention is the prosodic segment dictionary creating method according to claim 1,
The predetermined phrase includes a phrase having a relatively high appearance frequency.
For example, "it is a left turn to," which is frequently used in car navigation systems, phrases and such as "Turn right the ~", "Good morning", "Hello", "you will find a ~", "Please - The first and second prosodic information is extracted and the prosodic feature information is generated for a phrase that is used relatively frequently (such as a phrase with a relatively high frequency of appearance) such as a phrase that is regularly used. A prosodic segment dictionary is generated based on the first prosodic information and the prosodic feature information.

  Therefore, by adjusting the prosodic information sequence generated only from the standard prosodic dictionary using such a prosodic segment dictionary, the prosodic information sequence portion of a phrase that is likely to be used in a sentence is obtained from a predetermined speaker. The prosody information series part that reflects the prosodic utterance can be replaced, and the prosody information series part corresponding to the predetermined prosody unit before and after it can be adjusted to an appropriate content, so the sentence containing the phrase to be replaced and the phrases before and after it Can produce synthesized speech waveform data that is uttered (reproduced) with prosody (inflection, speech speed, volume, etc.) closer to natural speech.

Further, the prosodic segment dictionary creating method according to claim 3 according to the present invention is the prosodic segment dictionary creating method according to claim 1 or 2,
The second prosodic information and the third prosodic information include pitch frequency information,
The prosodic feature information includes information indicating a ratio between a pitch frequency indicated by the second prosodic information and a pitch frequency indicated by the third prosodic information.
That is, information indicating the ratio between the pitch frequency (second prosodic information) of a predetermined prosody unit extracted from the speech waveform data of the uttered sentence that is naturally uttered and the standard pitch frequency corresponding thereto, for example, both pitch frequencies When the ratio itself has a plurality of predetermined prosodic units, information such as the ratio of the average values of both pitch frequencies is used as the prosodic feature information.

  Therefore, by using such a prosodic segment dictionary, the prosodic information sequence part of the predetermined phrase corresponding to the prosodic segment dictionary in the prosodic information generated only from the standard prosodic dictionary is replaced, and the predetermined prosodic units before and after that When adjusting the prosodic information sequence part, it is possible to adjust to the appropriate content by adjusting the standard pitch frequency in the corresponding prosodic information series part to a frequency according to the ratio indicated by the prosodic feature information As a result, it is possible to easily perform adjustment in accordance with the prosody of the prosodic information sequence portion after replacement.

Furthermore, the invention according to claim 4 is the prosodic segment dictionary creating method according to any one of claims 1 to 3,
The second prosodic information and the third prosodic information include pitch frequency information,
The prosodic feature information includes information indicating a ratio between the intonation size obtained from the pitch frequency difference indicated by the second prosodic information and the intonation size obtained from the pitch frequency difference indicated by the third prosodic information. It is characterized by that.

  That is, the magnitude of intonation obtained from the difference in pitch frequency (second prosodic information) of a predetermined prosodic unit extracted from the speech waveform data of a naturally uttered sentence, and the corresponding standard pitch frequency (third prosodic information) ) Information indicating the ratio of the intonation size obtained from the height difference, for example, the difference between the maximum value and the minimum value of the pitch frequency is set as the intonation size, and the information of the ratio itself is used as the prosodic feature information.

  Therefore, using such a prosodic segment dictionary, the prosodic information sequence portion of the predetermined phrase corresponding to the prosodic segment dictionary in the prosodic information sequence generated only from the standard prosodic dictionary is replaced, and the predetermined prosody before and after that When adjusting the prosodic information sequence portion of a unit, for example, by adjusting the pitch frequency that determines the magnitude of intonation in the corresponding prosodic information sequence portion to a frequency according to the ratio indicated by the prosodic feature information, the content becomes appropriate. Therefore, it is possible to easily perform adjustment according to the prosody of the replaced prosodic information sequence portion.

In order to achieve the above object, a prosodic segment dictionary creation program according to claim 5 according to the present invention is provided.
A program for creating a prosodic segment dictionary used for speech synthesis,
A first prosodic information extraction step of extracting first prosodic information that is prosodic information of a speech waveform data portion of a predetermined phrase included in the speech waveform data from speech waveform data corresponding to an utterance sentence uttered by a predetermined speaker; ,
A second prosodic information extracting step for extracting second prosodic information as prosody information from at least one of the speech waveform data parts corresponding to a predetermined prosodic unit preceding and following the speech waveform data part of the predetermined phrase;
A third prosodic information generation step of generating third prosodic information that is prosodic information corresponding to the predetermined prosodic unit with reference to a standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit;
A characteristic of the second prosodic information and the third prosodic information based on the second prosodic information extracted in the second prosodic information extracting step and the third prosodic information generated in the third prosodic information generating step Prosodic feature information generating step for generating prosodic feature information indicating a difference,
A prosodic segment dictionary creating step for creating a prosodic segment dictionary based on the first prosodic information for each predetermined phrase extracted in the first prosodic information extracting step and the prosodic feature information generated in the prosodic feature information generating step; It is characterized by including a program for causing a computer to execute a process consisting of:
With this configuration, when the program is read by the computer and the computer executes processing according to the read program, the same effect as the prosodic segment dictionary creating method according to claim 1 can be obtained.

In order to achieve the above object, a speech synthesizer according to claim 6 according to the present invention comprises:
A speech synthesizer that generates synthesized speech waveform data corresponding to sentence text,
The prosodic segment dictionary creation method according to any one of claims 1 to 4 or the prosodic segment dictionary creation program according to claim 5 includes the first prosodic information and the prosodic feature information. A prosodic segment dictionary consisting of
Standard prosodic dictionary composed of standard prosodic information that is prosodic information for each voice unit;
A segment dictionary comprising spectral information for each speech unit and excitation source information for each speech unit;
Text analysis means for performing accent analysis and morphological analysis on the sentence text;
Prosodic information sequence generation means for generating a prosodic information sequence corresponding to the sentence text based on the analysis result of the text analysis means and the standard prosodic information of the standard prosodic dictionary;
When a phrase corresponding to the first prosodic information included in the prosodic segment dictionary is included in a phrase corresponding to the sentence text, the prosodic information series portion of the phrase is converted to the first prosodic information. A changing means for changing to the prosodic information sequence part generated based on
The prosodic segment corresponding to the changed prosodic information sequence portion with respect to the prosodic information sequence portion corresponding to at least one of the predetermined prosodic units preceding and following the phrase portion changed by the changing means Prosody information adjusting means for performing predetermined adjustment processing based on the prosodic feature information of the dictionary ;
Voice waveform generation means for generating synthesized voice waveform data corresponding to the text text based on the prosodic information series and the spectrum information and the excitation source information of the unit dictionary. .

With such a configuration, it is possible to perform accent analysis and morpheme analysis on the sentence text by the text analysis unit, and the analysis result of the text analysis unit and the standard by the prosodic information series generation unit It is possible to generate a prosodic information sequence corresponding to the sentence text based on the standard prosodic information possessed by the prosodic dictionary, and the changing means includes a phrase of the prosodic segment dictionary in the phrase corresponding to the sentence text. When a phrase corresponding to the first prosodic information is included, the prosodic information sequence portion of the phrase can be changed to a prosodic information sequence portion generated based on the first prosodic information, At least one of the predetermined prosodic units preceding and following the phrase part changed by the changing means by the prosodic information adjusting means Against prosodic information sequence portion corresponding to the prosodic units, it is possible to perform a predetermined adjustment processing based on the prosodic feature information included in the circuit prosody segment dictionary corresponding to prosodic information sequence portion of the modified parts The speech waveform generation means can generate synthesized speech waveform data corresponding to the sentence text based on the prosodic information series and the spectrum information and the excitation source information of the segment dictionary.
Therefore, by making appropriate adjustments based on the prosodic feature information according to the prosodic information sequence part after the change, the sentence including the predetermined phrase and the phrases before and after the prosodic sentence is more prosthetic (inflection, speech). Synthetic speech waveform data that is uttered (reproduced) at a high speed, volume, etc.) can be generated.
Here, based on the accent analysis result and the morphological analysis result, the text analysis means determines the reading, accent, and intonation of the input kanji mixed sentence (sentence text), for example, in the case of Japanese. In addition, based on the analysis result, a pronunciation / prosodic symbol string (intermediate language) that is reading information with prosodic symbols generated for the sentence text may be generated. Note that the intermediate language is a language that simply describes the control of the reading necessary for the synthesized output when converting the text text into speech waveform data. Hereinafter, the same applies to the speech synthesizer according to claims 7 to 8, the speech synthesis program according to claims 10 to 12, and the speech synthesis method according to claims 13 to 15.
The prosodic symbols are composed of exhalation paragraph breaks, accent phrase breaks, accent nucleus positions, and the like. For example, accents in Japanese are determined for each word, and are used to specify a vocabulary as well as stress in English. For example, homonyms such as “bridge” and “chopsticks” are distinguished by accents. The voice discrimination between the two is made by the position where the point where the pitch of the voice sharply drops (this is called the accent nucleus) is located. An accent phrase is a phrase having at most one accent kernel. The accent type is determined by the position of the accent kernel. For example, if “Hashi” is used, the word “shi” has an accent kernel, which is called type 2 or tail type. In the case of “chopsticks”, “ha” is a word having an accent core, which is called a type 1 or head height type, and those having no accent core are called a type 0 or a flat plate type. Also, when uttered as a phrase or sentence, each word is affected by the accent position when combined with an attached word or when forming a compound word with another word. It has been found that these can be classified into four types: subordinate, incompletely dominated, fused, and dominated. In addition, kanji is generally read in multiple ways, so-called sound reading and kanji reading. In addition, there are things such as rendaku and affix that change the way of reading depending on the combination of words. For example, as rendaku, there are Mezamashi + and Kei → Mezamashi Kei, and as an affix there is 1 + book → Japan. Hereinafter, the same applies to the speech synthesizer according to claims 7 to 8, the speech synthesis program according to claims 10 to 12, and the speech synthesis method according to claims 13 to 15.
The speech segment is composed of a series of phonemes, semiphones, multiple phonemes (for example, consonant-vowel, vowel-consonant-vowel, consonant-vowel-consonant). Hereinafter, the same applies to the speech synthesizer according to claims 7 to 8, the speech synthesis program according to claims 10 to 12, and the speech synthesis method according to claims 13 to 15.
The spectral information is information expressing the acoustic resonance characteristics of the vocal tract, and the excitation source information is information expressing the vocal cord vibration during voiced and the turbulent air flow from the lungs during unvoiced. For example, the excitation source information is expressed by a periodic pulse or the like when voiced, and is expressed by white noise or the like when silent. Hereinafter, the same applies to the speech synthesizer according to claims 7 to 8, the speech synthesis program according to claims 10 to 12, and the speech synthesis method according to claims 13 to 15.
Further, the spectrum information and the excitation source information may be one corresponding to each speech unit, or may be plural. Hereinafter, the same applies to the speech synthesizer according to claims 7 to 8, the speech synthesis program according to claims 10 to 12, and the speech synthesis method according to claims 13 to 15.

In order to achieve the above object, a speech synthesizer according to claim 7 according to the present invention comprises:
A speech synthesizer that generates synthesized speech waveform data corresponding to sentence text,
A first segment dictionary comprising spectral information for each speech unit and excitation source information for each speech unit;
Extracted from speech waveform data corresponding to the utterance sentence uttered by the predetermined speaker, the spectrum information for each predetermined phrase, the excitation source information for each predetermined phrase, and the predetermined preceding and following each predetermined phrase in the speech waveform data A second segment dictionary comprising information on the phonetic environment of
Standard prosodic dictionary composed of standard prosodic information that is prosodic information for each voice unit;
Text analysis means for performing accent analysis and morphological analysis on the sentence text;
Prosodic information sequence generation means for generating a prosodic information sequence corresponding to the sentence text based on the analysis result of the text analysis means and the standard prosodic information of the standard prosodic dictionary;
When the phrase corresponding to the sentence text includes a phrase that matches the phrase corresponding to the spectrum information and excitation source information of the second segment dictionary, the spectrum information and the excitation source information are A predetermined phoneme environment of the second segment dictionary preceding and succeeding the matching phrase in the speech waveform data at the time of extraction, and a predetermined preceding and succeeding a portion corresponding to the matching phrase in the prosodic information sequence Determining means for determining whether or not the phonological environment of
Based on the prosodic information series, the determination result of the determination means, and the spectrum information and the excitation source information of the first and second segment dictionaries, synthetic speech waveform data corresponding to the sentence text is generated. Voice waveform generation means,
The speech waveform generation means, when the determination means determines that they do not match, the synthesized speech waveform data generated from the spectrum information and excitation source information corresponding to the matching phrase of the second segment dictionary Based on the first synthesized speech waveform data excluding the speech segment data part at the side edge portion where the phoneme environments do not match, and the spectrum information and excitation source information of the first segment dictionary for the entire sentence text. Generating synthesized speech waveform data corresponding to the sentence text, which is synthesized from second synthesized speech waveform data obtained by excluding a portion corresponding to the first synthesized speech waveform data from the generated synthesized speech waveform data; It is a feature.

With such a configuration, it is possible to perform accent analysis and morpheme analysis on the sentence text by the text analysis unit, and the analysis result of the text analysis unit and the standard by the prosodic information series generation unit A prosodic information sequence corresponding to the sentence text can be generated based on standard prosodic information included in the prosodic dictionary, and the second unit dictionary is included in a phrase corresponding to the sentence text by a determination unit. spectral information and when it contains a match with the corresponding phrase in the excitation source information, preceding and succeeding the phrase that the matching of the voice data at the time of extracting the spectral information and the excitation source information possessed by and a predetermined phonetic environment having a second segment dictionary, pairs phrases that the match in the prosodic information sequence It is possible to determine whether or not a predetermined phoneme environment that precedes and follows the portion to be matched, and the prosody information sequence, the determination result of the determination unit, and the first by the speech waveform generation unit And synthesized speech waveform data corresponding to the sentence text can be generated based on the spectrum information and the excitation source information of the second segment dictionary.

  Further, the speech waveform generation means generates synthesized speech waveform data generated from spectrum information and excitation source information corresponding to the matching phrase of the second segment dictionary when the determination means determines that they do not match. Is generated based on the first synthesized speech waveform data excluding the speech segment data portion at the non-coincident side end portion of the text and the spectrum information and excitation source information of the first segment dictionary for the entire sentence text. Generating synthesized speech waveform data corresponding to the sentence text, which is obtained by synthesizing the synthesized speech waveform data with the second synthesized speech waveform data excluding the portion corresponding to the first synthesized speech waveform data. Is possible.

  That is, when it is determined by the determination means that they do not match, the speech data corresponding to the sentence text generated by combining the spectral information of the speech unit and the excitation source information of the first unit dictionary (hereinafter, full-speech speech) Audio data of a predetermined phrase portion (hereinafter referred to as data) (hereinafter referred to as first predetermined phrase audio data) is generated from the spectral information and excitation source information of the phrase unit of the second segment dictionary that matches this phrase. Speech unit data at a connection position with speech unit data of other phrases connected to the first predetermined phrase speech data in the full-speech speech data when the speech data is replaced with the voice data (hereinafter referred to as second predetermined phrase speech data). Is not the speech unit data of the second predetermined phrase voice data, but the first predetermined phrase sound. And to use a voice segment data of the data.

  For example, the second segment dictionary has spectrum information and excitation source information corresponding to “Waiting for (* 1)” following the reading in “Mr. It is assumed that synthesized speech waveform data is generated for a sentence with no punctuation, such as “Mr. XX is waiting”. In such a case, in the case of the above configuration, “Mr. XX” is the full-text voice data generated using the first segment dictionary. It is determined that the phonological environment of “ga” in the voice data part of ”and the phonological environment of“, ”preceding“ Waiting for you (* 1) ”do not match. (* 2) ”“ Oxsangao ”synthesized voice waveform data part and“ Waiting for (* 1) ”“ O ”is excluded,“ Sending ”. It is possible to generate synthesized speech waveform data of “O Mr. XX is waiting. (* 3)” composed of the waveform data part.

  Therefore, when there is a phrase that matches the phrase in the second segment dictionary in the phrase that corresponds to the sentence text, it precedes and follows the phrase that corresponds to the spectrum information and excitation source information that the second segment dictionary has. Even if the predetermined phoneme environment does not match the phoneme environment of the corresponding part of the prosodic information sequence generated by the prosodic information sequence generating means, that part is generated from the spectral information and excitation source information in units of speech units. Therefore, it is possible to generate synthesized speech waveform data in which the matching part corresponding to the sentence text and the sentences before and after the same are connected more smoothly. Is obtained.

Here, the spectrum information and excitation source information for each predetermined phrase included in the second unit dictionary include spectrum information and excitation source information for each speech unit constituting the predetermined phrase uttered by a predetermined speaker, and the predetermined phrase. And time information for each speech unit constituting the phrase. Hereinafter, the speech synthesizing apparatus according to claim 8, claim 1 1 and 1 2, wherein the speech synthesis program, as well, is the same in claim 1 4 and 1 5 speech synthesis method according.

The speech waveform generation means includes a predetermined phoneme environment that has a phrase that matches the phrase included in the sentence text in the second segment dictionary and precedes and follows the phrase in the speech waveform data at the time of extraction. And the phonological environment of the corresponding portion of the generated prosodic information sequence, the synthesized speech waveform data corresponding to all the matching phrases is generated using the second segment dictionary, and the other portions are It is also possible to generate using a single segment dictionary and synthesize these to generate synthesized speech waveform data corresponding to the sentence text. If there is no phrase in the second segment dictionary that matches the phrase included in the sentence text, it is also possible to create synthesized speech waveform data corresponding to the sentence text using only the first segment dictionary. It is. Hereinafter, the speech synthesizing apparatus according to claim 8, claim 1 1 and 1 2, wherein the speech synthesis program, as well, is the same in claim 1 4 and 1 5 speech synthesis method according.

Further, the speech unit data at the side end portion that does not match is at least one of the first speech unit data from the beginning and the end on the time axis in the synthesized speech waveform data. That is, if the side where the phoneme environment does not match is only the leading side, it is the first from the beginning. It becomes speech segment data. Hereinafter, the speech synthesizing apparatus according to claim 8, claim 1 1 and 1 2, wherein the speech synthesis program, as well, is the same in claim 1 4 and 1 5 speech synthesis method according.

In order to achieve the above object, a speech synthesizer according to claim 8 comprises:
A speech synthesizer that generates synthesized speech waveform data corresponding to sentence text,
The prosodic segment dictionary creation method according to any one of claims 1 to 4 or the prosodic segment dictionary creation program according to claim 5 includes the first prosodic information and the prosodic feature information. A prosodic segment dictionary consisting of
Standard prosodic dictionary composed of standard prosodic information that is prosodic information for each voice unit;
A first segment dictionary comprising spectral information for each speech unit and excitation source information for each speech unit;
Extracted from speech waveform data corresponding to an utterance sentence uttered by the predetermined speaker, spectrum information for each predetermined phrase, excitation source information for each predetermined phrase, and preceding and following each predetermined phrase in the speech waveform data A second segment dictionary including information on a predetermined phonetic environment ;
Text analysis means for performing accent analysis and morphological analysis on the sentence text;
Prosodic information sequence generation means for generating a prosodic information sequence corresponding to the sentence text based on the analysis result of the text analysis means and the standard prosodic information of the standard prosodic dictionary;
When a phrase corresponding to the first prosodic information included in the prosodic segment dictionary is included in a phrase corresponding to the sentence text, the prosodic information series portion of the phrase is converted to the first prosodic information. A changing means for changing to the prosodic information sequence part generated based on
The prosodic segment corresponding to the changed prosodic information sequence portion with respect to the prosodic information sequence portion corresponding to at least one of the predetermined prosodic units preceding and following the phrase portion changed by the changing means Prosody information adjusting means for performing predetermined adjustment processing based on the prosodic feature information of the dictionary ;
When the phrase corresponding to the sentence text includes a phrase that matches the phrase corresponding to the spectrum information and excitation source information of the second segment dictionary, the spectrum information and the excitation source information are A predetermined phoneme environment of the second segment dictionary preceding and succeeding the matching phrase in the speech waveform data at the time of extraction, and a predetermined preceding and succeeding a portion corresponding to the matching phrase in the prosodic information sequence Determining means for determining whether or not the phonological environment of
Based on the prosodic information series, the determination result of the determination means, and the spectrum information and the excitation source information of the first and second segment dictionaries, synthetic speech waveform data corresponding to the sentence text is generated. Voice waveform generation means,
The speech waveform generation means, when the determination means determines that they do not match, the synthesized speech waveform data generated from the spectrum information and excitation source information corresponding to the matching phrase of the second segment dictionary The first synthesized speech waveform data excluding the non-coincident side speech segment data part, and the synthesis generated based on the spectral information and excitation source information of the first segment dictionary for the entire sentence text Generating synthesized speech waveform data corresponding to the sentence text, which is synthesized from speech waveform data and second synthesized speech waveform data excluding a portion corresponding to the first synthesized speech waveform data; Yes.

With such a configuration, it is possible to perform accent analysis and morphological analysis on the sentence text by the text analysis unit, and the analysis result of the text analysis unit and the standard by the prosodic information series generation unit It is possible to generate a prosodic information sequence corresponding to the sentence text based on the standard prosodic information possessed by the prosodic dictionary, and a predetermined preceding and succeeding phrase portion changed by the changing means by the prosodic information adjusting means A predetermined adjustment process is performed on the prosodic information sequence portion corresponding to at least one of the prosodic units based on the prosodic feature information of the prosodic segment dictionary corresponding to the changed prosodic information sequence portion It is possible.

Furthermore, when the phrase corresponding to the sentence text includes a phrase that matches the spectrum information and the phrase corresponding to the excitation source information of the second segment dictionary, the spectrum information And a portion corresponding to the matching phrase in the prosodic information sequence and the predetermined phoneme environment of the second segment dictionary preceding and succeeding the matching phrase in the speech data at the time of extraction of the excitation source information It is possible to determine whether or not the preceding and succeeding predetermined phoneme environments match each other, and the prosody information sequence, the determination result of the determination unit, and the first and second by the speech waveform generation unit Generating synthesized speech waveform data corresponding to the sentence text based on the spectrum information and the excitation source information of the segment dictionary; Possible it is.

  Still further, the speech waveform generation means generates a synthesized speech waveform generated from spectrum information and excitation source information corresponding to the matching phrase of the second segment dictionary when the determination means determines that they do not match. Based on the first synthesized speech waveform data excluding the speech segment data portion at the non-matching side edge of the data, and the spectrum information and excitation source information of the first segment dictionary for the entire sentence text Generating synthesized speech waveform data corresponding to the sentence text by synthesizing the generated synthesized speech waveform data with second synthesized speech waveform data excluding a portion corresponding to the first synthesized speech waveform data; Is possible.

  Therefore, based on the prosodic feature information sequence, it is possible to appropriately adjust the prosodic information sequence portion before and after the changed phrase portion according to the prosodic information sequence portion of the changed phrase. Prosodic information that can make the utterance of a sentence including a given phrase and phrases before and after it more natural prosody (inflection, speech speed, volume, etc.) can be generated, so prosody closer to natural speech (inflection, speech speed) Synthetic speech waveform data that is uttered (reproduced) by sound volume, etc. can be generated.

  Furthermore, in addition to the above-described effect, when there is a phrase that matches the phrase in the second segment dictionary in the phrase corresponding to the sentence text, a predetermined number preceding and following the phrase in the speech waveform data Even when the phoneme environment does not match the phoneme environment of the corresponding part in the generated prosodic information sequence, the part is replaced with the speech unit data generated from the spectral information and the excitation source information in units of speech units. Since it is possible to synthesize, synthesized speech waveform data in which the matching part corresponding to the sentence text and the sentence before and after it are uttered (reproduced) with prosody (inflection, speech speed, volume, etc.) closer to natural speech The effect that can be generated is obtained.

Furthermore, the speech synthesizer according to claim 9 is the speech synthesizer according to claim 6 or 8 ,
When the prosody segment dictionary includes a plurality of first prosodic information corresponding to the same phrase as the phrase constituting the sentence text, the changing means converts the prosodic information series portion of the phrase into the prosody. The prosody information sequence part is generated based on the first prosodic information having the best connectivity with the prosodic information series part preceding and following the corresponding phrase in the information series part.

  With such a configuration, for example, the first prosody in which the cost (standard measure of harmony) between the standard prosodic information of a predetermined phoneme environment before and after the phrase to be changed and the first prosodic information of a plurality of change candidates is minimized. Since the prosodic information sequence of only the standard prosodic information can be changed by selecting information, synthesized speech waveform data is generated from the prosodic information sequence changed and adjusted by appropriate first prosodic information and prosodic feature information Therefore, it is possible to generate synthesized speech waveform data in which a sentence including the phrase to be replaced and the phrases before and after it are uttered (reproduced) with prosody (inflection, speech speed, volume, etc.) closer to natural speech. The effect is obtained.

In order to achieve the above object, the speech synthesis program according to claim 1 0, wherein the
A speech synthesis program for generating synthesized speech waveform data corresponding to sentence text,
A text analysis step for performing accent analysis and morphological analysis on the sentence text;
Prosodic information for generating prosodic information corresponding to the sentence text based on the analysis result in the text analyzing step and the standard prosodic information included in the standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit A sequence generation step;
In the phrase corresponding to the sentence text, the prosodic segment dictionary creation method according to any one of claims 1 to 4 or the prosodic segment dictionary creation program according to claim 5 , When a phrase corresponding to the first prosodic information included in the prosodic segment dictionary including the first prosodic information and the prosodic feature information is included, the prosodic information sequence portion of the phrase is converted to the first prosodic information sequence part. A change step to change to the prosodic information sequence part generated based on the information;
The prosodic segment corresponding to the changed prosodic information sequence portion with respect to the prosodic information sequence portion corresponding to at least one of the predetermined prosodic units preceding and following the phrase portion changed in the changing step A prosodic information adjustment step for performing a predetermined adjustment process based on the prosodic feature information of the dictionary ;
Corresponding to the text text based on the spectrum information and the excitation source information of the segment dictionary including the prosodic information series, the spectrum information for each speech unit and the excitation source information for each speech unit A program for causing a computer to execute a process including a voice waveform generation step for generating the synthesized voice waveform data.
With such a configuration, when the program is read by the computer and the computer executes processing according to the read program, the same operation and effect as those of the speech synthesizer according to claim 6 can be obtained.

In order to achieve the above object, the speech synthesis program according to claim 1 1, wherein the
A speech synthesis program for generating synthesized speech waveform data corresponding to sentence text,
A text analysis step for performing accent analysis and morphological analysis on the sentence text;
Prosodic information for generating a prosodic information sequence corresponding to the sentence text based on the analysis result in the text analysis step and the standard prosodic information included in the standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit A sequence generation step;
Spectral information for each predetermined phrase, excitation source information for each predetermined phrase, and the speech waveform data extracted from speech waveform data corresponding to an utterance sentence uttered by a predetermined speaker in the phrase corresponding to the sentence text Including a phrase corresponding to the spectrum information and excitation source information of the second segment dictionary including information of a predetermined phoneme environment preceding and following each of the predetermined phrases in Corresponding to a predetermined phoneme environment of the second segment dictionary preceding and following the matching phrase in the speech data when extracting the spectrum information and the excitation source information, and the matching phrase in the prosodic information series A determination step for determining whether or not a predetermined phonological environment preceding and succeeding the portion to be matched matches;
It said prosodic information sequence, the determination result, sentences first segment dictionary and said comprising an excitation source information of each of the speech segment and the spectral information for each speech segment to the entire text in the determination step A program for causing a computer to execute a process including a speech waveform generation step of generating synthesized speech waveform data corresponding to the sentence text based on the spectrum information and the excitation source information of the second segment dictionary ,
In the speech waveform generation step, when it is determined in the determination step that they do not match, the synthesized speech waveform data generated from the spectrum information and the excitation source information corresponding to the matching phrase of the second segment dictionary From the first synthesized speech waveform data excluding the speech segment data portion at the non-matching side edge, and the synthesized speech waveform data generated based on the spectrum information and excitation source information of the first segment dictionary , the first It is characterized in that synthesized speech waveform data corresponding to the sentence text is generated by synthesizing with second synthesized speech waveform data excluding a portion corresponding to one synthesized speech waveform data.
With such a configuration, when the program is read by the computer and the computer executes processing according to the read program, the same operation and effect as those of the speech synthesizer according to claim 7 can be obtained.

In order to achieve the above object, the speech synthesis program according to claim 1 wherein the
A speech synthesis program for generating synthesized speech waveform data corresponding to sentence text,
A text analysis step for performing accent analysis and morphological analysis on the sentence text;
Prosodic information for generating a prosodic information sequence corresponding to the sentence text based on the analysis result in the text analysis step and the standard prosodic information included in the standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit A sequence generation step;
In the phrase corresponding to the sentence text, the prosodic segment dictionary creation method according to any one of claims 1 to 4 or the prosodic segment dictionary creation program according to claim 5 , When a phrase corresponding to the first prosodic information included in the prosodic segment dictionary including the first prosodic information and the prosodic feature information is included, the prosodic information sequence portion of the phrase is converted to the first prosodic information sequence part. A change step to change to the prosodic information sequence part generated based on the information;
The prosodic segment corresponding to the changed prosodic information sequence portion with respect to the prosodic information sequence portion corresponding to at least one of the predetermined prosodic units preceding and following the phrase portion changed in the changing step A prosodic information adjustment step for performing predetermined adjustment processing based on the prosodic feature information of the dictionary ;
Spectral information for each predetermined phrase, excitation source information for each predetermined phrase, and the speech waveform data extracted from speech waveform data corresponding to an utterance sentence uttered by a predetermined speaker in the phrase corresponding to the sentence text Including a phrase corresponding to the spectrum information and excitation source information of the second segment dictionary including information of a predetermined phoneme environment preceding and following each of the predetermined phrases in and a predetermined phonetic environment preceding and subsequent to the phrase to the match in the audio data at the time of extracting the spectral information and the excitation source information, preceding and succeeding the portions corresponding to the phrases that the match in the prosodic information sequence A determination step for determining whether or not the predetermined phoneme environment of the second segment dictionary matches;
It said prosodic information sequence, the determination result, sentences first segment dictionary and said comprising an excitation source information of each of the speech segment and the spectral information for each speech segment to the entire text in the determination step A program for causing a computer to execute a process including a speech waveform generation step of generating synthesized speech waveform data corresponding to the sentence text based on the spectrum information and the excitation source information of the second segment dictionary ,
In the speech waveform generation step, when it is determined in the determination step that they do not match, the synthesized speech waveform data generated from the spectrum information and the excitation source information corresponding to the matching phrase of the second segment dictionary From the first synthesized speech waveform data excluding the speech segment data portion at the non-matching side edge, and the synthesized speech waveform data generated based on the spectrum information and excitation source information of the first segment dictionary , the first It is characterized in that synthesized speech waveform data corresponding to the sentence text is generated by synthesizing with second synthesized speech waveform data excluding a portion corresponding to one synthesized speech waveform data.
With such a configuration, when the program is read by the computer and the computer executes processing according to the read program, the same operation and effect as those of the speech synthesizer according to claim 8 can be obtained.

In order to achieve the above object, a method of speech synthesis according to claim 1 3, wherein the
A speech synthesis method for generating synthesized speech waveform data corresponding to sentence text,
A text analysis step for performing accent analysis and morphological analysis on the sentence text;
Prosodic information for generating prosodic information corresponding to the sentence text based on the analysis result in the text analyzing step and the standard prosodic information included in the standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit A sequence generation step;
In the phrase corresponding to the sentence text, the prosodic segment dictionary creation method according to any one of claims 1 to 4 or the prosodic segment dictionary creation program according to claim 5 , When a phrase corresponding to the first prosodic information included in the prosodic segment dictionary including the first prosodic information and the prosodic feature information is included, the prosodic information sequence portion of the phrase is converted to the first prosodic information sequence part. A change step to change to the prosodic information sequence part generated based on the information;
The prosodic segment corresponding to the changed prosodic information sequence portion with respect to the prosodic information sequence portion corresponding to at least one of the predetermined prosodic units preceding and following the phrase portion changed in the changing step A prosodic information adjustment step for performing a predetermined adjustment process based on the prosodic feature information of the dictionary ;
Corresponding to the text text based on the spectrum information and the excitation source information of the segment dictionary including the prosodic information series, the spectrum information for each speech unit and the excitation source information for each speech unit And a voice waveform generation step for generating the synthesized voice waveform data.
Thereby, the same effect as that of the speech synthesizer described in claim 6 can be obtained.

In order to achieve the above object, a method of speech synthesis Claim 1 4, wherein the
A speech synthesis method for generating synthesized speech waveform data corresponding to sentence text,
A text analysis step for performing accent analysis and morphological analysis on the sentence text;
Prosodic information for generating a prosodic information sequence corresponding to the sentence text based on the analysis result in the text analysis step and the standard prosodic information included in the standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit A sequence generation step;
Spectral information for each predetermined phrase, excitation source information for each predetermined phrase, and the speech waveform data extracted from speech waveform data corresponding to an utterance sentence uttered by a predetermined speaker in the phrase corresponding to the sentence text Including a phrase corresponding to the spectrum information and excitation source information of the second segment dictionary including information of a predetermined phoneme environment preceding and following each of the predetermined phrases in and a predetermined phonetic environment preceding and subsequent to the phrase to the match in the audio data at the time of extracting the spectral information and the excitation source information, preceding and succeeding the portions corresponding to the phrases that the match in the prosodic information sequence A determination step for determining whether or not the predetermined phoneme environment of the second segment dictionary matches;
It said prosodic information sequence, the determination result, sentences first segment dictionary and said comprising an excitation source information of each of the speech segment and the spectral information for each speech segment to the entire text in the determination step A speech waveform generation step of generating synthesized speech waveform data corresponding to the sentence text, based on the spectrum information and the excitation source information of the second unit dictionary,
In the speech waveform generation step, when it is determined in the determination step that they do not match, the synthesized speech waveform data generated from the spectrum information and the excitation source information corresponding to the matching phrase of the second segment dictionary From the first synthesized speech waveform data excluding the speech segment data portion at the non-matching side edge, and the synthesized speech waveform data generated based on the spectrum information and excitation source information of the first segment dictionary , the first It is characterized in that synthesized speech waveform data corresponding to the sentence text is generated by synthesizing with second synthesized speech waveform data excluding a portion corresponding to one synthesized speech waveform data.
Thereby, the same effect as that of the speech synthesizer described in claim 7 can be obtained.

In order to achieve the above object, a method of speech synthesis according to claim 1 5, wherein the
A speech synthesis method for generating synthesized speech waveform data corresponding to sentence text,
A text analysis step for performing accent analysis and morphological analysis on the sentence text;
Prosodic information for generating a prosodic information sequence corresponding to the sentence text based on the analysis result in the text analysis step and the standard prosodic information included in the standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit A sequence generation step;
In the phrase corresponding to the sentence text, the prosodic segment dictionary creation method according to any one of claims 1 to 4 or the prosodic segment dictionary creation program according to claim 5 , When a phrase corresponding to the first prosodic information included in the prosodic segment dictionary including the first prosodic information and the prosodic feature information is included, the prosodic information sequence portion of the phrase is converted to the first prosodic information sequence part. A change step to change to the prosodic information sequence part generated based on the information;
The prosodic segment corresponding to the changed prosodic information sequence portion with respect to the prosodic information sequence portion corresponding to at least one of the predetermined prosodic units preceding and following the phrase portion changed in the changing step A prosodic information adjustment step for performing a predetermined adjustment process based on the prosodic feature information of the dictionary ;
Spectral information for each predetermined phrase, excitation source information for each predetermined phrase, and the speech waveform data extracted from speech waveform data corresponding to an utterance sentence uttered by a predetermined speaker in the phrase corresponding to the sentence text Including a phrase corresponding to the spectrum information and excitation source information of the second segment dictionary including information of a predetermined phoneme environment preceding and following each of the predetermined phrases in and a predetermined phonetic environment preceding and subsequent to the phrase to the match in the audio data at the time of extracting the spectral information and the excitation source information, preceding and succeeding the portions corresponding to the phrases that the match in the prosodic information sequence A determination step for determining whether or not the predetermined phoneme environment of the second segment dictionary matches;
It said prosodic information sequence, the determination result, sentences first segment dictionary and said comprising an excitation source information of each of the speech segment and the spectral information for each speech segment to the entire text in the determination step A speech waveform generation step of generating synthesized speech waveform data corresponding to the sentence text, based on the spectrum information and the excitation source information of the second unit dictionary,
In the speech waveform generation step, when it is determined in the determination step that they do not match, the synthesized speech waveform data generated from the spectrum information and the excitation source information corresponding to the matching phrase of the second segment dictionary From the first synthesized speech waveform data excluding the speech segment data portion at the non-matching side edge, and the synthesized speech waveform data generated based on the spectrum information and excitation source information of the first segment dictionary , the first The synthesized speech waveform data corresponding to the sentence text is generated by synthesizing the synthesized speech waveform data excluding the portion corresponding to the 1 synthesized speech waveform data.
Thereby, the same effect as that of the speech synthesizer described in claim 8 can be obtained.

[First Embodiment]
A first embodiment of a prosodic segment dictionary creating method, prosodic segment dictionary creating program, speech synthesis apparatus, speech synthesis program, and speech synthesis method according to the present invention will be described below with reference to the drawings. 1 to 8 are diagrams showing a first embodiment of a prosodic segment dictionary creating method and prosodic segment dictionary creating program, a speech synthesizing device, a speech synthesizing program, and a speech synthesizing method according to the present invention.

First, the configuration of the speech synthesizer according to the first embodiment of the present invention will be described with reference to FIG. FIG. 1 is a block diagram showing a configuration of a speech synthesizer 100 according to the first embodiment of the present invention.
As shown in FIG. 1, the speech synthesizer 100 analyzes a text text to be synthesized, generates a pronunciation / prosodic symbol string, and a standard composed of pitch frequency information for each speech unit. A prosodic dictionary 11, a prosodic segment dictionary 12 composed of pitch frequency information in predetermined phrases and prosodic feature information described later, a standard prosodic dictionary 11, and a prosodic segment dictionary 12 based on pronunciation / prosodic symbol strings A prosody generating unit 13 for generating a prosody information sequence using the waveform synthesizing unit 14 for generating synthesized speech waveform data using a standard waveform segment dictionary 15 described later based on the prosodic information sequence, and spectrum information for each speech unit. And a standard waveform segment dictionary 15 composed of excitation source information.

The text analysis unit 10 performs accent analysis and morphological analysis on the text text to be synthesized using a word dictionary (not shown), and the input text text (for example, Japanese kana kanji spelling text) Reading, accent, and intonation are determined, and a pronunciation / prosodic symbol string that is reading information (intermediate language) with prosodic symbols is generated. Further, the generated pronunciation / prosodic symbol string is output to the prosody generation unit 13.
The standard prosody dictionary 11 has standard pitch frequency information that is pitch frequency information corresponding to each label information of speech units. From the standard prosody generation unit 13a described later, a speech unit label sequence corresponding to a pronunciation / prosodic symbol string. Is input to the standard prosody generation unit 13a.

The prosodic segment dictionary 12 is pitch frequency information in units of predetermined phrases created by prosody segment dictionary creation processing described later from speech waveform data corresponding to an utterance sentence composed of a plurality of phrases uttered by a predetermined speaker. It has certain prosodic segment pitch frequency information and corresponding prosodic feature information, and in response to a read request from the prosody segment selecting unit 13c described later, prosodic segment pitch frequency information of the requested phrase and Prosodic feature information is output to the prosodic segment selector 13c.
Here, the prosodic feature information is information indicating prosodic features generated from pitch frequency information of a predetermined phoneme environment preceding or succeeding the phrase when the pitch frequency information of the predetermined phrase is extracted from the speech waveform data. .

The prosody generation unit 13 includes a standard prosody generation unit 13a, a registered phrase collation unit 13b, a prosody segment selection unit 13c, and a prosody replacement shaping unit 13d.
The standard prosody generation unit 13a generates a speech unit label sequence based on the pronunciation / prosodic symbol string input from the text analysis unit 10, and outputs the speech unit label sequence to the standard prosody dictionary 11 to generate a speech unit label sequence. The standard pitch frequency information corresponding to the speech unit sequence corresponding to is acquired, and the standard prosodic information sequence including the standard pitch frequency sequence corresponding to the speech unit label sequence is generated.

Based on the pronunciation / prosodic symbol string input from the text analysis unit 10, the registered phrase matching unit 13 b determines whether the phrase that matches the phrase registered in the prosodic segment dictionary 12 is included in the pronunciation / prosodic symbol string. Determine whether. This determination result is output to the prosodic segment selection unit 13c.
When it is determined that the prosodic segment selection unit 13c includes a matching phrase based on the determination result input from the registered phrase matching unit 13b, all the prosodic segment units corresponding to the matched phrase from the prosodic segment dictionary The pitch frequency information and the prosodic feature information are read from the prosodic segment dictionary 12, the standard prosodic information sequence input from the standard prosody generation unit 13a is selected, and the selected prosodic segment pitch frequency is selected. Information and prosodic feature information are output to the prosody replacement shaping unit 13d. On the other hand, if it is determined that no matching phrase is included, this is notified to the prosody replacement shaping unit 13d.

  The prosody replacement shaping unit 13d receives standard prosody information input from the standard prosody generation unit 13a when prosodic segment pitch frequency information and prosodic feature information corresponding to the matching phrase are input from the prosody segment selection unit 13c. The prosodic information sequence corresponding to the predetermined phoneme environment before and after the matching phrase in the standard prosodic information sequence based on the prosodic feature information, while replacing the standard pitch frequency information of the matching phrase in the sequence with the prosodic segment pitch frequency information The final prosody information sequence is generated by shaping the part. The generated prosodic information sequence is output to the waveform synthesis unit 14.

On the other hand, when the prosody replacement shaping unit 13d receives a notification indicating that there is no matching phrase from the prosody segment selection unit 13c, the standard prosody information sequence input from the standard prosody generation unit 13a is directly input to the waveform synthesis unit 14. Output.
The waveform synthesizer 14 acquires spectrum information and excitation source information corresponding to the prosody information sequence input from the prosody generator 13 from the standard waveform segment dictionary 15, and generates a spectrum information sequence from the acquired spectrum information. At the same time, an excitation source information sequence is generated from the acquired excitation source information. Further, the waveform synthesis unit 14 includes a synthesis filter, and uses the generated spectrum information sequence as a parameter of the synthesis filter, and filters the excitation source signal generated based on the excitation source information sequence with the synthesis filter. Then, synthesized speech waveform data corresponding to the sentence text is generated. Then, based on the generated synthesized speech waveform data, synthesized speech is output.

The standard waveform segment dictionary 15 has spectrum information for each voice unit and excitation source information for each voice unit. In this embodiment, spectral envelope parameters extracted from speech in units of frames are used as spectral information, and the voicing degree (mixing ratio of pulse and noise) of the excitation source signal is used as excitation source information.
Here, in the present embodiment, the speech synthesizer 100, although not shown, stores a storage medium storing a program for controlling each of the above components, a processor for executing these programs, and execution of the programs. And a RAM for storing necessary data. And the process of each said component is implement | achieved by reading and running the program memorize | stored in the storage medium with the processor.

  The storage medium is a semiconductor storage medium such as RAM or ROM, a magnetic storage type storage medium such as FD or HD, an optical reading type storage medium such as CD, CDV, LD, or DVD, or a magnetic storage type / optical such as MO. Any storage medium can be used as long as it is a computer-readable storage medium regardless of electronic, magnetic, optical, or other reading methods.

Furthermore, a method for generating a speech unit label sequence in the prosody generation unit 13 will be described.
In the following, an explanation will be given by taking as an example a case where the sentence of the utterance content is “every reality was twisted towards me”. Here, a case where a speech unit is a phoneme and a combination of the following phonemic environments is used will be described.
-Position of the exhalation paragraph-Position of the accent phrase-Presence / absence of front and back poses-Mora length of the accent phrase-Accent type of the accent phrase-Accent type of the preceding accent phrase-Mora position within the accent phrase of the phoneme・ The phoneme

  When the phonological environment is associated with the utterance example, for example, in the first half of the utterance content, the phonemes are “a”, “r”, “a”, “y”, “u”, “r”, “u”. ”,“ G ”,“ e ”,“ N ”,“ j ”,“ i ”,“ ts ”,“ u ”,“ o ”, the exhalation paragraph becomes“ every reality ”, and the accent phrase (reading) Becomes “everything” and “reality”, and the position of the exhalation paragraph in the sentence is “1”.

Further, for example, the position of the accent phrase in the exhalation paragraph in “everything” is “1”, and the presence or absence of the preceding or following pause is “x” immediately before the accent phrase, and “0” immediately after the accent phrase. ”, The mora length of the accent phrase is“ 4 ”, the accent type is“ 3 ”in the accent phrase,“ x ”in the preceding accent phrase, and the mora position in the accent phrase is“ a ”when“ a ”is“ “1”, “r” and “a” are “2”, “y” and “u” are “3”, and “r” and “u” are “4”.
Here, the speech unit is a phoneme. However, the present invention is not limited to this, and it may be a semiphone or a mora. In addition, the phonetic environment is not limited to the above, but also uses the phonetic environment such as sentence mora length, expiratory paragraph mora length, part of speech, utilization form, utilization type, etc. You may do it.

In the present embodiment, the rules for assigning phoneme labels are as follows.
P_Bpos_Apos_Ppau_Npau_Alen_Atype_APtype_Mpos
Here, in the above-mentioned assignment rule, P is “the phoneme”, Bpos is “the position of the exhalation paragraph”, Apos is “the position of the accent phrase”, and Ppau is “the pose immediately before the accent phrase”. In the presence / absence field, Npau is set to “whether there is a pose immediately after the accent phrase”, Alen is set to “Mora length of the accent phrase”, ACtype is set to “Accent type of the accent phrase”, and APtype is set to “Present accent phrase”. The Mpos corresponds to the “accent type” and “mora position in the accent phrase of the phoneme”.

When a phoneme label is assigned to each phoneme of the utterance content using the above-mentioned assignment rule, for example, the phoneme label for “a” that is “any” phoneme is “a_1_1_x_0_4_3_x_1”. Here, x in the phoneme label indicates that there is no information. Further, a start time (second) and an end time (second) for each phoneme are extracted from the voice data corresponding to the utterance content.
Then, a speech unit label sequence corresponding to the speech / prosodic symbol string is generated from the timed label data composed of the phoneme label for each phoneme and the time information for each phoneme. The standard pitch frequency information corresponding to each phoneme label can be acquired from the standard prosody dictionary 11 from the phoneme label included in the speech unit label series.

Further, a method for creating a prosodic segment dictionary will be described with reference to FIGS. Here, FIG. 2 is a diagram showing a flow of a prosody segment dictionary creation process. Moreover, FIG. 3 is a figure which shows an example of the calculation method of pitch ratio and intonation ratio. FIG. 4A is a diagram showing a configuration example of a prosodic segment dictionary, and FIG. 4B is a diagram showing an example of pitch frequency data registered in the prosody segment dictionary.
In the present embodiment, the prosodic segment dictionary 12 is created from a large number of speech waveform data corresponding to many kinds of uttered sentences uttered by a specific speaker. It should be noted that these many speech waveform data are collected in advance in association with the utterance content information and stored in a storage device (HDD or the like).

  Hereinafter, the flow of the creation process of the prosodic segment dictionary 12 will be described with reference to FIG. In this embodiment, a dedicated prosodic segment dictionary creation program is activated by an information processing device such as a PC (not shown), and the prosody segment dictionary creation process shown in the flowchart of FIG. A segment dictionary 12 is created. Also, the utterance sentence corresponding to the collected speech waveform data is composed of two or more phrases.

As shown in FIG. 2, first, the process proceeds to step S100, and predetermined speech waveform data of a specific speaker is acquired from a storage device (not shown), and the process proceeds to step S102.
In step S102, it is determined whether or not a pre-registered fixed phrase is included in the speech waveform data from the utterance content information of the speech waveform data acquired in step S100. If (Yes), the process proceeds to step S104. If not (No), the process proceeds to step S122. Here, for example, in a car navigation system, the phrase is “turn right to”, “to turn right”, “to turn left”, “to turn left”, “to go straight” ", Etc.", and phrases that are used relatively frequently (relatively frequently appearing) such as "It is!", "Please do", "~ Please" are registered.

When the process proceeds to step S104, the pitch frequency of the phrase portion corresponding to the registered phrase is extracted from the speech waveform data as the first prosodic information, and the process proceeds to step S106.
In step S106, the phrase unit pitch frequency information (prosodic segment pitch frequency information) is generated based on the pitch frequency extracted in step S104, and the process proceeds to step S108.

In step S108, it is determined whether or not there is a phrase preceding the phrase portion corresponding to the registered phrase in the speech waveform data. If it is determined that there is (Yes), the process proceeds to step S110; No) moves to step S112.
When the process proceeds to step S110, the pitch frequency is extracted as second prosodic information from the speech waveform data part of a predetermined prosody unit preceding the corresponding phrase part, and the process proceeds to step S112.

In step S112, it is determined whether or not there is a phrase that follows the phrase portion corresponding to the registered phrase in the audio waveform data. If it is determined (Yes), the process proceeds to step S114; No) moves to step S116.
When the process proceeds to step S114, the pitch frequency is extracted as second prosodic information from the speech waveform data part of a predetermined prosody unit following the corresponding phrase part, and the process proceeds to step S116.

  Here, the predetermined prosody unit that precedes or follows the registered phrase, which is a pitch frequency extraction target in steps S110 and S114, will be described. For example, in the case where the phonetic symbol string of the registered phrase is a phrase that is spoken by a predetermined speaker with the preceding punctuation inside the accent phrase and the subsequent punctuation as the end of the exhalation paragraph, such as “Itasima's + Node” The predetermined prosodic unit preceding the registered phrase is the leading part in the expiratory paragraph and the leading part outside the expiratory paragraph, and the predetermined prosodic unit following the registered phrase is the trailing part in the expiratory paragraph and the trailing part outside the expiratory paragraph. Become. In addition, when the phonetic symbol string of the registered phrase is a phrase that is spoken by a predetermined speaker with the preceding punctuation within the accent phrase or the accent phrase and the subsequent punctuation as the end of the plain text, such as “Morsiagema's” The predetermined prosodic units preceding the registered phrase are the leading part in the expiratory paragraph and the leading part outside the expiratory paragraph, and there is no predetermined prosodic unit following the registered phrase. In addition, when the phonetic symbol string of a registered phrase is a phrase uttered by a predetermined speaker with the preceding punctuation as the end of the exhalation paragraph or accent phrase and the subsequent punctuation as the end of the plain sentence, such as “Migiho's Cordes +” The predetermined prosodic unit preceding the registered phrase is the leading part outside the expiratory paragraph, and there is no predetermined prosodic unit following the registered phrase.

Further, the predetermined prosodic unit is an accent phrase, an exhalation paragraph, or the like depending on the prosodic unit including the registered phrase and the position of the registered phrase with respect to the prosodic unit.
In step S116, the standard pitch frequency information corresponding to the pitch frequency of the prosodic unit preceding or following the registered phrase is generated as the third prosodic information by referring to the standard prosodic dictionary (here, the standard prosodic dictionary 11). The process proceeds to step S118.

In step S118, based on the standard pitch frequency information acquired in step S116 and the pitch frequency of the phoneme environment preceding or following the registered phrase, prosodic feature information corresponding to the registered phrase is generated, and the process proceeds to step S120.
Here, the prosodic feature information includes a ratio between the pitch frequency of the second prosodic information and the corresponding pitch frequency of the third prosodic information (hereinafter referred to as pitch ratio), and the pitch frequency of the second prosodic information. A ratio of the intonation (intonation) obtained from the above and the intonation (intonation) obtained from the pitch frequency of the third prosodic information corresponding thereto (hereinafter referred to as an intonation ratio) is generated.

  For example, as shown in FIG. 3, if the phoneme symbol string of the registered phrase is a phrase “Tenki Kiwa” in an utterance sentence such as “Today's weather is sunny”, for example, it is extracted from speech waveform data. The ratio between the average value of the pitch frequency of “Kyono”, which is the accent phrase of the preceding part in the exhalation paragraph, and the average value of “Kyono” of the standard pitch frequency corresponding thereto is the pitch ratio. For example, the pitch ratio with respect to the predetermined prosodic unit “Kyo '” is “1.2” as shown in FIG.

  On the other hand, as indicated by an arrow in FIG. 3, the difference value (intonation) between the maximum value and the minimum value of the pitch frequency of “Kyono” which is the accent phrase of the preceding part in the exhalation paragraph extracted from the speech waveform data The ratio of the difference between the maximum value and the minimum value of the standard pitch frequency corresponding to this (the size of intonation) is the inflection ratio. For example, the inflection ratio for the predetermined prosodic unit “Kyo '” is “1.5” as shown in FIG.

Also, as shown in FIG. 3, since there is a “Hale'Death” that is a subsequent part outside the exhalation paragraph, the pitch ratio and the inflection ratio are calculated in the same manner as described above. For example, the pitch ratio and the inflection ratio for the predetermined prosodic unit “Hare'Death” are “1.0” and “1.1”, respectively, as shown in FIG.
That is, the pitch ratio and the inflection ratio for each position of the predetermined prosodic unit with respect to the registered phrase is the prosodic feature information.

In step S120, the prosodic segment pitch frequency information generated in step S106 and the prosodic feature information generated in step S118 are associated with each other and stored in a storage device (not shown), and the process proceeds to step S122.
In step S122, it is determined whether or not the processing has been completed for all collected voice waveform data. If it is determined that the processing has ended (Yes), the process proceeds to step S124, and if not (No), The process proceeds to step S100.
In step S124, a prosodic segment dictionary is created based on the prosodic segment pitch frequency information and prosodic feature information stored in the storage device, and the process ends.

  In the present embodiment, when the prosodic segment dictionary is created, in addition to the prosodic segment pitch frequency information and prosodic feature information, the prosody that is the pitch frequency of the start position of the phrase corresponding to the prosodic segment pitch frequency information. Various information such as the segment start pitch (Hz), the prosody segment end pitch (Hz) which is the pitch frequency of the end position, and the prosody segment pitch information address indicating the storage address of the prosody segment pitch frequency data (waveform data) Generate. Further, based on the speech / prosodic symbol string corresponding to the speech waveform data, information on the speech / prosodic symbol string of the registered phrase, the preceding phrase cutting position and the subsequent phrase cutting position of the registered phrase is generated.

Then, information of prosodic feature information, speech / prosodic symbol string, leading phrase cutting position, trailing phrase cutting position, prosodic segment start pitch (Hz), prosodic segment end pitch (Hz), and prosodic segment pitch frequency information address From this, the data table shown in FIG.
That is, the prosodic segment dictionary is composed of the generated data table and prosodic segment pitch frequency data (waveform data). The prosodic segment pitch frequency data (waveform data) is stored in the storage area indicated by the prosodic segment pitch information address registered in the data table of FIG. Therefore, the prosodic segment dictionary according to the present embodiment can detect and acquire the various types of information using the speech / prosodic symbol string as an index.
In the present embodiment, the prosodic segment dictionary 12 of the speech synthesizer 100 is created by the prosody segment dictionary creating method.

Furthermore, the flow of the operation process of the speech synthesizer 100 will be described with reference to FIG. Here, FIG. 5 is a flowchart showing an operation process of the speech synthesizer 100. Note that the standard PF in FIG. 5 is the standard pitch frequency, and the prosodic side PF is the prosodic segment pitch frequency.
As shown in FIG. 5, first, the process proceeds to step S200, where the text analysis unit 10 determines whether or not a sentence text is input via an external device (not shown) or an input device (such as a keyboard). If it is determined (Yes), the process proceeds to step S202. If not (No), the determination process is continued until it is input.

When the process proceeds to step S202, the text analysis unit 10 performs accent analysis and morphological analysis on the text text input in step S200, generates a pronunciation / prosodic symbol string based on the analysis result, The generated pronunciation / prosodic symbol string is output to the prosody generation unit 13, and the process proceeds to step S204.
In step S204, the standard prosody generation unit 13a acquires the corresponding standard pitch frequency information from the standard prosody dictionary 11 based on the pronunciation / prosodic symbol string input from the text analysis unit 10, and uses the acquired standard pitch frequency information as the acquired standard pitch frequency information. Based on the standard prosody information series of the entire sentence text, the process proceeds to step S206.

In step S206, whether the phrase (symbol string) that matches the phrase (symbol string) included in the pronunciation / prosodic symbol string input from the text analysis unit 10 is included in the prosodic segment dictionary 12 in the registered phrase matching unit 13b. It is determined whether or not, and the process proceeds to step S208.
In step S208, if the registered phrase collation unit 13b determines that there is a matching phrase based on the determination result in step S206 (Yes), it notifies the prosody segment selection unit 13c that there is a matching phrase. Then, the process proceeds to step S210. Otherwise (No), the process proceeds to step S224.
In addition, although the process of step S204 and the process of step S206, S208 are performed in order, it is not restricted to this, You may perform these processes in parallel.

  When the process proceeds to step S210, the prosodic segment selection unit 13c selects from the prosodic segment dictionary 12 the prosodic segment pitch frequency information corresponding to the phrase that matches the registered phrase and having the best connection with the preceding and following standard pitch frequencies. Then, the prosodic feature information is acquired, and the acquired prosodic segment pitch frequency information and prosodic feature information are output to the prosodic replacement shaping unit 13d, and the process proceeds to step S212. Here, the connectivity is determined based on the prosodic segment start pitch (Hz) and prosodic segment end pitch (Hz) of the prosodic segment dictionary 12 and the standard pitch frequency before and after the matching phrase. Is the lowest (the smallest difference in frequency) is determined to have good connectivity.

  In step S212, in the prosody replacement shaping unit 13d, based on the prosodic feature information input from the prosodic segment selection unit 13c, the prosody corresponding to a predetermined prosodic unit preceding and following the phrase that matches the registered phrase in the standard prosodic information sequence. The information (standard pitch frequency) sequence portion is adjusted, and the process proceeds to step S214. In the present embodiment, the prosodic feature information is the pitch ratio and the inflection ratio described above. Therefore, the corresponding standard pitch frequency portion is the pitch ratio and the inflection ratio between the standard pitch frequency and the adjusted pitch frequency. Is adjusted to the same ratio as the pitch ratio and intonation ratio of the prosodic feature information.

In step S214, the prosody replacement shaping unit 13d replaces the phrase portion that matches the registered phrase in the standard prosodic information sequence with the prosody segment pitch frequency information input from the prosody segment selection unit 13c, and proceeds to step S216. To do.
In step S216, the prosody replacement shaping unit 13d performs shaping processing so that the prosody segment pitch frequency information input from the prosody segment selection unit 13c and the standard pitch frequency adjusted in step S212 are smoothly connected. Then, a final prosodic information sequence is generated, and the process proceeds to step S218.

In step S218, the prosody replacement shaping unit 13d outputs the prosodic information sequence generated in step S216 to the waveform synthesis unit 14 and proceeds to step S220.
In step S220, the text input in step S200 using the spectrum information and excitation source information of the standard waveform segment dictionary 15 based on the prosody information sequence input from the prosody replacement shaping unit 13d in the waveform synthesizer 14. Synthetic speech waveform data corresponding to the text is generated, and the process proceeds to step S222.

In step S222, the waveform synthesizer 14 outputs the synthesized text of the text text input in step S200 from an output device such as a speaker (not shown) based on the synthesized speech waveform data generated in step S220, and ends the processing. To do.
On the other hand, if there is no matching phrase in step S208 and the process proceeds to step S224, the standard prosody information sequence generated in step S204 is output to the waveform synthesis unit 14 as the final prosody information sequence, and the process proceeds to step S220. .

  Next, the operation of the present embodiment will be described with reference to FIGS. Here, FIG. 6A is a diagram illustrating an example of the pitch frequency sequence in the speech waveform data of the actual utterance sentence including the fixed phrase, and FIG. 6B is a standard pitch with respect to the actual utterance sentence of FIG. It is a figure which shows an example of the pitch frequency series produced | generated only by the frequency. Moreover, FIG. 7 is a figure which shows an example of the standard prosodic information series with respect to the text text containing a fixed phrase. FIG. 8A is a diagram showing an example of a prosodic information sequence in which a prosodic segment pitch frequency is used for a fixed phrase in a sentence text and an adjustment process based on prosodic feature information is performed. It is a figure which shows an example of the prosodic information series produced | generated by the method of the prior art with respect to the same text as (a).

  Here, a specific speaker A utters a sentence whose pronunciation / prosodic symbol string is “Kono 'Tabi, Itensurukoto'Ninarima'Sitade, Osilase Mosiagema's", and the registered phrase is “Mosiagema's”. The case where the prosodic segment dictionary 12 has the prosodic segment pitch frequency information and prosodic feature information generated from the speech waveform data of the utterance sentence will be described.

  When the specific speaker A utters the phrase “Morsia Gemma's”, which is a registered phrase, as understood from the prosodic information sequence generated from the speech waveform data of the actual utterance shown in FIG. In the exhalation paragraph that precedes (in the example of FIG. 6 (a), “Kono 'Tabi, Itensurukoto'Ninarima'Sitanode”) is slightly stronger, while in the same exhalation paragraph, the immediately preceding phrase (FIG. 6 (a In the example of), there is a tendency to suppress the inflection of “Oshirase”).

  On the other hand, when the same prosodic information sequence as in FIG. 6A is generated only with the standard pitch frequency information possessed by the standard prosodic dictionary 11, as shown in FIG. 6B, it is compared with the example of FIG. Since the prosodic information sequence part of each phrase is generated independently of the registered phrase, it can be seen that the intonation is monotonous. In addition, the waveform shown with the broken line in FIG.6 (b) becomes a standard pitch.

Here, prosodic feature information is generated for the exhalation paragraph preceding the registration phrase and the immediately preceding phrase in the same exhalation paragraph as the registration phrase. That is, as shown in FIG. 6A, the prosodic feature information (pitch ratio “1.1”, inflection ratio “1.5”) and “Oshirase” Prosodic feature information (pitch ratio “0.85”, intonation ratio “0.6”) is generated.
Therefore, the prosodic segment dictionary 12 has the same prosodic feature information for the prosodic segment pitch frequency information of “Mosia Gema's” and the exhalation paragraph “Kono 'Tabi, Itensurukoto'Ninarima'Sitade” preceding this phrase. The prosodic feature information for “Oshirase” preceding the phrase in the exhalation paragraph is registered.

Hereinafter, the actual operation of the speech synthesizer 100 in which the prosodic segment dictionary 12 has the prosodic segment pitch frequency information and the prosodic feature information of the “Mosia Gemas” will be described.
First, the text text “Thank you very much for your special assistance” is input to the text analysis unit 10 of the speech synthesizer 100 (step S200).
The text analysis unit 10 performs accent analysis and morphological analysis on the input sentence text by referring to the word dictionary, and generates a pronunciation / prosodic symbol string for the sentence text based on the analysis result (step S202). ). In this case, the pronunciation / prosodic symbol string is “Kakubetsu no Ohi + Kitateni Azuka ', Atsu + Quonley Mosia Gemma”. This pronunciation / prosodic symbol string is input to the standard prosody generation unit 13a and the registered phrase collation unit 13b of the prosody generation unit 13, respectively.

  The standard prosody generation unit 13 a generates a speech unit label sequence based on the input pronunciation / prosodic symbol string, and acquires standard pitch frequency information corresponding to the label information of the speech unit label sequence from the standard prosody dictionary 11. . Then, a standard prosodic information sequence corresponding to the pronunciation / prosodic symbol string is generated from the acquired standard pitch frequency information (step S204). The standard prosodic information sequence (standard pitch frequency sequence) generated in this way has a monotonous inflection as a whole, as shown in FIG. The generated standard prosodic information series is input to the prosodic replacement shaping unit 13d. In addition, the waveform shown with the broken line in FIG. 7 becomes a standard pitch.

  On the other hand, the registered phrase collation unit 13b determines whether or not a registered phrase is included in the pronunciation / prosodic symbol string based on the input pronunciation / prosodic symbol string (step S206). Since the pronunciation / prosodic symbol string includes “Morcia Gemma's” as a registered phrase (“Yes” branch of step S208), this is notified to the prosody segment selection unit 13c.

  When receiving the notification from the registered phrase collating unit 13b, the prosody segment selecting unit 13c has a plurality of prosody segment pitch frequency information and prosody feature information corresponding to the registered phrase “Mosia Gemma” from the prosody segment dictionary 12. In this case, the prosodic segment pitch frequency information and prosodic feature information of “Mosia Gemas” having the best connectivity with the standard pitch frequency in the standard prosodic information sequence generated by the standard prosody generating unit 13a is selected. As described above, as described above, the prosodic segment dictionary 12 has a plurality of “Mosia Gemas” prosodic segment start pitches (Hz) and prosodic segment end pitches (Hz), and standard prosody information. Based on the standard pitch frequencies before and after the phrase “Morcia Gemma's” in the series, the one with the smallest frequency difference is selected based on the judgment that the connectivity is the best. Here, the prosodic feature information (pitch ratio “1.1”, intonation ratio “1.5”) of the preceding expiratory paragraph and the prosodic feature information (pitch ratio) of the preceding phrase in the same expiratory paragraph, which are described as the premise, It is assumed that the prosodic segment pitch frequency information and prosodic feature information of “Mosia Gemma's” having “0.85” and intonation ratio “0.6”) is selected. Then, the prosodic segment selection unit 13c acquires the selected prosodic segment pitch frequency information and prosodic feature information from the prosody segment dictionary 12 and outputs it to the prosody replacement shaping unit 13d (step S210).

  When the prosodic segment pitch frequency information and prosodic feature information of “Morsia Gemma's” is input from the prosodic segment selection unit 13c, the prosodic replacement shaping unit 13d precedes “Morcia Gemma's” in the standard prosodic information sequence. Adjusted using the pitch ratio “1.1” and the inflection ratio “1.5”, which are the prosodic feature information, for the standard pitch frequency corresponding to “Kakubetsuno Ohi + Kitateni Azuka'ri” in the exhalation paragraph Is performed (step S212).

  Here, as in the example of FIG. 6A described above, the specific speaker A, when speaking the phrase of “Mosiagema's” selected above, has an exhalation paragraph (here “Kakubetsuno”) preceding this phrase. Ohi + Kitateni Azuka 'Li, ")) is slightly strengthened, while in the same exhalation paragraph, it tends to suppress the inflection of the immediately preceding phrase (here" Atsu + Quantley "). Therefore, as shown in FIG. 8 (a), “Kakubetsu no Ohi + Kitateni Azuka'ri, Atsu + Quantley” standard prosodic information sequence (standard pitch frequency sequence) is generated from actual speech data. By simply connecting the prosodic information sequence of “Mosia Gemma” (prosodic segment pitch frequency sequence), a standard prosodic information sequence is generated for each preceding phrase independently of the registered phrase as described above. Therefore, the standard prosodic information sequence having an inflection different from the inflection on the preceding phrase of the specific speaker A and the prosodic information sequence for the actual utterance are connected, resulting in an unnatural inflection as a whole. Therefore, in order to correct the unnatural inflection so that it becomes a natural intonation, the standard prosodic information series is adjusted based on the prosodic feature information. In addition, the waveform shown with the broken line in Fig.8 (a) becomes a standard pitch.

  Specifically, the series is such that the pitch ratio between the standard pitch frequency series corresponding to “Kakubetsunohi + Kitatenizazu'ri” and the adjusted pitch frequency series is 1.1 and the inflection ratio is 1.5. Adjust each standard pitch frequency in the. Similarly, in the standard prosodic information sequence, the pitch ratio between the standard pitch frequency sequence corresponding to “Atsu + Quantley”, which is the preceding phrase in the same exhalation paragraph as “Mosia Gema's”, and the adjusted pitch frequency sequence is Each standard pitch frequency in the sequence is adjusted so that the inflection ratio is 0.8.

  Further, the prosody replacement shaping unit 13d replaces the standard pitch frequency sequence corresponding to “Morsia Gemma's” in the standard prosodic information sequence with the prosodic segment pitch frequency information input from the prosodic segment selection unit 13c (Step S1). S214), performing a shaping process such as a smoothing process so that the prosodic segment pitch frequency information (prosodic segment pitch frequency sequence) after the replacement and the adjusted standard prosodic information sequence are smoothly connected, As shown in FIG. 8 (b), a final prosodic information sequence for “Kakubetsu Noh + Kitateni Azuka ', Atsu + Quonley Mosia Gemas” is generated (step S216). Then, the generated prosodic information sequence is output to the waveform synthesizer 14 (step S218). In addition, the waveform shown with the broken line in FIG.8 (b) becomes a standard pitch after adjustment.

  The waveform synthesizing unit 14 acquires spectrum information and excitation source information corresponding to the prosodic information sequence input from the prosody generating unit 13 from the standard waveform segment dictionary 15, and generates a spectrum information sequence from the acquired spectral information. Then, an excitation source information sequence is generated from the acquired excitation source information. Then, using the generated spectrum information sequence as a parameter of the synthesis filter, the synthesis filter filters the excitation source signal generated based on the excitation source information sequence, and generates synthesized speech waveform data corresponding to the sentence text (Step S220). Further, based on the generated synthesized speech waveform data, synthesized speech is output from a speaker (not shown) or the like (step S222).

  As described above, the speech synthesizer 100 according to the present embodiment uses the pitch frequency information (prosodic segment pitch frequency) of the registered phrase (standard phrase) extracted from the speech waveform data corresponding to the utterance sentence uttered by the predetermined speaker. Information), a prosodic feature information indicating a prosodic feature of a predetermined prosodic unit generated based on a pitch frequency for a predetermined prosodic unit preceding and following the fixed phrase in the speech waveform data, and a standard pitch frequency corresponding to the pitch frequency; The standard phrase part included in the standard prosodic information sequence generated from the standard pitch frequency information that is the pitch frequency information for each label information of the speech unit is used as the prosodic segment dictionary 12 having the prosodic segment dictionary 12. It replaces with the prosodic segment pitch frequency information it has. Further, the standard pitch frequency sequence portion of a predetermined prosody unit preceding and following the fixed phrase in the standard prosodic information sequence is adjusted based on the prosodic feature information corresponding to the prosodic segment pitch frequency information of the prosodic segment dictionary 12 Since the final prosodic information sequence is generated, it is possible to generate synthesized speech waveform data in which the sentence of the sentence text is uttered (reproduced and output) with an inflection closer to a natural utterance.

In the first embodiment, the text analysis unit 10 corresponds to the text analysis unit according to claim 6 , and the prosody segment dictionary 12 is the prosody segment according to any one of claims 1 to 5. Corresponding to the dictionary, the prosody generation unit 13 corresponds to the prosody information sequence generation means according to claim 6 and corresponds to the registered phrase by the registered phrase collation unit 13b, the prosody segment selection unit 13c, and the prosody replacement shaping unit 13d. The processing for replacing the standard prosodic information sequence portion with the prosodic information sequence portion composed of the prosodic segment pitch frequency information of the prosodic segment dictionary 12 corresponds to the changing means according to claim 6 or 9 , and the prosody replacement shaping unit 13d. process of adjusting, based on the prosodic feature information prosodic information sequence portion of a given prosody units preceding and succeeding the replacement part by corresponds to the prosody information adjustment unit according to claim 6, the waveform synthesizer 14, claim 6 Corresponding to the speech waveform generation means mounting, standard waveform segment dictionary 15 corresponds to segment dictionary according to any one of claims 6, 10 and 13.

  In the first embodiment, steps S100 to S104 correspond to the first prosodic information extraction step according to claim 1 or 5, and steps S108 to S114 correspond to the second prosodic information extraction step. Step S116 corresponds to the third prosodic information generation step according to claim 1 or 5, S118 corresponds to the prosodic feature information generation step according to claim 1 or 5, and step S124 corresponds to claim 1 or 5. This corresponds to the step of creating a prosodic segment dictionary.

In the first embodiment, steps S200 to S202 correspond to the text analysis step according to claim 10 or 13 , and steps S204 and S216 correspond to the prosodic information sequence generation step according to claim 10 or 13. Correspondingly, steps S208, S210, and S214 correspond to the changing step according to claim 10 or 13 , step S212 corresponds to the prosodic information adjustment step according to claim 10 or 12 , and step S220 corresponds to claim 10. Or it corresponds to the speech waveform generation step described in 13 .

[Second Embodiment]
Next, a second embodiment of the speech synthesizer, speech synthesis program, and speech synthesis method according to the present invention will be described with reference to the drawings. 9 to 13 are diagrams showing a second embodiment of the speech synthesizer, speech synthesis program, and speech synthesis method according to the present invention.
First, the configuration of the speech synthesizer according to the second embodiment of the present invention will be described with reference to FIGS. Here, FIG. 9 is a block diagram showing a configuration of speech synthesis apparatus 200 according to the second exemplary embodiment of the present invention. FIG. 10 is a block diagram showing a detailed configuration of the waveform synthesizer 23 of the speech synthesizer 200.

  As shown in FIG. 9, the speech synthesizer 200 analyzes a text text to be synthesized and generates a pronunciation / prosodic symbol string, and a standard composed of pitch frequency information for each speech unit. A standard prosody generation unit 22 that generates a standard prosody information sequence using the standard prosody dictionary 21 based on the prosody dictionary 21 and the pronunciation / prosodic symbol string; a standard waveform segment dictionary 24 that will be described later based on the standard prosody information sequence; A waveform synthesizer 23 for generating synthesized speech waveform data using a large waveform segment dictionary 25 described later, a standard waveform segment dictionary 24 composed of spectrum information and excitation source information for each speech unit, and for each phrase unit A large waveform segment dictionary 25 composed of spectrum information and excitation source information is included.

  The text analysis unit 20 performs accent analysis and morphological analysis on the text text to be synthesized using a word dictionary (not shown) and the like, and the input text text (for example, Japanese kana kanji characters) Reading, accent, and intonation are determined, and a pronunciation / prosodic symbol string that is reading information (intermediate language) with prosodic symbols is generated. Further, the generated pronunciation / prosodic symbol string is output to the prosody generation unit 22.

The standard prosody dictionary 21 has standard pitch frequency information which is pitch frequency information corresponding to each label information of speech units, and a speech unit label sequence corresponding to a pronunciation / prosodic symbol string is input from the standard prosody generation unit 22. Then, the corresponding standard pitch frequency information is output to the standard prosody generation unit 22.
The standard prosody generation unit 22 generates a speech unit label sequence based on the pronunciation / prosodic symbol string input from the text analysis unit 20 and outputs the speech unit label sequence to the standard prosody dictionary 21 to generate a speech unit label sequence. Is obtained, and a standard prosodic information sequence including a standard pitch frequency sequence corresponding to the speech unit label sequence is generated.

As shown in FIG. 10, the waveform synthesis unit 23 includes a unit selection unit 23a, a unit connection unit 23b, and a synthesis unit 23c.
Based on the standard prosodic information sequence input from the standard prosody generation unit 22 and the phoneme environment information included in the large waveform segment dictionary 25, the segment selection unit 23a has a spectrum for each speech unit corresponding to the standard prosodic information sequence. Information and excitation source information are selected from the standard waveform segment dictionary 24 and the large waveform segment dictionary 25. Information on the selection result is output to the segment connection unit 23b.

  Specifically, it is determined whether or not the pronunciation / prosodic symbol string of the standard prosodic information sequence includes a phrase that matches the registered phrase of the large waveform segment dictionary 25, and the matching phrase is included. For the first and last mora of the phrase, following and preceding the first and last mora in the speech waveform data (hereinafter referred to as original sound data) at the time of extraction of the spectrum information and excitation source information of the phrase. It is determined whether the phoneme environment matches the phoneme environment following and preceding the first and last mora of the matching phrase in the standard prosodic information sequence. When the phoneme environments match, spectrum information and excitation source information corresponding to the first or last mora are selected from the large waveform segment dictionary 25, and when the phoneme environments do not match, the standard waveform segment dictionary 24 Spectral information and excitation source information corresponding to the first or last mora are selected. For the parts other than the matching phrases, spectrum information and excitation source information are selected from the standard waveform segment dictionary 24 for all mora. That is, when a phrase that matches the pronunciation / prosodic symbol string is not included, spectrum information and excitation source information are selected from the standard waveform segment dictionary 24 for all standard prosodic information sequences.

  Based on the information on the selection result from the element selection unit 23a, the element connection unit 23b obtains spectrum information for each speech unit corresponding to the information on the selection result from the standard waveform element dictionary 24 and the large waveform element dictionary 25. The excitation source information is acquired from the standard waveform segment dictionary 24 and the large waveform segment dictionary 25. Then, the acquired spectrum information and excitation source information for each speech unit are connected to generate a spectrum information sequence and an excitation source information sequence. The generated spectrum information sequence and excitation source information sequence are output to the synthesis unit 23c.

The synthesizing unit 23c includes a synthesizing filter, and uses the spectrum information sequence input from the unit connection unit 23b as a parameter of the synthesizing filter. The synthesizing filter generates an excitation source signal generated based on the excitation source information sequence. Filter processing is performed to generate synthesized speech waveform data corresponding to the sentence text. Then, based on the generated synthesized speech waveform data, synthesized speech is output.
The standard waveform segment dictionary 24 has spectrum information for each voice unit and excitation source information for each voice unit. In the present embodiment, spectral envelope parameters extracted from speech in units of frames are used as spectral information, and the voicing level of the excitation source signal is used as excitation source information. Here, the audio unit is “Mora”.

  The large waveform segment dictionary 25 has spectrum information for each phrase (large segment) unit and excitation source information for each phrase (large segment) unit. In the present embodiment, spectral envelope parameters extracted from speech in units of frames are used as spectral information, and the voicing level of the excitation source signal is used as excitation source information. Note that the spectrum information for each phrase unit and the excitation source information for each phrase unit are the time information of the spectrum information for the speech unit that constitutes the spectrum information for each phrase and the speech unit (mora) that constitutes the excitation source information for each phrase. The duration length information of the excitation source information is included. Therefore, it is possible to extract spectrum information and excitation source information for each voice unit from spectrum information for each phrase unit and excitation source information for each phrase unit.

The large waveform segment dictionary 25 also has mora information of a predetermined phoneme environment that precedes and follows the first and last mora of each phrase in the original sound data. This information is used by the segment selection unit 23a, and thereby, it is determined whether the phoneme environment in the original sound data matches the phoneme environment in the standard prosodic information sequence.
Hereinafter, based on FIG. 11, the detailed structure of the large waveform segment dictionary 25 is demonstrated. Here, FIG. 11 is a diagram illustrating a configuration example of the large waveform segment dictionary.

  As shown in FIG. 11, the large waveform segment dictionary 25 includes, for example, original sound data of a sentence “the next intersection is in the right direction” including a registered (standard) phrase “in the right direction”. The spectrum information and excitation source information of the phrase portion “rightward” extracted from the above, and information of a predetermined phoneme environment preceding and following the phrase. In the case of “to the right”, since it is the phrase that comes after the reading and at the end of the sentence, the information “pau (pause)” of the mora preceding and following the phrase in the original sound data is here. It becomes information of a predetermined phoneme environment. Note that n and m in FIG. 11 are the subband orders of the spectrum information and the excitation source information, respectively. Furthermore, as shown in FIG. 11, the large waveform segment dictionary 25 includes mora sequence information, which is information of voice units (mora units) constituting the registered phrase, and a duration length for each voice unit (mora) described above. Information (number of frames). That is, the large-waveform segment dictionary 25 has, for each of a plurality of types of registered phrases, mora sequence information including spectral information, excitation source information, preceding and succeeding mora information, and each voice unit (mora). It becomes the structure with each duration time information.

  Here, in the present embodiment, the speech synthesizer 200, although not shown, stores a storage medium storing a program for controlling each of the above components, a processor for executing these programs, and execution of the programs. And a RAM for storing necessary data. And the process of each said component is implement | achieved by reading and running the program memorize | stored in the storage medium with the processor.

  The storage medium is a semiconductor storage medium such as RAM or ROM, a magnetic storage type storage medium such as FD or HD, an optical reading type storage medium such as CD, CDV, LD, or DVD, or a magnetic storage type / optical such as MO. Any storage medium can be used as long as it is a computer-readable storage medium regardless of electronic, magnetic, optical, or other reading methods.

Furthermore, the flow of the operation process of the speech synthesizer 200 will be described with reference to FIG. Here, FIG. 12 is a flowchart showing an operation process of the speech synthesizer 200.
As shown in FIG. 12, first, the process proceeds to step S300, where the text analysis unit 20 determines whether or not text text is input via an external device (not shown), an input device (such as a keyboard), and the like. If it is determined (Yes), the process proceeds to step S302. If not (No), the determination process is continued until it is input.

When the process proceeds to step S302, the text analysis unit 20 performs accent analysis and morphological analysis on the sentence text input in step S300, generates a pronunciation / prosodic symbol string based on the analysis result, The generated pronunciation / prosodic symbol string is output to the standard prosody generation unit 22, and the process proceeds to step S304.
In step S304, the standard prosody generation unit 22 acquires the corresponding standard pitch frequency information from the standard prosody dictionary 21 based on the pronunciation / prosodic symbol string input from the text analysis unit 20, and uses the acquired standard pitch frequency information as the acquired standard pitch frequency information. Based on this, a standard prosodic information sequence for the entire text is generated, and the generated standard prosodic information sequence is output to the waveform synthesis unit 23, and the process proceeds to step S306.

In step S306, the segment selection unit 23a selects from the standard prosody information sequence input from the standard prosody generation unit 22 the standard prosody information sequence part of the mora unit for which segment selection has not been processed, and proceeds to step S308. .
In step S308, the segment selection unit 23a determines whether or not the mora selected in step S306 is a mora in the registered phrase based on the standard prosody information sequence input from the standard prosody generation unit 22. If it is determined that the mora is in the phrase (Yes), the process proceeds to step S310. If not (No), the process proceeds to step S326.

When the process proceeds to step S310, the segment selection unit 23a determines whether the mora selected in step S306 is the first or last mora that constitutes the registered phrase, and is the first or last mora. If it is determined (Yes), the process proceeds to step S312; otherwise (No), the process proceeds to step S314.
When the process proceeds to step S312, in the segment selection unit 23a, in the original sound data corresponding to the registered phrase including the mora selected in step S306, a predetermined phoneme environment preceding or succeeding the selected mora, and a standard prosodic information sequence It is determined whether or not a predetermined phoneme environment preceding or succeeding the selected mora in step S3 is determined. If it is determined to match (Yes), the process proceeds to step S314; otherwise (No), step S314 is performed. The process proceeds to S326.

When the process proceeds to step S314, the segment selection unit 23a determines to select the spectrum information and excitation source information corresponding to the standard prosodic information sequence of the mora selected in step S306 from the large waveform segment dictionary 25. Then, the process proceeds to step S316.
In step S316, the segment connection unit 23b acquires spectrum information and excitation source information corresponding to the standard prosodic information sequence of the mora selected in step S306 from the large waveform segment dictionary 25 based on the selection result in step S314. Then, the process proceeds to step S318.

In step S318, the segment connection unit 23b connects the spectrum information and the excitation source information acquired in step S316 or step S328 to generate a spectrum information sequence and an excitation source information sequence, and the process proceeds to step S320.
In step S320, the segment connecting unit 23b determines whether the mora selected in step S306 is the last mora. If it is determined that the last mora is the last mora (Yes), the spectrum information generated in step S318 is determined. The sequence and the excitation source information sequence are output to the synthesizer 23c, and the process proceeds to step S322. Otherwise (No), the process proceeds to step S306.

When the process proceeds to step S322, the synthesis unit 23c generates synthesized speech waveform data based on the spectrum information sequence and the excitation source information series input from the segment connection unit 23b, and the process proceeds to step S324.
In step S324, the waveform synthesizer 23 outputs a synthesized sound based on the synthesized speech waveform data generated in step S322, and the process ends.

On the other hand, if the mora selected in step S306 is not in the registered phrase, or if the phonetic environment preceding or succeeding the selected mora does not match and the process proceeds to step S326, the standard prosodic information of the mora selected in step S306 It is determined that the spectrum information and the excitation source information corresponding to the series are selected from the standard waveform segment dictionary 24, and the process proceeds to step S328.
In step S328, the segment connection unit 23b acquires spectrum information and excitation source information corresponding to the standard prosodic information sequence of the mora selected in step S306 from the standard waveform segment dictionary 24 based on the selection result in step S326. Then, the process proceeds to step S318.

  Next, the operation of the present embodiment will be described with reference to FIG. Here, FIG. 13A is a diagram showing the structure of the original sound data for the example sentence, and FIG. 13B is a synthesis example in the case where the phoneme environment of the preceding mora of the registered phrase matches the example sentence of FIG. (C) is a figure which shows the example of a synthesis | combination in case the phonetic environment of the preceding mora of a registration phrase does not correspond with the example sentence of (a).

  Hereinafter, as shown in FIG. 13A, the large waveform segment dictionary 25 extracts spectral information extracted from the phrase “right” in the real voice data “next intersection is right”. The actual operation of the speech synthesizer 200 having source information, mora sequence information including preceding and subsequent mora information, and duration time information for each voice unit (mora) will be described. In this case, the preceding mora information is “o_m”. Here, “_m” in “o_m” indicates that the mora subsequent to the mora “o” is “m”.

First, the text text “In front of elementary school is in the right direction” is input to the text analysis unit 20 of the speech synthesizer 200 (step S300).
The text analysis unit 20 performs accent analysis and morphological analysis on the input sentence text with reference to the word dictionary, and generates a pronunciation / prosodic symbol string for the sentence text based on the analysis result (step S302). ). This pronunciation / prosodic symbol string is input to the standard prosody generation unit 22.

  The standard prosody generation unit 22 generates a speech unit label sequence based on the input pronunciation / prosodic symbol string, and acquires standard pitch frequency information corresponding to the label information of the speech unit label sequence from the standard prosody dictionary 21. . Then, a standard prosodic information sequence corresponding to the pronunciation / prosodic symbol string is generated from the acquired standard pitch frequency information (step S304). The standard prosodic information sequence (standard pitch frequency sequence) generated in this way is input to the segment selection unit 23 a of the waveform synthesis unit 23.

  When the standard prosodic information sequence is input, the segment selecting unit 23a sequentially extracts the standard prosodic information sequence portion (standard pitch frequency sequence portion) of mora units in which the segment selection processing is not processed in order from the beginning of the standard prosodic information sequence. It is selected (step S306), and it is determined whether or not the mora of the selected standard prosodic information sequence part is included in the registered phrase of the large waveform segment dictionary 25 (step S308). Here, assuming that the registered phrase is not included in the phrase “Before Elementary School”, it is determined that the mora of the standard prosodic information sequence portion of this phrase is not included in the registered phrase (“No” in step S308). Branch), it is determined that the spectrum information and excitation source information corresponding to each selected standard prosodic information sequence portion are selected from the standard waveform segment dictionary 24 (step S326). Based on the selection result, the segment connecting unit 23b acquires spectrum information and excitation source information corresponding to each selected standard prosody information sequence portion from the standard waveform segment dictionary 24 (step S328), and sequentially connects them. Then, a spectrum information sequence and an excitation source information sequence are generated (step S318).

  On the other hand, since the phrase “rightward” in the input sentence text is a registered phrase, the mora of the standard prosodic information sequence portion selected for this phrase is determined to be a mora in the registered phrase. ("Yes" branch of step S308). It should be noted that the mora sequence information of “is rightward” including the preceding and succeeding mora of the large waveform segment dictionary 25 is “o_m” “mi” “gi” “ho” “o” “ko” “o”. “De” “su” “pau”. When the standard prosodic information sequence portion corresponding to “mi” is selected from the standard prosodic information sequence, this mora becomes the first mora of the registered phrase (the branch of “Yes” in step S310). Whether or not the phoneme environment information preceding the phrase “It is in the right direction” in the waveform segment dictionary 25 (above “o_m”) matches the information of the phoneme environment preceding the phrase in the standard prosodic information series. Is determined (step S312). Here, the mora sequence information for the standard prosodic information sequence of “is the right direction” in the input sentence text “in front of elementary school” is “o_m”, “mi_g”, “gi_h”, “ho_o”. “O_k”, “ko_o”, “o_d”, “de_s”, “su”, “pau”, and the preceding mora becomes “o_m”. Therefore, since it matches “o_m” that is the preceding mora of “It is in the right direction” of the large waveform segment dictionary 25 (“Yes” branch of step S312), “mi” of the selected standard prosodic information sequence is selected. It is determined that the spectrum information and the excitation source information corresponding to the standard prosodic information sequence portion corresponding to are selected from the large waveform segment dictionary 25 (step S314). Then, the segment connection unit 23b acquires the spectrum information and excitation source information corresponding to the standard prosodic information sequence portion corresponding to “mi” from the large waveform segment dictionary 25 (step S316), and the spectrum information and excitation The source information is connected to the already generated spectrum information series and excitation source information series of “Before Elementary School” (step S318).

  Further, since “gi”, “ho”, “o”, “ko”, “o”, and “de” following “mi” are not the first or last mora of the registered phrase, the spectrum information and the excitation source corresponding to these are not included. The information is determined to be selected from the large waveform segment dictionary 25 (step S314), and the segment connection unit 23b acquires the spectrum information and the excitation source information corresponding to these from the large waveform segment dictionary 25. (Step S316), it connects to the already generated spectrum information sequence and excitation source information sequence (Step S318). Note that since the end “su” comes to the end of the sentence, in this case, it is determined in step S312 that the phoneme environments are unconditionally matched, and the spectrum information and the excitation source information are set to the large waveform segment. It is decided to select from the dictionary 25. Then, when spectrum information and excitation source information corresponding to “su” which is a mora at the end of the sentence are acquired from the large waveform segment dictionary 25 and connected to the already generated spectrum information sequence and excitation source information sequence, FIG. As shown in (b), for the text “Before Elementary School”, spectral information and excitation source information acquired from the standard waveform segment dictionary 24 are used (corresponding to the mora sequence enclosed by the dotted line in the figure). ), A spectrum information sequence and an excitation source information sequence are generated, and for the text “in the right direction”, the spectrum information and the excitation source information obtained from the large waveform segment dictionary 25 are used (indicated by a solid line in the figure). A spectral information sequence and an excitation source information sequence are generated. Since “su” is the last mora of the registered phrase, the segment connection unit 23b outputs the generated spectrum information sequence and excitation source information sequence to the synthesis unit 23c (“Yes” branch of step S320). .

  Further, a case where the text text “To the right” is input to the text analysis unit 20 will be described. When the standard prosodic information sequence “to the right” is generated in the same procedure as described above (step S304), the segment selecting unit 23a performs unselected segment selection processing in order from the beginning of the standard prosodic information sequence. A standard prosodic information sequence portion (standard pitch frequency sequence portion) in units of mora is selected (step S306), and whether the mora of the selected standard prosodic information sequence portion is included in the registered phrase of the large waveform segment dictionary 25 or not. Is determined (step S308). Here, assuming that the registered phrase is not included in the phrase “heading”, the segment connecting unit 23b, as in the case of “Before Elementary School”, the spectrum information corresponding to each selected standard prosodic information sequence portion. The excitation source information is acquired from the standard waveform segment dictionary 24 (step S328), and these are sequentially connected to generate the spectrum information sequence and the excitation source information sequence (step S318).

  On the other hand, since the phrase “rightward” in the input sentence text is a registered phrase, the mora of the standard prosodic information sequence portion selected for this phrase is determined to be a mora in the registered phrase. ("Yes" branch of step S308). Here, the mora sequence information for the standard prosodic information sequence of the text sentence that is input is “to the right” is “te_m”, “mi_g”, “gi_h”, “ho_o”, “o_k”, “ko_o”, “o_d”. “De_s”, “su”, “pau”, and the preceding mora is “te_m”. Therefore, since it does not coincide with “o_m” that is the preceding mora of “It is in the right direction” of the large waveform segment dictionary 25 (“No” branch in step S312), “0” of the selected standard prosodic information sequence is selected. It is determined that the spectrum information and the excitation source information corresponding to the standard prosodic information sequence portion corresponding to “mi” are selected from the standard waveform segment dictionary 25 (step S326). Then, the segment connection unit 23b acquires the spectrum information and excitation source information corresponding to the standard prosody information sequence portion corresponding to “mi” from the standard waveform segment dictionary 25 (step S328), and the spectrum information and excitation. The source information is connected to the already generated “toward” spectrum information sequence and excitation source information sequence (step S318). As for “gi”, “ho”, “o”, “ko”, “o”, “de”, and “su” following “mi”, it is not the first or last mora of the registered phrase or the mora at the end of the sentence. Therefore, it is determined that the spectrum information and the excitation source information corresponding to these are selected from the large waveform segment dictionary 25 (step S314), and the segment connection unit 23b selects the spectrum information and the excitation source information corresponding thereto. Is obtained from the large waveform segment dictionary 25 (step S316) and connected to the already generated spectrum information sequence and excitation source information sequence (step S318).

  As shown in FIG. 13B, this connection result is obtained by using the spectrum information and excitation source information obtained from the standard waveform segment dictionary 24 for the text text “Look” (indicated by the dotted line in the figure). The spectrum information sequence and the excitation source information sequence are generated, and the spectrum information and excitation obtained from the large waveform segment dictionary 25 for the sentence text “in the right direction” are generated. Using the source information (corresponding to a mora sequence surrounded by a solid line in the figure), a spectrum information sequence and an excitation source information sequence are generated. Then, the element connection unit 23b outputs the generated spectrum information sequence and excitation source information sequence to the synthesis unit 23c (“Yes” branch of step S320).

  Based on the spectrum information sequence and the excitation source information sequence input from the unit connection unit 23b, the synthesis unit 23c uses the spectrum information sequence as a parameter for the synthesis filter, and is generated by the synthesis filter based on the excitation source information sequence. The excitation source signal is filtered to generate synthesized speech waveform data corresponding to the sentence text (step S322). Further, based on the generated synthesized speech waveform data, synthesized speech is output from a not-shown speaker or the like (step S324).

  As described above, the speech synthesizer 200 according to the present embodiment includes the spectrum information and excitation source information of the registered phrase (standard phrase) extracted from the speech waveform data corresponding to the utterance sentence uttered by the predetermined speaker, A standard prosody using a large waveform segment dictionary 25 having information on the phoneme environment preceding or succeeding the phrase part, duration information for each speech unit constituting the phrase, and mora sequence information of the phrase The combined waveform data of the fixed phrase part included in the information series can be generated using the spectrum information and excitation source information of the large waveform segment dictionary 25. Furthermore, when the information of the phoneme environment preceding or following the fixed phrase and the information of the phoneme environment preceding and following the fixed phrase in the standard prosodic information sequence do not match, the first or last mismatched side edge of the fixed phrase Can be generated using the spectrum information and excitation source information included in the standard waveform segment dictionary 24, and the other mora can be generated using the spectrum information and excitation source information included in the large waveform segment dictionary 25. is there. As a result, even when the phoneme environments do not match as described above, it is possible to generate synthesized speech waveform data in which the sentence of the sentence text is uttered (reproduced and output) with an inflection closer to a natural utterance.

In the second embodiment, the text analysis unit 20 corresponds to the text analysis unit described in claim 7 , and the standard prosody generation unit 22 corresponds to the prosody information sequence generation unit described in claim 7 , and waveform synthesis is performed. part 23 corresponds to the speech waveform generation means according to claim 7, wherein the standard waveform segment dictionary 24 corresponds to the first segment dictionary according to any one of claims 7, 11 and 14, the large The waveform segment dictionary 25 corresponds to the second segment dictionary according to any one of claims 7 , 11 and 14 .

In the second embodiment, when the mora selected from the standard prosody information sequence in the segment selection unit 23a is the first or last mora constituting the registered phrase, the original sound data (large Whether the predetermined phoneme environment preceding or succeeding the selected mora in the waveform segment dictionary 25) matches the predetermined phoneme environment preceding or succeeding the selected mora in the standard prosodic information sequence The processing to be performed corresponds to the determination means described in claim 7 .

Further, in the second embodiment, steps S300~S302 correspond to the text analysis step according to claim 11, step S304 corresponds to the prosody information sequence generating step according to claim 11 or 14, wherein the step S306 to S328 correspond to the speech waveform generation step according to claim 11 or 14 .
Moreover, in the said 2nd Embodiment, step S312 respond | corresponds to the determination step of Claim 11 or 14 .

[Third Embodiment]
Next, a third embodiment of the prosody segment dictionary creation method and prosody segment dictionary creation program, speech synthesizer, speech synthesis program, and speech synthesis method according to the present invention will be described with reference to the drawings. 14 to 17 are diagrams showing a third embodiment of the prosody segment dictionary creation method and prosody segment dictionary creation program, and the speech synthesizer, speech synthesis program, and speech synthesis method according to the present invention.

Here, FIG. 14 is a block diagram showing a configuration of a speech synthesizer 300 according to the third embodiment of the present invention.
As shown in FIG. 14, the speech synthesizer 300 includes a text analysis unit 10, a standard prosody dictionary 11, a prosody segment dictionary 12, a prosody generation unit 13, a waveform synthesis unit 23, and a standard waveform segment dictionary 24. And a large waveform segment dictionary 25.

That is, the speech synthesizer 300 is the same as the speech synthesizer 100 in the first embodiment, the text analysis unit 10, the standard prosody dictionary 11, the prosody segment dictionary 12, and the prosody generation unit 13, and the second Similar to the speech synthesizer 200 in this embodiment, the waveform synthesizer 23, the standard waveform segment dictionary 24, and the large waveform segment dictionary 25 are included.
Therefore, based on the prosody information sequence generated by the prosody generation unit 13, the waveform synthesis unit 23 generates synthesized speech waveform data using the standard waveform segment dictionary 24 and the large waveform segment dictionary 25, and outputs synthesized speech. It becomes composition. Hereinafter, only points different from the first embodiment and the second embodiment will be described in detail.

  The prosody replacement shaping unit 13d of the prosody generation unit 13 receives input from the standard prosody generation unit 13a when prosodic segment pitch frequency information and prosodic feature information corresponding to the matching phrase is input from the prosody segment selection unit 13c. In the standard prosodic information sequence, the standard pitch frequency information of the matching phrase is replaced with the prosodic segment pitch frequency information, and based on the prosodic feature information, the predetermined prosodic unit before and after the matching phrase in the standard prosodic information sequence is replaced. The corresponding prosodic information sequence portion is shaped to generate a final prosodic information sequence. The generated prosodic information sequence is output to the waveform synthesis unit 23.

On the other hand, when the prosody replacement shaping unit 13d receives a notification indicating that there is no matching phrase from the prosody segment selection unit 13c, the standard prosody information sequence input from the standard prosody generation unit 13a is directly input to the waveform synthesis unit 23. Output.
The segment selection unit 23 a of the waveform synthesis unit 23 is based on the prosody information sequence input from the prosody generation unit 13 and the phoneme environment information included in the large waveform segment dictionary 25, and the speech unit corresponding to the standard prosody information sequence. Each spectrum information and excitation source information is selected from the standard waveform segment dictionary 24 and the large waveform segment dictionary 25. Information on the selection result is output to the segment connection unit 23b.

  That is, the prosodic information sequence input from the prosody generating unit 13 to the waveform synthesizing unit 23 is a fixed phrase part included in the standard prosodic information sequence generated from the standard pitch frequency information that is the pitch frequency information for each piece of speech label information. Is replaced with the prosodic segment pitch frequency information of the prosodic segment dictionary 12. Further, the standard pitch frequency sequence of a predetermined prosody unit preceding and following the registered (standard) phrase in the standard prosodic information sequence is adjusted based on the prosodic feature information corresponding to the prosodic segment pitch frequency information of the prosodic segment dictionary 12 Has been. For such a prosodic information series, for the standard phrase part, the synthesized waveform data is generated using the spectrum information extracted from the original sound data and the excitation source information, while the standard phrase in the original sound data is preceded or succeeded. If the information on the phonetic environment to be matched does not match the information on the phoneme environment preceding and following the standard phrase in the standard prosodic information sequence, the mora at the first or last non-matching side edge of the standard phrase is used as the standard waveform segment. It is generated using the spectrum information and excitation source information that the dictionary 24 has.

  Here, in the present embodiment, the speech synthesizer 300, although not shown, stores a storage medium storing a program for controlling each of the above components, a processor for executing these programs, and execution of the programs. And a RAM for storing necessary data. And the process of each said component is implement | achieved by reading and running the program memorize | stored in the storage medium with the processor.

  The storage medium is a semiconductor storage medium such as RAM or ROM, a magnetic storage type storage medium such as FD or HD, an optical reading type storage medium such as CD, CDV, LD, or DVD, or a magnetic storage type / optical such as MO. Any storage medium can be used as long as it is a computer-readable storage medium regardless of electronic, magnetic, optical, or other reading methods.

Further, the flow of the operation process of the speech synthesizer 300 will be described with reference to FIG. Here, FIG. 15 is a flowchart showing an operation process of the speech synthesizer 300. Note that the standard PF in FIG. 15 is the standard pitch frequency, and the prosodic side PF is the prosodic segment pitch frequency.
In addition, since the process of step S400-step S418 and step S424 is the same as that of step S200-step S218 in the flowchart of FIG. 5 of the said 1st Embodiment, and step S224, description is abbreviate | omitted.

In step S420, the waveform synthesis unit 23 generates synthesized speech waveform data based on the prosody information sequence input from the prosody generation unit 13, and the process proceeds to step S422.
In step S422, the waveform synthesizer 23 outputs the synthesized text of the text text input in step S400 from an output device such as a speaker (not shown) based on the synthesized speech waveform data generated in step S420, and the process ends. To do.

Next, based on FIG. 16, the flow of the synthetic speech waveform data generation process in step S420 will be described. Here, FIG. 16 is a flowchart showing a process of generating synthesized speech waveform data in the waveform synthesizer 23.
As shown in FIG. 16, first, the process proceeds to step S500, and the segment selection unit 23a selects a prosodic information sequence portion of a mora unit for which segment selection is not processed from the prosody information sequence input from the prosody generation unit 13. Then, the process proceeds to step S502.

In step S502, the segment selection unit 23a determines whether or not the mora selected in step S500 is a mora in the registered phrase based on the prosody information sequence input from the prosody generation unit 13. If it is determined that the mora is (No), the process proceeds to step S504. If not (No), the process proceeds to step S518.
When the process proceeds to step S504, the unit selection unit 23a determines whether the mora selected in step S500 is the first or last mora that constitutes the registered phrase, and is the first or last mora. If determined (Yes), the process proceeds to step S506, and if not (No), the process proceeds to step S508.

  When the process proceeds to step S506, in the segment selection unit 23a, in the original sound data corresponding to the registered phrase including the mora selected in step S500, the predetermined phoneme environment preceding or succeeding the selected mora, and the prosody information sequence It is determined whether or not a predetermined phoneme environment preceding or succeeding the selected mora matches. If it is determined that they match (Yes), the process proceeds to step S508, and if not (No), step S518 is determined. Migrate to

When the process proceeds to step S508, the unit selection unit 23a determines that the spectrum information and excitation source information corresponding to the prosodic information sequence portion of the mora selected in step S500 are selected from the large waveform unit dictionary 25. Then, the process proceeds to step S510.
In step S510, the segment connection unit 23b acquires spectrum information and excitation source information corresponding to the prosodic information sequence portion of the mora selected in step S500 from the large waveform segment dictionary 25 based on the selection result in step S508. Then, the process proceeds to step S512.

In step S512, the unit connection unit 23b connects the spectrum information and the excitation source information acquired in step S510 or step S520 to generate a spectrum information sequence and an excitation source information sequence, and the process proceeds to step S514.
In step S514, the segment connecting unit 23b determines whether the mora selected in step S500 is the last mora. If it is determined that the last mora is (Yes), the spectrum information generated in step S512 is determined. The sequence and the excitation source information sequence are output to the synthesizing unit 23c, and the process proceeds to step S516. Otherwise (No), the process proceeds to step S500.
When the process proceeds to step S516, the synthesis unit 23c generates synthesized speech waveform data based on the spectrum information sequence and the excitation source information sequence input from the unit connection unit 23b, ends the series of processing, and returns to the original processing. Return to processing.

On the other hand, if the mora selected in step S500 is not in the registered phrase, or if the phoneme environment preceding or succeeding the selected mora does not match and the process proceeds to step S518, the prosodic information sequence of the mora selected in step S500 It is determined that the spectrum information and the excitation source information corresponding to the part are selected from the standard waveform segment dictionary 24, and the process proceeds to step S520.
In step S520, the segment connection unit 23b acquires spectrum information and excitation source information corresponding to the prosodic information sequence portion of the mora selected in step S500 from the standard waveform segment dictionary 24 based on the selection result in step S518. Then, the process proceeds to step S512.

Next, the operation of the present embodiment will be described based on FIG. Here, FIG. 17A is a diagram showing a synthesized speech waveform generated using the conventional technique when the phoneme environments of the preceding mora of the registered phrase are inconsistent, and FIG. It is a figure which shows the synthetic | combination audio | voice waveform produced | generated using this invention when the phonetic environment of mora is inconsistent.
First, the text text “Before, Hirakawamon is in the right direction” is input to the text analysis unit 10 of the speech synthesizer 300 (step S400).

  As in the first embodiment, the text analysis unit 10 performs accent analysis and morphological analysis on the input sentence text by referring to the word dictionary, and based on the analysis result, A pronunciation / prosodic symbol string is generated (step S402). The pronunciation / prosodic symbol string in this case is “Sonosaki +, Hirakawamonma 'Eomigiho'-Cordes +." This pronunciation / prosodic symbol string is input to the standard prosody generation unit 13a and the registered phrase collation unit 13b of the prosody generation unit 13, respectively.

  The standard prosody generation unit 13 a generates a speech unit label sequence based on the input pronunciation / prosodic symbol string, and acquires standard pitch frequency information corresponding to the label information of the speech unit label sequence from the standard prosody dictionary 11. . Then, a standard prosodic information sequence corresponding to the pronunciation / prosodic symbol string is generated from the acquired standard pitch frequency information (step S404). The generated standard prosody information sequence is input to the prosody replacement shaping unit 13d.

  On the other hand, the registered phrase collation unit 13b determines whether or not a registered phrase is included in the pronunciation / prosodic symbol string based on the input pronunciation / prosodic symbol string (step S406). Here, the pronunciation / prosodic symbol string includes “Sonosaki +” and “Migiho's Cordes +” as registration phrases (the branch of “Yes” in step S208), and this is the prosodic segment selection unit. 13c is notified.

  When receiving the notification from the registered phrase collating unit 13b, the prosodic segment selecting unit 13c receives from the prosody segment dictionary 12 the prosody segment pitch frequency information corresponding to the registered phrases “Sonosaki +” and “Migiho's Cordes +” and When there are a plurality of prosodic feature information, as in the first embodiment, the prosodic segment of “Sonosaki +” and “Migiho's Cordes +” having the best connectivity with the standard pitch frequency in the standard prosodic information sequence Select pitch frequency information and prosodic feature information. Then, the prosodic segment selection unit 13c acquires the selected prosodic segment pitch frequency information and prosodic feature information from the prosody segment dictionary 12 and outputs it to the prosody replacement shaping unit 13d (step S410).

  The prosodic replacement shaping unit 13d, when the prosodic segment pitch frequency information and prosodic feature information of “Sonosaki +” and “Migiho's chords +” are input from the prosodic segment selecting unit 13c, The standard pitch frequency corresponding to “Hirakawamonma'eo” in the exhalation paragraph following “+” is adjusted using the pitch ratio and the inflection ratio, which are the prosodic feature information corresponding thereto (step S412). In this case, “Hirakawamonma'Eo” also becomes an exhalation paragraph preceding “Migiho's Cordes +”, so the standard pitch frequency corresponding to “Hirakawamonma'eo” is adjusted using two types of prosodic feature information. .

  Specifically, for example, the pitch ratio and the inflection ratio for “Sonosaki +” are “0.8” and “0.6”, and the pitch ratio and the inflection ratio for “Migiho's Cordes +” are “0.8”. And “0.6”, the two match, so the pitch ratio between the standard pitch frequency sequence corresponding to “Hirakawamonma'eo” and the adjusted pitch frequency sequence is 0.8, and the inflection ratio is 0. Each standard pitch frequency in the sequence is adjusted to be .6. If the two do not match, for example, the average value of the pitch ratio and the inflection ratio of the two is used, or the exhalation paragraph of “Hirakawamonma'eo” is divided into two phrases, one of which is “Sonosaki +” The other is adjusted with the pitch ratio and the inflection ratio with respect to “Migiho's Cordes +”.

  Further, the prosodic replacement shaping unit 13d receives the standard pitch frequency sequence corresponding to “Sonosaki +” and “Migiho's Cordes +” in the standard prosodic information sequence, and the prosodic segment pitch frequency information input from the prosodic segment selection unit 13c. (Step S414) and shaping such as smoothing processing so that the replaced prosodic segment pitch frequency information (prosodic segment pitch frequency sequence) and the adjusted standard prosodic information are smoothly connected. Processing is performed to generate a final prosodic information sequence for “Sonosaki +, Hirakawamonma 'Eomigiho-Cordes +.'” (Step S416). Then, the generated prosodic information sequence is output to the segment selection unit 23a of the waveform synthesis unit 23 (step S418).

  When a prosodic information sequence is input, the segment selecting unit 23a selects a prosodic information sequence portion (standard pitch frequency sequence portion) of mora units in which the segment selection processing has not been processed in order from the beginning of the prosodic information sequence ( Step S500), it is determined whether or not the mora of the selected prosodic information sequence portion is included in the registered phrase of the large waveform segment dictionary 25 (Step S502). Here, “Sonosaki +” and “Migiho'-Cordes +” are registered phrases. Therefore, it is determined that the mora of the prosodic information sequence portion selected for these phrases is the mora in the registered phrase (“Yes” branch of step S504). The “Sonosaki +,” mora sequence information of the large waveform segment dictionary 25 is “so”, “no”, “sa”, “ki”, “pau (pause)”, and the mora of “Omigiho's Cordes +”. The series information is “pau”, “mi”, “gi”, “ho”, “o”, “ko”, “o”, “de”, “su”.

  In addition, when a plurality of prosodic segments correspond to a registered phrase, individual large waveform segments (spectrum information and excitation source information) based on the same speech waveform data are large for each prosodic segment. You may store in the waveform segment dictionary 25. FIG. In this case, as shown in the example of FIG. 4, when a prosodic segment piece corresponding to the prosodic segment pitch frequency information registered in the prosodic segment dictionary 12 is actually spoken, The storage address information of the extracted large waveform segment in the large waveform segment dictionary 25 is stored in advance in the prosodic segment dictionary 12. As a result, an appropriate large waveform segment corresponding to the prosody segment pitch frequency information registered in the prosody segment dictionary 12 can be understood, so that a more natural synthesized sound that more accurately reflects the speech waveform data at the time of utterance can be obtained. Is possible.

  When the prosodic information sequence portion corresponding to “so” is selected from the prosodic information sequence, this mora becomes the first mora of the registered phrase (the branch of “Yes” in step S504), but the mora at the beginning of the sentence. Therefore, it is determined that the phoneme environments are unconditionally matched (branch “Yes” in step S506), and the spectrum information and the excitation source information corresponding to the prosodic information sequence portion of this mora are converted into large waveform segments. It is determined to be selected from the dictionary 25 (step S508). Then, the segment connecting unit 23b acquires the spectrum information and the excitation source information corresponding to the prosodic information sequence portion corresponding to “so” from the large waveform segment dictionary 25 (Step S510), and the spectrum information and the excitation source. The information is set to the head of the spectrum information series and the excitation source information series (step S512). Since “no” and “sa” following “so” are mora in the registered phrase and not the first and last mora of the phrase, the spectrum information and the excitation source corresponding to these prosodic information sequence parts The information is determined to be selected from the large waveform segment dictionary 25 (step S508), and the segment connection unit 23b converts the spectrum information and excitation source information corresponding to the prosodic information sequence portion of “no” and “sa” into the large waveform. Obtained from the segment dictionary 25 (step S510), the spectrum information and the excitation source information are connected to the spectrum information sequence and the excitation source information sequence corresponding to “so” (step S512).

  Furthermore, when the prosodic information sequence portion corresponding to “ki” is selected from the prosodic information sequence, this mora becomes the last mora of the registered phrase (the branch of “Yes” in step S504). It is determined whether or not the phonological environment information following the phrase “Sonosaki +” possessed by the piece dictionary 25 (above “pau”) matches the phonological environment information following the phrase in the prosodic information sequence ( Step S506). Here, the mora sequence information for the prosodic information sequence of “beyond,” which is the input sentence text is “so_n”, “no_s”, “sa_k”, “ki”, “pau”, and the subsequent mora is “pau”. . Accordingly, since it matches “pau” that is a subsequent mora of “Sonozaki +” in the large waveform segment dictionary 25 (“Yes” branch in step S506), it corresponds to “ki” in the selected prosodic information sequence portion. It is determined that the spectrum information and the excitation source information corresponding to the prosodic information sequence portion to be selected are selected from the large waveform segment dictionary 25 (step S508). Then, the segment connecting unit 23b acquires the spectrum information and excitation source information corresponding to the standard prosodic information sequence portion corresponding to “ki” from the large waveform segment dictionary 25 (step S510), and the spectrum information and the excitation information. The source information is connected to the already generated “sonosa” spectrum information sequence and excitation source information sequence (step S512).

  Since the prosodic information sequence portion corresponding to each mora of “Hirakawamonma 'Eo” following “Sonosaki +” is not included in the registered phrase (the branch of “No” in step S502), the prosody of these mora It is determined that spectrum information and excitation source information corresponding to the information series portion are selected from the standard waveform segment dictionary 24 (step S518). Then, based on this selection result, the segment connection unit 23b obtains spectrum information and excitation source information corresponding to each selected prosodic information sequence portion from the standard waveform segment dictionary 24 (step S520), and “Sonosaki + The spectrum information sequence and the excitation source information sequence are connected in order to generate a spectrum information sequence and an excitation source information sequence (step S512).

  Furthermore, when the prosodic information sequence portion corresponding to the mora “mi” in “Migiho's Cordes +” is selected from the prosodic information sequence, this mora becomes the first mora of the registered phrase, so this prosodic information sequence portion Is determined to be a mora in the registered phrase ("Yes" branch of step S502). It should be noted that the mora sequence information of “is in the right direction” including the preceding mora of the large waveform segment dictionary 25 is “pau”, “mi”, “gi”, “ho”, “o”, “ko”, “o”, “de”. “Su”. In addition, the mora sequence information for the prosodic information sequence portion of “Omigiho's Cordes +” in the input sentence text “Beyond Hirakawamon Mae” is “o_m” “mi_g” “gi_h” “ ho_o "" o_k "" ko_o "" o_d "" de_s "" su ". That is, “pau”, which is the preceding mora included in the large waveform segment dictionary 25, does not match “o_m”, which is the preceding mora for the prosodic information sequence portion of “Migiho's Cordes +” (“No” in step S506). Branch).

  Accordingly, it is determined that the spectrum information and the excitation source information corresponding to the prosodic information sequence portion corresponding to “mi” of the selected prosodic information sequence are selected from the standard waveform segment dictionary 25 (step S518). Then, the segment connecting unit 23b acquires the spectrum information and the excitation source information corresponding to the prosody information sequence portion corresponding to “mi” from the standard waveform segment dictionary 25 (step S520), and the spectrum information and the excitation source. The information is connected to the already generated spectrum information sequence and excitation source information sequence of “Sonosaki +, Hirakawamonma'eo” (step S512). In addition, “gi”, “ho”, “o”, “ko”, “o”, “de”, and “su” following “mi” are not the first or last mora of the registered phrase, or become a mora at the end of the sentence. Then, it is determined to select the spectrum information and the excitation source information corresponding to these from the large waveform segment dictionary 25 (step S508), and the segment connection unit 23b selects the spectrum information and the excitation source information corresponding to these, Obtained from the large waveform segment dictionary 25 (step S510) and connected to the already generated spectrum information sequence and excitation source information sequence (step S512).

  This connection result shows that the spectrum information sequence and the excitation source information sequence are generated using the spectrum information and the excitation source information acquired from the large waveform segment dictionary 25 for the pronunciation / prosodic symbol string “Sonosaki +”. For the pronunciation / prosodic symbol string “Hirakawamon Omi”, the spectrum information sequence and the excitation source information sequence are generated using the spectrum information and the excitation source information acquired from the standard waveform segment dictionary 24. For the prosodic symbol string “Giho's Cordes +”, the spectrum information sequence and the excitation source information sequence are generated using the spectrum information and the excitation source information acquired from the large waveform segment dictionary 25. Then, the element connection unit 23b outputs the generated spectrum information sequence and excitation source information sequence to the synthesis unit 23c (“Yes” branch of step S514).

Based on the spectrum information sequence and the excitation source information sequence input from the unit connection unit 23b, the synthesis unit 23c uses the spectrum information sequence as a parameter for the synthesis filter, and is generated by the synthesis filter based on the excitation source information sequence. The excitation source signal is filtered to generate synthesized speech waveform data corresponding to the sentence text (step S420).
As shown in FIG. 17B, the synthesized speech waveform data generated in this way is a synthesized speech waveform of “H” of “Hirakawamonma'eo” and a synthesized speech of “mi” of “Migiho's Cordes +”. The waveform is connected smoothly. When synthesized speech is output from a speaker or the like (not shown) based on the generated synthesized speech waveform data (step S422), the text of the sentence is pronounced with a prosody closer to natural utterance, and “O” of “Hirakawamonma'Eo”. And “Mi” of “Migiho's Cordes +” are pronounced continuously.

  On the other hand, as in the conventional method, the spectrum information sequence and the excitation source information sequence are obtained using the spectrum information and the excitation source information acquired from the large waveform segment dictionary 25 for the entire registered phrase “Migiho's Cordes +”. When generating the synthesized speech waveform data, as shown in the synthesized speech waveform of FIG. 17A, the synthesized speech waveform of “H” of “Hirakawamonma'Eo” and “Mikiho's Cordes +” of “M ”And the synthesized voice waveform becomes discontinuous. Accordingly, when a synthesized sound having such a synthesized speech waveform is output, “H” of “Hirakawamonma'Eo” and “Mi” of “Migiho'-Cordés +” are discontinuously pronounced, resulting in an uncomfortable feeling. As described above, the speech synthesizer 300 according to the present embodiment has the pitch frequency information (prosody segment pitch frequency information) of the registered phrase (standard phrase) extracted from the speech waveform data corresponding to the utterance sentence uttered by the predetermined speaker. And prosodic feature information indicating the prosodic features of the predetermined prosodic unit generated based on the pitch frequency for the predetermined prosodic unit preceding and following the fixed phrase in the speech waveform data and the standard pitch frequency corresponding thereto. Using the prosodic segment dictionary 12, the prosodic segment dictionary 12 has a fixed phrase part included in a standard prosodic information sequence generated from standard pitch frequency information that is pitch frequency information for each piece of label information in speech units. Replace with fragment pitch frequency information. Further, the standard pitch frequency sequence of a predetermined prosodic unit preceding and following the fixed phrase in the standard prosodic information sequence is adjusted based on the prosodic feature information corresponding to the prosodic segment pitch frequency information of the prosodic segment dictionary 12, It is possible to generate a final prosodic information sequence.

  Further, the spectrum information and excitation source information of the registered phrase (standard phrase) part extracted from the speech waveform data corresponding to the utterance sentence uttered by the predetermined speaker, information on the phoneme environment preceding or succeeding the phrase part, Using the large waveform segment dictionary 25 having the duration length information for each voice unit constituting the phrase and the mora sequence information of the phrase, the synthesized waveform data of the fixed phrase part included in the prosodic information sequence is obtained. It is possible to generate using the spectrum information and excitation source information of the large waveform segment dictionary 25. This makes it possible to generate synthesized speech waveform data using the spectrum information and excitation source information of the naturally-sound original sound data in addition to the prosody close to that of natural utterances. It is possible to generate synthesized speech waveform data in which a sentence is uttered (reproduced and output) with an inflection closer to a natural utterance.

  Furthermore, when the information on the phonetic environment preceding or following the fixed phrase and the information on the phonemic environment preceding and following the fixed phrase in the prosodic information sequence do not match, the first or last mismatched side edge of the fixed phrase Can be generated using the spectrum information and excitation source information included in the standard waveform segment dictionary 24, and the other mora can be generated using the spectrum information and excitation source information included in the large waveform segment dictionary 25. is there. As a result, even when the phonological environments do not match, it is possible to generate synthesized speech waveform data in which the sentence of the sentence text is uttered (reproduced and output) with an inflection closer to natural utterance.

In the third embodiment, the text analysis unit 10 corresponds to the text analysis unit according to claim 8 , and the prosodic segment dictionary 12 is added to the prosody dictionary according to any one of claims 1 to 5. Correspondingly, the prosody generation unit 11 corresponds to the prosody information sequence generation means according to claim 8 , and the standard prosody corresponding to the registered phrase by the registered phrase collation unit 13b, the prosody segment selection unit 13c, and the prosody replacement shaping unit 13d. The processing for replacing the information sequence portion with the prosodic information sequence portion configured from the prosodic segment pitch frequency information of the prosodic segment dictionary 12 corresponds to the changing means according to claim 8 , and is replaced with the replacement portion by the prosodic replacement shaping unit 13d. leading and process adjustments to basis prosodic information sequence portion of the subsequent predetermined prosodic units prosodic feature information corresponds to a prosodic information adjustment unit according to claim 8, the waveform synthesis section 23, the audio according to claim 8 Corresponding to the shape generating means, the standard waveform segment dictionary 24 corresponds to the first segment dictionary according to claim 8, large waveform segment dictionary 25 corresponds to the second segment dictionary according to claim 8 .

In the third embodiment, when the mora selected from the standard prosodic information sequence in the segment selection unit 23a is the first or last mora constituting the registered phrase, the original sound data (large Whether the predetermined phoneme environment preceding or succeeding the selected mora in the waveform segment dictionary 25) matches the predetermined phoneme environment preceding or succeeding the selected mora in the standard prosodic information sequence The processing to be performed corresponds to the determination means described in claim 8 .

In the third embodiment, steps S400 to S402 correspond to the text analysis step according to claim 12 or 15 , and steps S404 and S416 correspond to the prosodic information sequence generation step according to claim 12 or 15. correspondingly, the step S408, S410, S414 corresponds to the changing step according to claim 12 or 15, wherein, step S412 corresponds to the prosody information adjustment step of claim 12 or 15, wherein, step S420 is claim 12 Or it corresponds to the speech waveform generation step described in 15 .

In the third embodiment, step S506 corresponds to the determination step according to claim 12 or 15 .
In the first to third embodiments, Japanese has been described as an example. However, the present invention is not limited to this, and the present invention may be applied to languages other than Japanese.
In the first and third embodiments, the example has been described in which the prosodic information included in the prosodic segment dictionary is the pitch frequency. In addition to the above, or in place of the pitch frequency information, other information related to prosody such as speech speed information and volume information may be included.

  In the first and third embodiments, the example has been described in which the prosodic feature information included in the prosodic segment dictionary is the pitch ratio and the inflection ratio. In addition to the pitch ratio and the inflection ratio, or in place of the pitch ratio and the inflection ratio, the information may include other information obtained from the pitch frequency of the original sound data and the standard pitch frequency. May include information obtained from the speech speed information and volume information of the original sound data, and the standard speech speed information and standard volume information. In other words, in addition to the pitch frequency information, the corresponding prosodic information sequence part is adjusted by the speech speed information and the volume information, so that the text of the sentence is uttered (reproduced) with an inflection closer to natural speech than the adjustment by the pitch frequency alone. Synthesized speech waveform data can be generated.

It is a block diagram which shows the structure of the speech synthesizer 100 which concerns on the 1st Embodiment of this invention. It is a figure which shows the flow of a creation process of a prosodic segment dictionary. It is a figure which shows an example of the calculation method of pitch ratio and intonation ratio. (A) is a figure which shows the structural example of a prosodic segment dictionary, (b) is a figure which shows an example of the pitch frequency data registered into a prosodic segment dictionary. 3 is a flowchart showing an operation process of the speech synthesizer 100. (A) is a figure which shows an example of the pitch frequency series in the speech waveform data of the actual speech sentence containing a fixed phrase, (b) was produced | generated only with the standard pitch frequency with respect to the actual speech sentence of (a). It is a figure which shows an example of a pitch frequency series. It is a figure which shows an example of the standard prosodic information series with respect to the text text containing a fixed phrase. (A) is a figure which shows an example of the prosodic information series which used the prosodic segment pitch frequency for the fixed phrase in sentence text, and performed the adjustment process based on prosodic feature information, (b) It is a figure which shows an example of the prosodic information series produced | generated by the method of the prior art with respect to the same sentence text. It is a block diagram which shows the structure of the speech synthesizer 200 which concerns on the 2nd Embodiment of this invention. 3 is a block diagram illustrating a detailed configuration of a waveform synthesis unit 23 of the speech synthesizer 200. FIG. It is a figure which shows the structural example of a large waveform segment dictionary. 4 is a flowchart showing an operation process of the speech synthesizer 200. (A) is a figure which shows the structure of the original sound data with respect to an example sentence, (b) is a figure which shows the example of a synthesis | combination in case the phonetic environment of the preceding mora of a registration phrase corresponds with the example sentence of (a). (C) is a figure which shows the example of a synthesis | combination when the phonetic environment of the preceding mora of a registration phrase does not correspond with the example sentence of (a). It is a block diagram which shows the structure of the speech synthesizer 300 which concerns on the 3rd Embodiment of this invention. 3 is a flowchart showing an operation process of the speech synthesizer 300. 7 is a flowchart showing a process of generating synthesized speech waveform data in a waveform synthesis unit 23. (A) is a figure which shows the synthetic | combination speech waveform produced | generated using the prior art, when the phonetic environment of the preceding mora of a registration phrase is inconsistent, (b) is the phonetic environment of the preceding mora of a registration phrase. It is a figure which shows the synthetic | combination audio | voice waveform produced | generated using this invention in the case of mismatching.

Explanation of symbols

10, 20 Text analysis unit 11, 21 Standard prosody dictionary 12 Prosody segment dictionary 13 Prosody generation unit 13a, 22 Standard prosody generation unit 13b Registered phrase matching unit 13c Prosody segment selection unit 13d Prosody replacement shaping unit 14, 23 Waveform synthesis unit 23a Segment selection unit 23b Segment connection unit 23c Synthesizer 15, 24 Standard waveform segment dictionary 25 Large waveform segment dictionary 100-300 Speech synthesizer

Claims (15)

A prosodic segment dictionary creation method for creating a prosodic segment dictionary used for speech synthesis,
A first prosodic information extraction step of extracting first prosodic information that is prosodic information of a speech waveform data portion of a predetermined phrase included in the speech waveform data from speech waveform data corresponding to an utterance sentence uttered by a predetermined speaker; ,
A second prosodic information extracting step for extracting second prosodic information as prosody information from at least one of the speech waveform data parts corresponding to a predetermined prosodic unit preceding and following the speech waveform data part of the predetermined phrase;
A third prosodic information generation step of generating third prosodic information that is prosodic information corresponding to the predetermined prosodic unit with reference to a standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit;
Based on the second prosodic information extracted in the second prosodic information extracting step and the third prosodic information generated in the third prosodic information generating step, characteristic features of the second prosodic information and the third prosodic information are characterized. Prosodic feature information generation step for generating prosodic feature information indicating the difference,
A prosodic segment dictionary creating step of creating a prosodic segment dictionary based on the first prosodic information for each predetermined phrase extracted in the first prosodic information extracting step and the prosodic feature information generated in the prosodic feature information generating step; The prosodic segment dictionary creation method characterized by including.   The prosody segment dictionary creation method according to claim 1, wherein the predetermined phrase includes a phrase having a relatively high appearance frequency. The second prosodic information and the third prosodic information include pitch frequency information,
3. The prosodic element according to claim 1, wherein the prosodic feature information includes information indicating a ratio between a pitch frequency indicated by the second prosodic information and a pitch frequency indicated by the third prosodic information. Single dictionary creation method. The second prosodic information and the third prosodic information include pitch frequency information,
The prosodic feature information includes information indicating a ratio between the intonation size obtained from the pitch frequency difference indicated by the second prosodic information and the intonation size obtained from the pitch frequency difference indicated by the third prosodic information. The prosodic segment dictionary creation method according to any one of claims 1 to 3. A prosodic segment dictionary creation program for creating a prosodic segment dictionary used for speech synthesis,
A first prosodic information extraction step of extracting first prosodic information that is prosodic information of a speech waveform data portion of a predetermined phrase included in the speech waveform data from speech waveform data corresponding to an utterance sentence uttered by a predetermined speaker; ,
A second prosodic information extracting step for extracting second prosodic information as prosody information from at least one of the speech waveform data parts corresponding to a predetermined prosodic unit preceding and following the speech waveform data part of the predetermined phrase;
A third prosodic information generation step of generating third prosodic information that is prosodic information corresponding to the predetermined prosodic unit with reference to a standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit;
A characteristic of the second prosodic information and the third prosodic information based on the second prosodic information extracted in the second prosodic information extracting step and the third prosodic information generated in the third prosodic information generating step Prosodic feature information generating step for generating prosodic feature information indicating a difference,
A prosodic segment dictionary creating step for creating a prosodic segment dictionary based on the first prosodic information for each predetermined phrase extracted in the first prosodic information extracting step and the prosodic feature information generated in the prosodic feature information generating step; A prosodic segment dictionary creating program characterized by including a program for causing a computer to execute a process comprising: A speech synthesizer that generates synthesized speech waveform data corresponding to sentence text,
The prosodic segment dictionary creation method according to any one of claims 1 to 4 or the prosodic segment dictionary creation program according to claim 5 includes the first prosodic information and the prosodic feature information. A prosodic segment dictionary consisting of
Standard prosodic dictionary composed of standard prosodic information that is prosodic information for each voice unit;
A segment dictionary comprising spectral information for each speech unit and excitation source information for each speech unit;
Text analysis means for performing accent analysis and morphological analysis on the sentence text;
Prosodic information sequence generation means for generating a prosodic information sequence corresponding to the sentence text based on the analysis result of the text analysis means and the standard prosodic information of the standard prosodic dictionary;
When a phrase corresponding to the first prosodic information included in the prosodic segment dictionary is included in a phrase corresponding to the sentence text, the prosodic information series portion of the phrase is converted to the first prosodic information. A changing means for changing to the prosodic information sequence part generated based on
The prosodic segment corresponding to the changed prosodic information sequence portion with respect to the prosodic information sequence portion corresponding to at least one of the predetermined prosodic units preceding and following the phrase portion changed by the changing means Prosody information adjusting means for performing predetermined adjustment processing based on the prosodic feature information of the dictionary ;
Voice waveform generation means for generating synthesized voice waveform data corresponding to the sentence text based on the prosodic information series and the spectrum information and the excitation source information of the segment dictionary. Speech synthesizer. A speech synthesizer that generates synthesized speech waveform data corresponding to sentence text,
A first segment dictionary comprising spectral information for each speech unit and excitation source information for each speech unit;
Extracted from speech waveform data corresponding to the utterance sentence uttered by the predetermined speaker, the spectrum information for each predetermined phrase, the excitation source information for each predetermined phrase, and the predetermined preceding and following each predetermined phrase in the speech waveform data A second segment dictionary comprising information on the phonetic environment of
Standard prosodic dictionary composed of standard prosodic information that is prosodic information for each voice unit;
Text analysis means for performing accent analysis and morphological analysis on the sentence text;
Prosodic information sequence generation means for generating a prosodic information sequence corresponding to the sentence text based on the analysis result of the text analysis means and the standard prosodic information of the standard prosodic dictionary;
When the phrase corresponding to the sentence text includes a phrase that matches the phrase corresponding to the spectrum information and excitation source information of the second segment dictionary, the spectrum information and the excitation source information are A predetermined phoneme environment of the second segment dictionary preceding and succeeding the matching phrase in the speech waveform data at the time of extraction, and a predetermined preceding and succeeding a portion corresponding to the matching phrase in the prosodic information sequence Determining means for determining whether or not the phonological environment of
Based on the prosodic information series, the determination result of the determination means, and the spectrum information and the excitation source information of the first and second segment dictionaries, synthetic speech waveform data corresponding to the sentence text is generated. Voice waveform generation means,
The speech waveform generation means, when the determination means determines that they do not match, the synthesized speech waveform data generated from the spectrum information and excitation source information corresponding to the matching phrase of the second segment dictionary Based on the first synthesized speech waveform data excluding the speech segment data part at the side edge portion where the phoneme environments do not match, and the spectrum information and excitation source information of the first segment dictionary for the entire sentence text. Generating synthesized speech waveform data corresponding to the sentence text, which is synthesized from second synthesized speech waveform data obtained by excluding a portion corresponding to the first synthesized speech waveform data from the generated synthesized speech waveform data; A featured voice synthesizer. A speech synthesizer that generates synthesized speech waveform data corresponding to sentence text,
The prosodic segment dictionary creation method according to any one of claims 1 to 4 or the prosodic segment dictionary creation program according to claim 5 includes the first prosodic information and the prosodic feature information. A prosodic segment dictionary consisting of
Standard prosodic dictionary composed of standard prosodic information that is prosodic information for each voice unit;
A first segment dictionary comprising spectral information for each speech unit and excitation source information for each speech unit;
Extracted from speech waveform data corresponding to an utterance sentence uttered by the predetermined speaker, spectrum information for each predetermined phrase, excitation source information for each predetermined phrase, and preceding and following each predetermined phrase in the speech waveform data A second segment dictionary including information on a predetermined phonetic environment ;
Text analysis means for performing accent analysis and morphological analysis on the sentence text;
Prosodic information sequence generation means for generating a prosodic information sequence corresponding to the sentence text based on the analysis result of the text analysis means and the standard prosodic information of the standard prosodic dictionary;
When a phrase corresponding to the first prosodic information included in the prosodic segment dictionary is included in a phrase corresponding to the sentence text, the prosodic information series portion of the phrase is converted to the first prosodic information. A changing means for changing to the prosodic information sequence part generated based on
The prosodic segment corresponding to the changed prosodic information sequence portion with respect to the prosodic information sequence portion corresponding to at least one of the predetermined prosodic units preceding and following the phrase portion changed by the changing means Prosody information adjusting means for performing predetermined adjustment processing based on the prosodic feature information of the dictionary ;
When the phrase corresponding to the sentence text includes a phrase that matches the phrase corresponding to the spectrum information and excitation source information of the second segment dictionary, the spectrum information and the excitation source information are A predetermined phoneme environment of the second segment dictionary preceding and succeeding the matching phrase in the speech waveform data at the time of extraction, and a predetermined preceding and succeeding a portion corresponding to the matching phrase in the prosodic information sequence Determining means for determining whether or not the phonological environment of
Based on the prosodic information series, the determination result of the determination means, and the spectrum information and the excitation source information of the first and second segment dictionaries, synthetic speech waveform data corresponding to the sentence text is generated. Voice waveform generation means,
The speech waveform generation means, when the determination means determines that they do not match, the synthesized speech waveform data generated from the spectrum information and excitation source information corresponding to the matching phrase of the second segment dictionary The first synthesized speech waveform data excluding the non-coincident side speech segment data part, and the synthesis generated based on the spectral information and excitation source information of the first segment dictionary for the entire sentence text Generating synthesized speech waveform data corresponding to the sentence text, synthesized from speech waveform data and second synthesized speech waveform data excluding a portion corresponding to the first synthesized speech waveform data; A speech synthesizer. When the prosody segment dictionary includes a plurality of first prosodic information corresponding to the same phrase as the phrase constituting the sentence text, the changing means converts the prosodic information series portion of the phrase into the prosody. claim and changes the generated prosodic information sequence portion based on the connectivity is the best first prosody information with the preceding and succeeding prosodic information sequence portion corresponding phrases in the information sequence portion 6 or claim 8 The speech synthesizer described. A speech synthesis program for generating synthesized speech waveform data corresponding to sentence text,
A text analysis step for performing accent analysis and morphological analysis on the sentence text;
Prosodic information for generating prosodic information corresponding to the sentence text based on the analysis result in the text analyzing step and the standard prosodic information included in the standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit A sequence generation step;
In the phrase corresponding to the sentence text, the prosodic segment dictionary creation method according to any one of claims 1 to 4 or the prosodic segment dictionary creation program according to claim 5 , When a phrase corresponding to the first prosodic information included in the prosodic segment dictionary including the first prosodic information and the prosodic feature information is included, the prosodic information sequence portion of the phrase is converted to the first prosodic information sequence part. A change step to change to the prosodic information sequence part generated based on the information;
The prosodic segment corresponding to the changed prosodic information sequence portion with respect to the prosodic information sequence portion corresponding to at least one of the predetermined prosodic units preceding and following the phrase portion changed in the changing step A prosodic information adjustment step for performing a predetermined adjustment process based on the prosodic feature information of the dictionary ;
Corresponding to the text text based on the spectrum information and the excitation source information of the segment dictionary including the prosodic information series, the spectrum information for each speech unit and the excitation source information for each speech unit A speech synthesis program comprising: a program for causing a computer to execute a process comprising a speech waveform generation step for generating synthesized speech waveform data. A speech synthesis program for generating synthesized speech waveform data corresponding to sentence text,
A text analysis step for performing accent analysis and morphological analysis on the sentence text;
Prosodic information for generating a prosodic information sequence corresponding to the sentence text based on the analysis result in the text analysis step and the standard prosodic information included in the standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit A sequence generation step;
Spectral information for each predetermined phrase, excitation source information for each predetermined phrase, and the speech waveform data extracted from speech waveform data corresponding to an utterance sentence uttered by a predetermined speaker in the phrase corresponding to the sentence text Including a phrase corresponding to the spectrum information and excitation source information of the second segment dictionary including information of a predetermined phoneme environment preceding and following each of the predetermined phrases in Corresponding to a predetermined phoneme environment of the second segment dictionary preceding and following the matching phrase in the speech data when extracting the spectrum information and the excitation source information, and the matching phrase in the prosodic information series A determination step for determining whether or not a predetermined phonological environment preceding and succeeding the portion to be matched matches;
It said prosodic information sequence, the determination result, sentences first segment dictionary and said comprising an excitation source information of each of the speech segment and the spectral information for each speech segment to the entire text in the determination step A program for causing a computer to execute a process including a speech waveform generation step of generating synthesized speech waveform data corresponding to the sentence text based on the spectrum information and the excitation source information of the second segment dictionary ,
In the speech waveform generation step, when it is determined in the determination step that they do not match, the synthesized speech waveform data generated from the spectrum information and the excitation source information corresponding to the matching phrase of the second segment dictionary From the first synthesized speech waveform data excluding the speech segment data portion at the non-matching side edge, and the synthesized speech waveform data generated based on the spectrum information and excitation source information of the first segment dictionary , the first A speech synthesis program for generating synthesized speech waveform data corresponding to the sentence text, which is synthesized with second synthesized speech waveform data excluding a portion corresponding to one synthesized speech waveform data. A speech synthesis program for generating synthesized speech waveform data corresponding to sentence text,
A text analysis step for performing accent analysis and morphological analysis on the sentence text;
Prosodic information for generating a prosodic information sequence corresponding to the sentence text based on the analysis result in the text analysis step and the standard prosodic information included in the standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit A sequence generation step;
In the phrase corresponding to the sentence text, the prosodic segment dictionary creation method according to any one of claims 1 to 4 or the prosodic segment dictionary creation program according to claim 5 , When a phrase corresponding to the first prosodic information included in the prosodic segment dictionary including the first prosodic information and the prosodic feature information is included, the prosodic information sequence portion of the phrase is converted to the first prosodic information sequence part. A change step to change to the prosodic information sequence part generated based on the information;
The prosodic segment corresponding to the changed prosodic information sequence portion with respect to the prosodic information sequence portion corresponding to at least one of the predetermined prosodic units preceding and following the phrase portion changed in the changing step A prosodic information adjustment step for performing predetermined adjustment processing based on the prosodic feature information of the dictionary ;
Spectral information for each predetermined phrase, excitation source information for each predetermined phrase, and the speech waveform data extracted from speech waveform data corresponding to an utterance sentence uttered by a predetermined speaker in the phrase corresponding to the sentence text Including a phrase corresponding to the spectrum information and excitation source information of the second segment dictionary including information of a predetermined phoneme environment preceding and following each of the predetermined phrases in and a predetermined phonetic environment preceding and subsequent to the phrase to the match in the audio data at the time of extracting the spectral information and the excitation source information, preceding and succeeding the portions corresponding to the phrases that the match in the prosodic information sequence A determination step for determining whether or not the predetermined phoneme environment of the second segment dictionary matches;
It said prosodic information sequence, the determination result, sentences first segment dictionary and said comprising an excitation source information of each of the speech segment and the spectral information for each speech segment to the entire text in the determination step A program for causing a computer to execute a process including a speech waveform generation step of generating synthesized speech waveform data corresponding to the sentence text based on the spectrum information and the excitation source information of the second segment dictionary ,
In the speech waveform generation step, when it is determined in the determination step that they do not match, the synthesized speech waveform data generated from the spectrum information and the excitation source information corresponding to the matching phrase of the second segment dictionary From the first synthesized speech waveform data excluding the speech segment data portion at the non-matching side edge, and the synthesized speech waveform data generated based on the spectrum information and excitation source information of the first segment dictionary , the first A speech synthesis program for generating synthesized speech waveform data corresponding to the sentence text, which is synthesized with second synthesized speech waveform data excluding a portion corresponding to one synthesized speech waveform data. A speech synthesis method for generating synthesized speech waveform data corresponding to sentence text,
A text analysis step for performing accent analysis and morphological analysis on the sentence text;
Prosodic information for generating prosodic information corresponding to the sentence text based on the analysis result in the text analyzing step and the standard prosodic information included in the standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit A sequence generation step;
In the phrase corresponding to the sentence text, the prosodic segment dictionary creation method according to any one of claims 1 to 4 or the prosodic segment dictionary creation program according to claim 5 , When a phrase corresponding to the first prosodic information included in the prosodic segment dictionary including the first prosodic information and the prosodic feature information is included, the prosodic information sequence portion of the phrase is converted to the first prosodic information sequence part. A change step to change to the prosodic information sequence part generated based on the information;
The prosodic segment corresponding to the changed prosodic information sequence portion with respect to the prosodic information sequence portion corresponding to at least one of the predetermined prosodic units preceding and following the phrase portion changed in the changing step A prosodic information adjustment step for performing a predetermined adjustment process based on the prosodic feature information of the dictionary ;
Corresponding to the text text based on the spectrum information and the excitation source information of the segment dictionary including the prosodic information series, the spectrum information for each speech unit and the excitation source information for each speech unit A speech waveform generation step of generating the synthesized speech waveform data. A speech synthesis method for generating synthesized speech waveform data corresponding to sentence text,
A text analysis step for performing accent analysis and morphological analysis on the sentence text;
Prosodic information for generating a prosodic information sequence corresponding to the sentence text based on the analysis result in the text analysis step and the standard prosodic information included in the standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit A sequence generation step;
Spectral information for each predetermined phrase, excitation source information for each predetermined phrase, and the speech waveform data extracted from speech waveform data corresponding to an utterance sentence uttered by a predetermined speaker in the phrase corresponding to the sentence text Including a phrase corresponding to the spectrum information and excitation source information of the second segment dictionary including information of a predetermined phoneme environment preceding and following each of the predetermined phrases in and a predetermined phonetic environment preceding and subsequent to the phrase to the match in the audio data at the time of extracting the spectral information and the excitation source information, preceding and succeeding the portions corresponding to the phrases that the match in the prosodic information sequence A determination step for determining whether or not the predetermined phoneme environment of the second segment dictionary matches;
It said prosodic information sequence, the determination result, sentences first segment dictionary and said comprising an excitation source information of each of the speech segment and the spectral information for each speech segment to the entire text in the determination step A speech waveform generation step of generating synthesized speech waveform data corresponding to the sentence text, based on the spectrum information and the excitation source information of the second unit dictionary,
In the speech waveform generation step, when it is determined in the determination step that they do not match, the synthesized speech waveform data generated from the spectrum information and the excitation source information corresponding to the matching phrase of the second segment dictionary From the first synthesized speech waveform data excluding the speech segment data portion at the non-matching side edge, and the synthesized speech waveform data generated based on the spectrum information and excitation source information of the first segment dictionary , the first A speech synthesis method comprising: generating synthesized speech waveform data corresponding to the sentence text, which is synthesized with second synthesized speech waveform data excluding a portion corresponding to one synthesized speech waveform data. A speech synthesis method for generating synthesized speech waveform data corresponding to sentence text,
A text analysis step for performing accent analysis and morphological analysis on the sentence text;
Prosodic information for generating a prosodic information sequence corresponding to the sentence text based on the analysis result in the text analysis step and the standard prosodic information included in the standard prosodic dictionary composed of standard prosodic information that is prosodic information for each speech unit A sequence generation step;
In the phrase corresponding to the sentence text, the prosodic segment dictionary creation method according to any one of claims 1 to 4 or the prosodic segment dictionary creation program according to claim 5 , When a phrase corresponding to the first prosodic information included in the prosodic segment dictionary including the first prosodic information and the prosodic feature information is included, the prosodic information sequence portion of the phrase is converted to the first prosodic information sequence part. A change step to change to the prosodic information sequence part generated based on the information;
The prosodic segment corresponding to the changed prosodic information sequence portion with respect to the prosodic information sequence portion corresponding to at least one of the predetermined prosodic units preceding and following the phrase portion changed in the changing step A prosodic information adjustment step for performing a predetermined adjustment process based on the prosodic feature information of the dictionary ;
Spectral information for each predetermined phrase, excitation source information for each predetermined phrase, and the speech waveform data extracted from speech waveform data corresponding to an utterance sentence uttered by a predetermined speaker in the phrase corresponding to the sentence text Including a phrase corresponding to the spectrum information and excitation source information of the second segment dictionary including information of a predetermined phoneme environment preceding and following each of the predetermined phrases in and a predetermined phonetic environment preceding and subsequent to the phrase to the match in the audio data at the time of extracting the spectral information and the excitation source information, preceding and succeeding the portions corresponding to the phrases that the match in the prosodic information sequence A determination step for determining whether or not the predetermined phoneme environment of the second segment dictionary matches;
It said prosodic information sequence, the determination result, sentences first segment dictionary and said comprising an excitation source information of each of the speech segment and the spectral information for each speech segment to the entire text in the determination step A speech waveform generation step of generating synthesized speech waveform data corresponding to the sentence text, based on the spectrum information and the excitation source information of the second unit dictionary,
In the speech waveform generation step, when it is determined in the determination step that they do not match, the synthesized speech waveform data generated from the spectrum information and the excitation source information corresponding to the matching phrase of the second segment dictionary From the first synthesized speech waveform data excluding the speech segment data portion at the non-matching side edge, and the synthesized speech waveform data generated based on the spectrum information and excitation source information of the first segment dictionary , the first A speech synthesis method comprising: synthesizing second synthesized speech waveform data excluding a portion corresponding to one synthesized speech waveform data to generate synthesized speech waveform data corresponding to the sentence text.

Images