JPH11249678A - Voice synthesizer and its text analytic method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice synthesizer and its text analytic method making a synthetic voice generated by the supplied data the natural synthetic voice easy to hear and outputting it. SOLUTION: The voice synthesizer 10 is constituted so that a division candidate generation part 460 of a vocal sound processing part 40 divides an output of a language processing part 30 based on a (first) rule of ICRLB (immediate constituent with recursively left-branching structure) division, and judges (second rule) whether this divided rhythm phrase is the number of prescribed beats or below, and makes a combination of the rhythm phrases generated according to positions dividing this rhythm phrase when this rhythm phrase is re-divided a division candidate, and a valuation value showing division suitability is calculated by using information of a voice expression incorporated in the division candidate obtained in a division candidate selection part 462 and a parameter supplied from a parameter storage part 464, and further, and the division candidate satisfying a (third) rule, selecting the minimum valuation value is selected by using this valuation value by the division candidate selection part 462, and the selected division candidate becomes the rhythm phrase divided to an optimum length.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizer for performing various analyzes on a supplied sentence and artificially generating a synthesized speech from an intermediate language generated according to a prosody rule and the like. The text analysis method relates to a speech synthesis method, in particular, which analyzes a morpheme and a syntax of a supplied Japanese sentence, and then generates a synthesized speech based on an intermediate language generated by phonological processing performed on the result. The present invention relates to an apparatus and a text analysis method suitable for application to the speech synthesis apparatus.

[0002]

2. Description of the Related Art To artificially synthesize speech, there are a recording / reproducing method in which sounds recorded in advance are combined and reproduced in order to form a waveform, and a pure synthesizing method in which sounds are artificially synthesized purely. For the control of waveform formation in this recording / reproducing method, there are various methods such as an edit control method using a combination of information prepared in advance, or a rule control method for purely artificially generating a control signal. Among these methods, recently, e.g., e-mail reading and weather forecasting services,
2. Description of the Related Art A speech synthesizer that provides a high-quality synthesized speech by applying a rule control method to various fields such as news such as the results of professional baseball has attracted attention.

In order to obtain a synthesized speech with high quality, it is necessary to construct a prosodic rule that can accurately reflect linguistic information such as meaning, syntactic structure, and discourse structure in Japanese and generate a natural prosody. Must. An example of a method for generating a synthesized speech in Japanese by applying such a rule control method is disclosed in the Journal of the Acoustical Society of Japan, Vol. 50, No. 6 (“Prosody rules for synthesis of Japanese sentence speech”). . The method disclosed herein performs a morphological analysis and a syntactic analysis on a supplied sentence to generate a prosodic phrase, and when the generated prosodic phrase is too large, divides the prosodic phrase almost equally. Thus, the intermediate language is generated so that the sentence can be read with smooth prosody while considering the semantic structure of the sentence.

The speech synthesizer generates a control parameter based on the generated intermediate language, and synthesizes and outputs a speech corresponding to the control parameter.

[0005]

By the way, when the generated prosody phrase is too large as described above, by applying rules, the generated prosody phrase may be too finely divided, or may be read in a manner contrary to the original prosody. Prosodic phrase division may be performed. When such prosodic phrases are divided, the output synthesized speech may sound unnatural.

The present invention solves the above-mentioned drawbacks of the prior art, and provides a speech synthesis apparatus and a text analysis method capable of outputting a synthesized speech generated by supplied data as a more natural and easily synthesized speech. The purpose is to provide.

[0007]

According to the present invention, in order to solve the above-mentioned problems, a language level feature is determined based on a morpheme included in a sentence supplied as information and a language used for processing the syntax of the sentence. A linguistic analysis means for analyzing, and a phonological analysis for performing an analysis based on features of a speech language level on an output of the linguistic analysis means and generating an intermediate language serving as a command for speech synthesis based on the obtained analysis result. Means,
A control parameter generating means for generating a control parameter corresponding to an output of the phoneme analyzing means; and a voice signal generating means for synthesizing a voice signal based on an output of the control parameter generating means. In the speech synthesizer for synthesizing, the first rule for dividing the output of the language analysis means according to the modification relation of the prosodic phrase in which a plurality of prosodic words which are a chain of morphemes are collected, and the prosody phrase obtained by this division. Candidate generation for generating, as a division candidate, a combination of prosodic phrases obtained at positions separating the prosodic phrases when subdividing the prosodic phrases in accordance with a second rule for determining whether is smaller than a predetermined size. Means for calculating an evaluation value for evaluating the validity of the division using the information of the speech expression included in the division candidate generated by the division candidate generation means,
A division candidate selection unit for selecting a division candidate according to a third rule for selecting a division candidate having the smallest evaluation value, and a parameter storage unit for storing parameters used for calculating an evaluation value in the division candidate selection unit. It is characterized in that it is included in the analysis means.

Here, the division candidate generating means is a first rule dividing means for dividing the output of the linguistic analysis means using a first rule, and is divided at the output of the first rule processing means. A second rule processing means for judging the length by the second rule, and a plurality of division candidates generated by subdividing the prosodic phrase output from the first rule division means in accordance with the judgment of the division length judging means It is desirable to include subdivision means for performing the division.

[0009] The parameter storage means may store a first weighting factor based on information of speech expression in the division candidate and a second weighting factor multiplied by the average number of prosodic words according to the number of divisions of the division candidate. . As a result, it is possible to generate an intermediate language capable of accurately expressing the designation of emphasis / suppression of the generated prosodic phrase.

A divided example storing means for storing an example of a set position of a prosodic phrase which satisfies a rule, and an example in which a supplied sentence is optimally divided from examples stored in the divided example storing means. It is preferable to include a division example search unit that performs the division. With such a configuration, it is possible to output an intermediate language that outputs an optimum synthesized speech according to an actual example.

[0011] The divided example retrieval means includes a search for an example stored in the divided example storage means, the emphasis information, the prosodic word element, the number of beats of the prosodic word element, the total number of beats, the type of prosodic word element dependency, and the accent. Preferably, a command and / or part of speech type is used. It is possible to improve the accuracy of matching of examples by searching for examples and shorten the search time.

[0012] It is preferable that the linguistic analysis means sets emphasis information rules and includes emphasis information setting means for storing the set emphasis information.

In the speech synthesis apparatus according to the present invention, the output of the language analysis means is divided by the division candidate generation means based on the first rule, and the divided prosody phrase is determined by the second rule to determine the prosody phrase. When subdividing, a combination of prosodic phrases generated based on the positions separating the prosodic phrases is used as a division candidate, and evaluation is performed to evaluate the validity of the division using the information of the speech expression included in the division candidate obtained by the division candidate selection means. The value is calculated using the parameters supplied from the parameter storage means, and further, by using this evaluation value, the division candidate selection means selects a division candidate satisfying the third rule. Since the prosodic phrase is divided into various lengths, the phonological analysis means can perform intermediate processing in a speech synthesis device that generates an intermediate language corresponding to the selected division candidate.

Further, the text analysis method of the speech synthesis apparatus according to the present invention analyzes the features at the language level based on a morpheme included in a sentence supplied as information and a language used for processing the syntax of the sentence. Based on the analysis result, the analysis based on the features of the spoken language level of the prosody of the sentence is performed, and based on the obtained analysis result, an intermediate language serving as a command for speech synthesis is generated, and according to the generated intermediate language, In the text analysis method of the speech synthesizer for artificially synthesizing a speech corresponding to the control parameter after generating the control parameter according to the modification relation of the prosodic phrase in which a plurality of prosodic words which are a chain of morphemes are grouped. A rule dividing step of dividing an analysis result for a sentence using a first rule of dividing information by using a first rule, wherein the size of a prosodic phrase is smaller than a predetermined size. Is determined as a second rule, a division length determining step of determining the result of the rule dividing step by the second rule, and a prosody phrase obtained by re-dividing the information in accordance with the determination result of the division length determining step And a parameter storage for storing parameters used for evaluating the validity of the division based on the information of the speech expression included in the division candidates obtained in the division candidate generation step. A step, an evaluation value calculating step of calculating an evaluation value from information of the speech expression included in the division candidate generated in the division candidate generating step and the parameter, and a division candidate indicating a minimum among the evaluation values obtained in the evaluation value calculating step Is a third rule, and a division candidate selection step of selecting a division candidate based on the third rule is performed, and the supplied sentence is analyzed.

Here, when the beat length included in the prosodic phrase is larger than the preset re-divided beat length indicating the re-division, the prosodic phrase is generated by subdividing the prosodic phrase, and It is preferable to enter a symbol for instructing a tone, which is one of the features of the speech language level. By defining the division candidates in this way, it is possible to avoid consideration of a long prosodic phrase, for example, to divide the prosodic phrase too finely.

The evaluation value is obtained by calculating a sum of absolute values of differences between a beat length of a section obtained by dividing the prosodic phrase and a beat length obtained by equally dividing the beat length of the entire prosodic phrase; A feature amount sum calculating step of calculating a sum of beat lengths of sections obtained by dividing the prosodic phrase according to the presence of a feature amount represented by information of a speech expression included in a feature of a speech language level in the phrase;
A weight calculating step of calculating a weight coefficient based on information of a speech expression included in the prosodic word element of the division candidate, and multiplying a result of the feature amount sum calculating step by the weight coefficient obtained in the weight calculating step; A weighted feature amount sum calculating step of calculating the sum of the multiplication results, and a product calculating step of calculating the product of the average beat length after dividing the prosodic phrase and the beat length obtained by equally dividing the entire prosodic phrase. And adding the result of the error sum calculation step and the result of the weighted feature amount sum calculation step, and subtracting the result of the product calculation step from the result obtained by the addition to calculate the evaluation value of the target division candidate. Advantageously, it is calculated using a value calculation step.

It is preferable that the information of the above-mentioned speech expression is classified into values for emphasizing and suppressing prosodic words and for excluding both, and values are set. The feature amount is obtained based on the set value.

As a criterion for classification, emphasis is classified by proper nouns or numerals included in the center of the prosodic phrase represented by the division candidate, and suppression is included in the center of the prosodic phrase represented by the division candidate. It is desirable to classify this prosodic phrase by being located at the beginning of a formal noun, verb, or sentence, and by placing a particle at the end of the prosodic phrase, and assigning preset values to emphasis and suppression.

The emphasis is classified by including a proper noun or a numeral in the prosody of the division candidate, and the suppression is located at the formal noun, the verb or the beginning of the prosody of the division candidate, and the final part of the prosody. It is desirable to classify this prosodic clause by including the following, and to assign predetermined values to emphasis and suppression.

A sentence division candidate to which the text analysis method is applied stores information to be supplied and / or information added to an appropriately divided example as a search key in advance, and a division length determination step includes: An output destination is selected according to the judgment result of the supplied prosodic phrase, and thereafter, an example corresponding to the division of the prosodic phrase supplied at the selected output destination is searched, and the prosodic phrase is selected according to the search result. It is preferable to include an actual example search step of dividing and giving various instructions. By this search, accurate sentence analysis can be performed.

In the example storing step, it is desirable to store various examples in a learning manner in advance. With this memory, it is possible to shorten the period of gaining experience and to cope with it more widely.

Preferably, the parameters are obtained by statistical analysis or multivariate analysis. Thereby, the parameter can be changed semi-automatically.

The language to which this text analysis method is applied is
Advantageously, it is Japanese. This facilitates generation of an intermediate language based on Japanese sentence analysis having ambiguous expressions and syntax.

The text analysis method for speech synthesis according to the present invention comprises:
The analysis result for the sentence is processed in the order of the first rule of the rule division step and the second rule of the division length determination step, and a combination obtained when information is redivided according to the result is set as a division candidate. An evaluation value of division validity is calculated based on the parameters obtained in the parameter storage step based on the information of the speech expression included in the division candidate and the information of the speech expression included in the division candidate (evaluation value calculation step). By selecting a division candidate satisfying the third rule in the candidate selection step, even if the prosodic phrase of the supplied sentence is further subdivided and a phonological analysis is performed, an optimal division candidate is selected to cope with this division. It is possible to generate an intermediate language.

[0025]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing an embodiment of a speech synthesizing apparatus according to the present invention;

The speech synthesizing apparatus of the present invention analyzes the morpheme and syntax of the supplied Japanese sentence, and then generates control parameters based on the intermediate language generated by phonemic processing performed on the result. This is a device that generates and outputs synthesized speech based on the control parameters. This speech synthesizer and its text analysis method are shown in FIGS.
This will be described with reference to FIG. As shown in FIG. 1, the speech synthesizer 10 includes a data input unit 20, a language analysis unit 30, a phoneme processing unit 40, a control parameter generation unit 50, and a speech signal generation unit 60.

The data input section 20 converts various forms of text into a data format that can be processed by the speech synthesizer 10 and inputs the data. This sentence has various forms, such as a form of a handwritten manuscript and a form recorded on a recording medium in a predetermined data format created by sentence creation software such as a word processor. In the case of a handwritten document, the data input unit 20 is provided with a device for optically reading a sentence related to the paper. When reading stored data from a floppy disk, the data input section 20
, A disk driver is disposed inside the device as a data reading device.

As shown in FIG. 1, the language analysis unit 30 includes a morphological analysis unit 32, a syntax analysis unit 34, and an emphasis information setting unit 36. In the language analysis unit 30, data supplied from the data input unit 20, that is, digital data of an input sentence is supplied to the morphological analysis unit 32. The morphological analysis unit 32 includes a word division analysis unit 32a and a word dictionary unit 32b shown in FIG. A morpheme is the smallest form that has meaning. The word division analysis unit 32a stores a processing program for performing word division, and performs a search process so as to match words included in an input sentence with words in the word dictionary unit 32b. The morphological analysis unit 32 includes a word division analysis unit 32a
The input sentence is divided into words while searching the word dictionary section 32b according to the program stored in.

Here, terms of language analysis will be described. Sentences are classified into processing units, prosodic units, syntactic units, and intermediate units between them.
(see (a)). Prosodic units include prosodic words, prosodic phrases, prosodic clauses, and prosodic sentences. A prosodic word is a chain of phonemes (that is, minimum units that can be discriminated as sounds) that correspond to one accent component and indicate a certain accent type. A prosodic phrase is a chain of prosodic words corresponding to one phrase component. A prosodic clause is a chain of prosodic phrases separated by pauses, and indicates that the phrase component has not been reset (that is, a negative phrase command has been generated) in the last part of the prosodic phrase. Finally, the prosodic sentence is a chain of prosodic phrases, and the phrase component is not reset in the last part as in the definition of the prosodic phrase. For example,
The sentence "The cyclone is south of the Izu Peninsula and is showering in eastern Japan" is classified according to these definitions as shown in Figure 3 (b).

The syntactic units include clauses in order of size,
(Phrases), clauses and sentences. In particular, a clause is a predicate that does not modify other terms (verb predicate, adjective predicate, noun predicate, etc.)
Is defined as a chain consisting of all the phrases that directly and indirectly modify the predicate. However, in the sentence "It is a substation that supplies electricity to 500 households in the vicinity", the description of "to do" qualifies "substation" as a union, so in this sentence, Are not categorized into sections.

The intermediate units are prosodic words, ICRLB
There is. A prosodic word is a chain of morphemes that is not divided into a plurality of prosodic words by syntactic conditions and emphasis / suppression of reading. ICRLB (Immediate Constituent with Rec
Ursively Left-Branching Structure) is a sequence of prosodic words separated by right branching boundaries and including only left branching boundaries. In the parse tree, a left branch boundary is a boundary between prosodic words having a modification relationship, and a right branch boundary is defined as a boundary between prosody words having no modification relationship.
This is called the ICRLB boundary. ICRLB may be a single word, and if the parse tree is divided at the right branch boundary, one sentence
ICRLB chain. The first rule described later is a rule for performing ICRLB division.

According to such a definition, for example, a boundary of a clause can be distinguished from a normal modification relation even if there is a modification relation, so that it is also an ICRLB boundary at the same time. Therefore, the definition of the boundary can be summarized as follows: the node boundary is the boundary between nodes, ICRLB
The boundary is a boundary between ICRLBs, and the enumeration boundary is a word boundary after a reading point in an expression in which nouns are arranged in parallel with reading points, for example, “judicial, legislative, administrative”.

The syntax analyzer 34 that performs an analysis process using such a definition receives the morphological analysis result from the morphological analyzer 32 as data and analyzes the syntax tree or the dependency structure of the input sentence. Department. Also, the parser
Numeral 34 determines ICRLB, clauses, parallel relations, and the like from the results of analysis by the above-described functions, and performs syntax analysis, which is a feature at the language level, on input sentences. If the result of this parsing is not a parse tree but a dependency structure, the ICRLB boundary is defined based on the modification relation defined above, and if there is a compound word containing multiple accent nuclei, the immediately preceding compound word is added. Defined as ICRLB boundary.

As shown in FIG. 2, the emphasis information setting unit 36 sets an emphasis / suppression emphasis information for a word or a phrase based on the analysis result of the morphological analysis unit 32 and the analysis result of the syntax analysis unit 34. And a standard storage unit 36b for storing rules of emphasis information, which is a reference for setting to be supplied to the information setting unit 36a. The information setting section 36a sets emphasis information in the input data by comparing it with the rules supplied from the standard storage section 36b.

Here, the emphasis information is information used for setting an accent and a part of a phrase command, and is one of information of voice expression. The emphasis information is classified into three types of information, + emph for emphasis of information, -emph for suppression of information, and 0emph when not classified as both emphasis and suppression. As an example of setting emphasis information, for example, clauses centered on proper nouns, numbers, etc. should be + emph, and "koto", "mono"
For example, a phrase centered on a verb such as "Naru" or "Iru" or a phrase such as "ha" (participatory) case at the beginning of a sentence is -e
mph.

The language analysis unit 30 outputs to the phonological processing unit 40 the analysis result from the morphological analysis unit 32 and the syntax analysis unit 34 and the emphasis information set from the emphasis information setting unit 36 for the input sentence. The phoneme processing unit 40 generates an intermediate language for speech synthesis based on the output of the language analysis unit 30. As shown in FIG. 1, the phonological processing unit 40 includes a prosodic word generation processing unit 42, an accent command generation unit 44, and a pause / phrase command generation unit 46.
Is provided.

The prosodic word generation unit 42 generates a prosodic word by performing accent combination of words included in a phrase. The prosodic word generation unit 42 also generates a prosodic word by performing accent connection between the phrases as necessary. The prosodic word generation unit 42 also sets phrase emphasis information in the generated prosodic word.

As shown in FIG. 4, an accent command generation unit 44a generates an accent command for instructing how to apply an accent based on the ICRLB information in the output of the language analysis unit 30, as shown in FIG. And a standard storage unit 44b for storing accent command rules that serve as a reference for setting supplied to the accent command setting unit 44a. The accent command setting unit 44a sets the accent command in the result of the linguistic analysis by comparing it with the rule of the accent command supplied from the standard storage unit 44b.

Here, the accent command is a command indicating the size of the accent and the time of the start and the end of the accent. In the generation of the accent command, the processing target range is the range of accent deformation. Accent deformation (accent sandhi) refers to a plurality of continuous prosody according to an accent type such as a flat type of a flat accent, an undulating type with an undulating accent, a syntactic condition or a discourse condition that divides a sentence into boundaries. It is the interaction between prosodic words that affect the accent components of words. As can be seen from such a relationship, the range of accent deformation is ICRLB.

The pause / phrase command generation unit 46 has a function of generating a pause length / phrase command. Pause / phrase command generation unit to demonstrate this function
46 includes a pause / phrase command setting unit 46A for generating a pause command / phrase command shown in FIG. 4 and an ICRLB dividing unit 46B for dividing the output of the language analysis unit 30 into ICRLBs of an optimal length.
And are provided.

Here, the pause command is composed of pause symbols S ₁ , S ₂ ,
Represented by S _3. The pause symbols S ₁ , S ₂ , S ₃ are statements,
Clause and ICRLB. A phrase command is represented by phrase symbols P ₀ , P ₁ , P ₂ , and P ₃ . Phrase symbol
P ₀ , P ₁ , P ₂ , and P ₃ are used for resetting, recovering, adding at the beginning of a phrase component, and adding between ICRLBs and within ICRLB, respectively.

The pause / phrase command setting section 46A includes a command setting section 46a for generating a pause command / phrase command,
There is a standard storage unit 46b for storing rules of a pause command / phrase command as a reference for setting supplied to the command setting unit 46a. The command setting unit 46a is not supplied with the output from the language analysis unit 30 as it is, but in this embodiment, ICRLB
The output from the dividing unit 46B is supplied.

Further, the ICRLB dividing section 46B includes a division candidate generation section 460, a division candidate selection section 462, and a parameter storage section 464. The division candidate generation unit 460 includes:
An output from the language analysis unit 30 is supplied. The division candidate generation unit 460 includes a rule division unit that divides the supplied data (clause) at the boundary of ICRLB (that is, the first rule).
460a and a division for judging whether the output of the rule division unit 460a (the length of the prosodic phrase divided by ICRLB: hereinafter, ICRLB prosodic phrase) is shorter than a preset length (that is, the second rule). There is a length determining unit 460b and a re-dividing unit 460c that re-divides the prosodic phrase divided by the rule dividing unit 460a according to the determination of the division length determining unit 460b to generate a plurality of division candidates.

Here, the ICRLB prosodic phrase is a combination of prosodic phrases obtained by subdividing the prosodic phrase by dividing the prosodic phrase, and this combination is called a division candidate.

When the division length judgment unit 460b satisfies the judgment condition of the second rule when supplied first, it supplies an output to the command setting unit 46a. In the first case, when the determination condition is not satisfied, the command setting unit 46a sets the output destination to the re-division unit 460c. By the way, at the time of the second and subsequent determinations, the command setting unit 46a makes a determination in addition to the determination condition, as well as a condition as to whether or not there is still a combination of division candidates for the divided ICRLB prosody.

The subdivision unit 460c returns the divided data to the subdivision unit 460c to determine the second rule, and repeats the subdivision processing until there are no more combinations of plural types of division candidates. The re-dividing unit 460c selects the second among the divided candidates.
Only the division candidates satisfying the determination condition of the rule are output to the calculation unit 462a of the division candidate selection unit 462.

However, if the information is always finely divided according to the division rule at the boundary of the ICRLB before being supplied to the ICRLB division unit 46c, the rule division unit 460a can omit the arrangement. The division length determination unit 460b determines the divided ICRL
B When the prosodic phrase is longer than a predetermined length, the output is supplied to the re-segmentation unit 460c; otherwise (ie, when the input sentence is first supplied, the second rule is satisfied), only the pause This output is supplied to the command setting section 46a of the phrase command setting section 46A.

Therefore, when the supplied ICRLB prosodic phrase is longer than a predetermined length, the division candidate generating section 460 re-divides the ICRLB prosodic phrase by the built-in subdivision section 460c until the ICRLB prosodic phrase becomes shorter than the predetermined length. While repeating, this output is supplied to the division candidate selection unit 462.

The division candidate selection unit 462 includes a calculation unit 462a that calculates a cost as an evaluation value indicating the validity of division obtained based on the information of the speech expression included in the generated division candidate.
A selection unit 362b that selects a division candidate having the minimum value among the costs calculated by the calculation unit 462a (that is, the third rule).
There is. This cost calculation corresponds to the sum of the errors in the conventional method, and partially includes the sum of the errors. The details will be described later. The division candidate selection unit 462 supplies an output of the division candidate selected from the selection unit 462b to the command setting unit 46a of the pause / phrase command setting unit 46A. The parameter storage unit 464 has a memory for storing parameters used for cost calculation in the calculation unit 462a of the division candidate selection unit 462.

The parameter storage unit 464 also has a function unit (not shown) for changing parameters by applying a method such as statistical analysis or multivariate analysis. When the simplification of the apparatus is taken into consideration, the parameter may be changed outside the apparatus, and the parameter to be changed may be simply supplied to the parameter storage unit 464. In the statistical analysis or the multivariate analysis, a rule for digitizing information of a speech expression is provided in advance, and the tendency of a sentence is evaluated using a numerical value obtained corresponding to the information based on the rule. At this time, the value corresponding to the evaluation is a parameter indicating a tendency due to the subdivision.

The ICRLB division unit 46B selects the minimum cost from the division candidates when the length of the long ICRLB prosodic phrase is reduced to a predetermined length or less, and sets this division candidate in the command setting of the pause / phrase instruction setting unit 46A. It has a function of supplying a pause command / phrase command to the section 46a to make it more accurate and output it. Thus, the ICRLB prosodic phrase can be divided into appropriate ranges. The phonological processing unit 40, after determining the word reading, the accent command, the pause command, and the phrase command, outputs data of an intermediate language (not shown) created by performing processes such as lengthening and prompting to the control parameter generating unit 50. send. By this phonological process, generation of an intermediate language that results in unnatural speech synthesis can be avoided.

The control parameter generation unit 50 includes a phoneme processing unit 40
A control parameter used for speech synthesis is generated based on the data in the intermediate language supplied from. The generated control parameters are supplied to the audio signal generation unit 60. The audio signal generator 60 includes an audio waveform generator 62 and an audio output unit 64. The audio waveform generation unit 62 performs a D / A conversion process on the control parameters supplied from the control parameter generation unit 50, generates an audio waveform, and outputs the generated audio waveform to the audio output unit 64. The sound output unit 64 outputs a sound waveform as information (for example, a sentence) input through a speaker, for example. With such a configuration, the speech composition device 10 generates accurate intermediate language data and outputs a synthesized speech based on the generated data.

Next, the control and operation of the speech synthesizer 10 of this embodiment will be described with reference to the flowcharts of FIGS. 5 to 13 and tables based on various examples. The flowchart of FIG. 5 describes the control of the speech synthesizer 10 and the main operation procedure under the control. When the power is turned on to the speech synthesizer 10, the operation of the speech synthesizer 10 is started, the initialization is performed, and the process proceeds to step S10. In step S10, the sentence supplied via the data input unit 20 of the speech synthesizer 10 is digitized, and the digitized data is temporarily stored in the memory. After this data input, the process proceeds to subroutine SUB1.

In a subroutine SUB1, a language analysis process is performed by the language analysis unit 30 on the supplied data. The language analysis processing performed here includes morphological analysis processing, syntax analysis processing, emphasis information setting processing, and the like. The result of this language analysis processing is sent to the subroutine SUB2.

Next, in subroutine SUB2,
Based on the result of SUB1, the phoneme processing unit 40 performs phoneme analysis processing.
The phonological analysis processing includes prosody word generation processing, accent command generation processing, pause / phrase command generation processing, and the like.
The final analysis result of the phonological analysis (that is, the data in the intermediate language) is sent to the control parameter generation unit 50, and the processing procedure proceeds to step S11.

Rules are prepared in the subroutine SUB2 so that a pause / phrase command can be issued accurately. This rule, the number of beats shortest preset as L _1, the number of beats included limits are in the prosodic phrase if the L _2, longer than the pulse rate limit L ₂
If there is ICRLB prosodic phrase, all the prosodic phrase inserts phrase symbol P ₃ to prosodic word containing boundary so as to limit L ₂ less than or equal to the number of beats. However, if the distance is the number of L ₁ following the shortest beat from phrase symbol P ₁ / P ₂ / P ₃ immediately before omitted insertion phrase symbol P _3. Further, when there are a plurality of long ICRLB prosodic phrase division methods, a cost function described later is applied to prosodic phrases that can be divided, and division that minimizes the value of the cost function is selected. Therefore, a pause phrase command will be generated for this selected prosodic phrase.

In step S11, the control parameter generation unit 50 generates control parameters required for speech synthesis based on the result of the phoneme analysis processing of the subroutine SUB2. The control parameters generated here proceed to step S12.

In step S12, the audio signal generation unit 60 performs an audio signal generation process based on the control parameters obtained in step S11. By this processing, a synthesized speech corresponding to the sentence finally supplied is generated and output as speech. With this voice output, this series of processes of the voice synthesizer 10 ends.

The above-described subroutines SUB1 and SUB2 will be further described. When performing the language analysis process in the subroutine SUB1 of FIG.
Proceed to SS10. In sub-step SS10, morphological analysis processing is performed. The morphological analysis unit 32 recognizes the boundaries of sentences from the data supplied by the word decomposition analysis unit 32a and divides the sentences into sentences. The morphological analysis unit 32 further converts words that match the partial character strings forming the elements of the character string of the divided sentence into the word dictionary 32 of FIG.
Search from b. The morphological analysis unit 32 also checks grammatical connection possibility and divides the sentence into word strings.
After this processing, the flow advances to sub-step SS11.

In sub-step SS11, sub-step SS10
The syntax analysis unit 34 performs syntax analysis using the processing result (word string). In this parsing, the supplied word strings are combined into phrases, and the modification relationship between the phrases is analyzed. According to this analysis, for example, when "Low pressure is south of the Izu Peninsula and rain is showering in eastern Japan" by ICRLB, the syntax tree shown in Fig. 7 (a) or the relationship shown in Fig. 7 (b) A receiving structure is generated. Here, the symbol “/” in FIG. 7 indicates the ICRLB boundary of the sentence. At this time, the syntax analyzer 34 also determines ICRLB, clauses, enumeration expressions, parallel relations, and the like. After this processing, the flow advances to sub-step SS12.

In sub-step SS12, the results of the morphological analysis process and the syntax analysis process are supplied to the emphasis information setting unit 36, and the supplied data is subjected to the emphasis information setting process.
The setting rule storage unit 36b of the emphasis information setting unit 36 stores rules for classification into three cases other than emphasis, suppression, and both. The information setting unit 36a compares the supplied phrase with the rule from the setting rule storage unit 36b, and sets emphasis information corresponding to the matching rule in this phrase.

Next, in sub-step SS13,
The processing results obtained in SS10, SS11, and SS12 are supplied to the phoneme processing unit 40. After this supply, the process returns to the subroutine SUB1 and ends.

Immediately after the end of the subroutine SUB1, the process proceeds to the subroutine SUB2 of FIG. In subroutine SUB2, phoneme analysis processing is performed based on the result of the language analysis processing of subroutine SUB1 described above. The phonological analysis processing includes a prosodic word generation processing, an accent command generation processing, a pause / phrase command generation processing, and the like. First, the process proceeds to sub-step SS20.

In the sub-step SS20, the data from the language analysis unit 30 is supplied to the prosody word generation unit 42 to perform processing for generating a prosody word. The prosodic word generation unit 42 performs accent connection of words in the supplied data (phrase). Further, the prosodic word generation unit 42 also generates a prosodic word by performing accent connection between the phrases as necessary. The generated prosodic word also has a function of setting phrase emphasis information, not shown in the prosodic word generation unit 42 of FIG. After this processing, the flow advances to sub-step SS21.

In sub-step SS21, the supplied ICRLB
To determine the magnitude of the accent of each prosodic word on the basis of the information of. This processing is performed by the accent command generation unit 44. In the accent command generation unit 44, the standard storage unit 44
The accent command setting unit 44a compares a rule stored in advance in b, for example, a rule for accent transformation with the supplied ICRLB information. Accent command setting section 44a
Is set by assigning accent marks that match the rules to ICRLB information. After this setting, the process proceeds to subroutine SUB3.

In a subroutine SUB3 shown in FIG. 9, a pause command / phrase command representing a prosodic feature is set in the supplied data (pause / phrase command generation processing). In this pause / phrase command generation process, first, in order to perform a process of generating a pause command / phrase command, first,
ICRLB division processing is performed in subroutine SUB4. ICRL
The B division process further includes subroutines SUB5, SUB6, and SUB7 for performing a division candidate generation process, a parameter storage process, and a division selection process described later. The ICRLB prosodic phrase (that is, the division candidate) that has been optimally divided through these processes is supplied to the command setting unit 46a. Thereafter, the process proceeds to sub-step SS30.

In sub-step SS30, the command setting section 46a sets a pause command and a phrase command for the supplied division candidates and outputs them to an intermediate language creating section (not shown). After this processing, the process proceeds to the return. Upon return, the process of setting the pause command / phrase command is completed, and the flow shifts to sub-step SS22.

In sub-step SS22, sub-step SS2
0, SS21, and SUB3 are processed based on the data obtained, such as prolonged sounding and prompting, and then an intermediate language is created. The created intermediate language data is supplied to the control parameter generation unit 50. After this output, the process proceeds to the return, and the processing of this subroutine SUB3 is terminated.

The ICRLB division processing of the subroutine SUB4 performed in the above-described subroutine SUB3 will be briefly described with reference to the flowchart of FIG. When the processing shifts to the subroutine SUB3 in FIG. 9, the subroutine SUB4 is started to perform the ICRLB division processing, and the division candidate generation processing, the parameter storage processing, and the division selection processing are sequentially performed. The division candidate generation process is a subroutine SUB5 (see FIG. 11), the parameter storage process is a subroutine SUB6 (see FIG. 12),
The division selection process is performed in a subroutine SUB7 (see FIG. 13). After these series of processing, the data is stored in subroutine SUB3
Pass to. At this time, data is supplied to the command setting section 46a of FIG. 4 corresponding to the processing of the subroutine SUB3.
The command setting unit 46a sets a pause command / phrase command for the input data as described above in accordance with the rules of the standard storage unit 46b.

Next, a subroutine for performing division candidate generation processing
SUB5 will be described with reference to FIG. In the division candidate generation processing, when the processing shifts to the subroutine SUB4, the processing of the subroutine SUB5 is started immediately and the process proceeds to the sub-step SS50.

In sub-step SS50, processing is divided according to the type of data supplied from language analysis unit 30. When the data is, for example, a phrase (Yes), the process proceeds to sub-step SS51. If the data has already been divided at an appropriate ICRLB boundary, the process proceeds to sub-step SS52 without performing this process.

In sub-step SS51, the supplied data is divided into ICRLB. This processing is performed by the rule dividing unit 460a according to the above-described first rule.

In sub-step SS52, the length of the supplied data (ICRLB prosodic phrase) is compared with a preset division length. The basic unit of the length of the division length is one beat given corresponding to the reading of words. Therefore, this comparison is made with the number of beats of this division length and the number of beats of the ICRLB prosodic phrase. The comparison condition is based on whether the number of beats of the ICRLB prosodic phrase is equal to or less than the number of beats of the division length in the second rule. In the comparison processing, for example, a flag F may be provided to change the supply destination of the processing result depending on the number of processing. A comparison process is performed on the data supplied from the language analysis unit 30 (flag F = 0
When). At this time, if the condition is satisfied (Yes), FIG.
The division length determination unit 60b supplies the supplied ICRLB prosodic phrase to the command setting unit 46a. When the condition is not satisfied, that is, when the number of beats of the ICRLB prosodic phrase is longer than the number of beats of the division length (N
o), supply the ICRLB prosodic phrase to the subdivision unit 460c. Thereafter, the process proceeds to sub-step SS53.

In sub-step SS53, the supplied ICRLB
The prosodic phrase is subdivided, and the ICRL divided by the division length judgment unit 460b
B Output data consisting of prosodic phrases.

Next, in sub-step SS54, it is determined whether or not the returned data satisfies the comparison condition. When the comparison condition is satisfied (Yes), the process proceeds to sub-step SS55. If the comparison condition is not satisfied (No), the process returns to sub-step SS53. This determination is made by the division length determination unit 460b.

In sub-step SS55, it is determined whether there is a division candidate that can be a combination of division candidates. When the combination as a division candidate is in the ICRLB prosodic phrase (Yes), the division candidate satisfying the comparison condition is delivered to the subroutine SUB7, and after this delivery process, the process returns to the sub-step SS53. Thereby, the division candidate generation unit 460 supplies data from the re-division unit 460c to the division candidate selection unit 462. If the combination that is a candidate for division is not included in the ICRLB prosody (No), the process proceeds to return.

Here, the presence / absence of a division candidate is obtained by calculation. In this calculation, first, after the ICRLB division, the number of beats of the ICRLB prosodic phrase is divided below a set value, and the minimum number of phrase commands obtained by the division is obtained. That is,
An integer value obtained by dividing the total number of beats by the above setting value,
If there is a remainder, this minimum number is an integer value +1. This minimum number is set as a variable PH_NUM. In addition, the maximum number of phrase commands MAX_PH_NUM obtained when the data is finely divided by ICRLB division is obtained. When the variable PH_NUM is determined to require division, the division length determination unit 460b sets the variable PH
Steps up the value of _NUM by +1. With this setting, the actually re-divided data is passed to the subroutine SUB7 after the variable PH_NUM is divided. The combination of division candidates is continued until the variable PH_NUM exceeds the maximum number of phrase commands MAX_PH_NUM. In the flowchart of FIG. 11, the data transfer is a subroutine for convenience.
This is indicated by the display of SUB7. The data delivery is not limited to this method, and the data may be stored in a memory and a combination of division candidates may be delivered in a lump. These determination processes are performed by the division length determination unit 460b. By this series of processing, a combination of division candidates is generated.

Next, a subroutine SU for the division selection process
Before explaining B7, parameters used in this division selection process will be briefly described with reference to FIG. The parameters are set in subroutine SUB6. Here, the parameters include a weighting coefficient A for the feature value I used in the expression (1) described later and a coefficient B for multiplying the average beat after division.

Here, the feature amount I is indicated by, for example, emphasis information (+ emp, -emp, 0emp) of the prosodic word at the head of the prosodic phrase. In sub-step SS60, the rule of the feature amount I, that is, the relationship between the emphasis information (+ emp, -emp, 0emp) is stored.

Next, in sub-step SS61, a weighting factor A is stored according to the numerical value of the feature value I. In this embodiment, 7 and -3 are assigned to the weighting factors A for the numerical values -1 and -2 of the feature amount I, and the numerical values are stored.

Next, in sub-step SS62, the coefficient
Stores the value of B. Here, the above-mentioned weight coefficient A and weight coefficient B may be calculated by the parameter storage unit 464 shown in FIG. These weighting factors A and B may be calculated by making full use of, for example, a statistical analysis process of the feature value I in speech synthesis or a multivariate analysis. The parameter storage unit 464 calculates the parameter value in advance outside the device,
It is also possible to simply store a numerical value.

The processing proceeds in such a procedure, and the process proceeds to the return. The storage of the parameters is completed, and the process proceeds to the subroutine SUB7 as shown in FIG. In this subroutine SUB7,
As shown in FIG. 13, cost calculation is performed for each combination of division candidates (ICRLB prosodic phrases) supplied from the subroutine SUB5. Further, the subroutine SUB7 performs a process of selecting an optimal division candidate from the costs obtained according to the procedure of FIG. Here, the cost is an error that is an index of an accurate speech expression when an evaluation is made on a phrase that can be divided.

First, in the sub-step SS70 of FIG. 13, the number of beats in each range divided by the ICRLB boundary for each combination of the supplied division candidates is counted and stored by the calculation unit 462a of FIG. This storage is performed, and the flow advances to sub-step SS71. In sub-step SS71, a variable MIN used for selecting a division candidate
Set the value of. This set value is, for example, 9999.
Further, the contents of the memory MIN_DIV [] for storing the position of the ICRLB boundary of the smallest division candidate and the variable VAL for storing the cost value are cleared. Thereafter, the process proceeds to sub-step SS72.

In the sub-step SS72, the cost is calculated for each of the division candidates with the division number m divided at the ICRLB boundary. This calculation is performed by the calculation unit 462a according to the cost function F (D). The cost function F (D) is given by the equation (1)

[0085]

(Equation 1) Where D (m, n) is a variable indicating that the division candidate is divided into m at the nth in the supplied combination of division candidates, L is the number of beats of the entire ICRLB, L (m , n, i) is a variable representing the number of beats of the i-th prosodic phrase when the n-th combination candidate is divided into m, and Ph (m, n, k) is a variable that represents the n-th combination candidate. k-th prosodic phrase when divided into m, count (Ph)) is k
A _k is a function representing the presence or absence of feature I in the
In the k-th prosodic phrase, the weighting factor of the feature amount I, and B is a weighting factor relating to the average beat after division. In equation (1) expressed using these various variables and weighting factors, the first term is the sum of the above-described errors, the second term is the cost of the feature I, and finally the third term is the average after the division. It represents the cost of beats.
The operation result of each of these terms is the cost of the n-th division candidate.

Next, in sub-step SS73, the calculated cost is stored in a variable VAL. After storing, sub-step SS74
Proceed to.

In sub-step SS74, variable VAL and variable MI
Compare the value of N. In this comparison, when the value of the variable MIN is larger than the value of the variable VAL (VAL <MIN), the process proceeds to sub-step SS75. When the value of the variable VAL is equal to or more than the value of the variable MIN (VAL ≧ MIN), the process proceeds to the sub-step SS76.

In sub-step SS75, the value of variable MIN is replaced with the value of variable VAL, and data indicating the ICRLB boundary position indicating the position at which the n-th division candidate has been divided is stored in memory MIN_DIV []. Replace with data.

In sub-step SS76, it is determined whether or not there is a combination of division candidates. If there are still division candidates (Yes), the flow advances to sub-step SS77. When there are no more division candidates (No), the process proceeds to sub-step SS78.
Here, the determination is made based on the relationship between the variable MAX_PH_NUM and the variable PH_NUM used in the presence or absence of the division candidate described above. That is, when the variable MAX_PH_NUM becomes equal to the variable PH_NUM, it means that there are no more division candidates.

In sub-step SS77, the new supplied
Count the number of beats of the (n + 1) th division candidate. After this counting process, the process returns to sub-step SS71. Further, in sub-step SS78, the cost calculation of the division candidate is completed,
At this point, the smallest division candidate is selected, and the data of the ICRLB boundary position in the smallest determined memory MIN_DIV [] is set as the optimal division candidate as the command setting unit 46a in FIG.
To supply. After this supply, go to return. The process of selecting the division candidates is terminated via the return.

By performing the above-described processing, the division candidate to which the optimal division has been performed is selected, and the pause instruction / phrase instruction is set in the selected division candidate by comparing them with the rules of those instructions.

Next, a description will be given using a more specific example and a comparison with a conventional process will be described. The text of the input sentence is analyzed based on the rules described above. here,
Shortest beat, which is a basic parameter used in the second rule
L ₁ = 5, and the beat limit L ₂ = 15 included in the prosodic phrase. The symbol "↓" inserted between prosodic words indicates an accent command, the symbol "," indicates a prosodic word boundary, and the symbol "P" indicates a phrase command. As a preparation for the arithmetic processing, the parameter storage unit 464 stores the stored feature values I = -1, I =
-2 weighting factors A ₁ = 7, A ₂ = -3, and weighting factor B = 0.
2 is supplied to the calculation unit 462a.

When the input sentence “This tool is inseparable from our lives.” Supplied through the data input unit 20 and the language analysis unit 30 in FIG. The input sentence is divided by the command and the prosodic word element into the following: "Watashi ↓ Tatino, Sekatsu ヲ, Kirihanasenai,
Dog ↓ g, na ↓ mass. Also, when the number of beats of each prosodic word is represented by a number in parentheses, the number of beats of the prosodic word is
They are (6), (6), (7), (4) and (6), respectively. In this case, the maximum number of divisions is 5, and the total number of beats of ICRLB is (29).
The prosodic phrase of the input sentence is divided between the minimum number 2 and the maximum number 5 of phrase commands as shown in FIG. The minimum number is
It is clear from the relationship between the overall number of beats of ICRLB and the beat rate limit L ₂ = 15. The table in FIG. 14 shows the partition positions at which the cost is minimum with the set number of divisions.

As for the cost calculation, the calculation unit 462a calculates a value for each term of the cost function F, and obtains the sum of these values (see equation (1)). Furthermore, as a result of determining at which division position the cost is minimized for a certain number of divisions,
The table in FIG. 14 shows that the cost is minimized when the number of divisions is 2 and the division is made into (6 + 6, 7 + 4 + 6). When a pause command / phrase command is set in the command setting unit 46a in accordance with this division, the data generated for the input sentence is
"P ₁ Watashi ↓ Tachino, Sekatsuwo, P ₃ Kirihanasenai,
Dog ↓ ，, ↓ マ ス マ ス テ イ P _{0 0 0} . In particular, it has been found that by considering the length of the prosodic phrase after division in the third term of the expression (1), data having a natural prosody is generated.

By the way, when the number of divisions is determined using only the first term of the cost function, the sum of the errors is 5 divisions,
That is, when (6,6,7,4,6) was set, the minimum value was 3.6. As a result, the input sentence "P ₁ Watashi ↓ Tachino, P ₃ Sekatsuwo, P ₃ Kirihanasenai, P ₃ Dogu ↓ door, P ₃ Na ↓ Tteimasu P _0" phrase command and is added. However, since the input sentence is divided too finely, the prosody of the synthesized speech obtained based on the generated data is unnatural. By performing processing in consideration of the length of the prosody phrase after division in this way, it is possible to finally generate a synthesized speech accompanied by an accurate prosody.

Further, another input sentence “Mochizuki that supported the opening streak has been struck.” Was supplied to the speech synthesizer 10. In this case, by prosody processing and accent command, "Kai-mak, re-show ↓, sasaeta, mochi ↓ zukiga, uta ↓ letter"
Becomes Also, when the number of beats of each prosodic word element is represented by a number in parentheses, the number of beats of the prosodic word element is (4), (5), (4), (5),
In (4), the minimum number of divisions in this case is 2, the maximum number is 3, and the total number of beats in ICRLB is (22). The table shown in FIG. 15 shows the positional relationship of the phrase command that minimizes the cost for each division number.

Also in this case, the cost function F calculates a value for each term and obtains the sum of these values. As a result, it can be seen that when the number of divisions is 2 and the division is (4 + 5 + 4,5 + 4) among the minimum values of the divisions, the cost is minimized to -1.2. If the setting of the pause command phrase command in the command setting section 46a in accordance with this division, data generated with respect to the input sentence, "P ₁ opening, Les ↓ Nshowo, Sasaeta, P ₃ waxy ↓ Zukiga, Songs ↓ Letter P ₀ ". This is especially true for the post-prosodic emphasis information and the prosodic phrase length after splitting.
It has been shown that by taking into account the second and third terms of (1), natural prosody data is generated.

When the number of divisions is determined using only the first term of the cost function as in the above-described comparison, the total error is divided into two, and the division candidates are (4 + 5,4 + 5 +4) or
When the prosodic phrase was delimited by (4 + 5 + 4,5 + 4), the minimum value was 4. This input sentence can be divided into two types. (4 + 5,4 + 5 +
Data generated by the division candidate 4), "P ₁ opening,
Les ↓ Nshowo, P ₃ Sasaeta, waxy ↓ Zukiga, Songs ↓ Letter P ₀ ". On the other hand, data generated by the division candidate (4 + 5 + 4,5 + 4)," P ₁ opening, Les ↓ Nshowo, Sasaeta, P ₃ mochi ↓ Zukiga, and Uta ↓ Letter P ₀ ".

Here, the phrase command is set before the verb “supported” and the proper noun “mochizuki” is added to the generated data of the (4 + 5,4 + 5 + 4) division candidate. There is no phrase command. This indicates that the undulating word is emphasized by increasing the accent command, but the emphasized word is not emphasized because the phrase command is added to the previous word. That is, this is caused by performing ICRLB division in which the emphasized word does not come at the beginning of the prosodic phrase. Therefore, when the synthesized speech is generated using the former division candidate, the synthesized speech is uttered with an unnatural prosody. In the case of this input sentence, since the sum of the errors is the same as 4, there is a possibility that the former candidate for division may be adopted only in the judgment of the sum of the errors.
From this point of view, the speech synthesizer 10 to which the present invention is applied
Can appropriately select a division candidate that naturally reproduces the prosody, so that the adoption of this inappropriate division candidate can be avoided.

In some cases, ICRLB division is performed so that a suppressed word comes at the head of a prosodic phrase. An example sentence in such a case is, for example, "Phone may be connected." In this example, the result is "denwaga, tsunagatteshimau, koto] ga, alcamo, shiremasen" by the prosody processing and the accent command. When the number of beats of each prosodic word is represented by a number in parentheses, the number of beats of the prosodic word is (4), (8), (3), (4), (5), respectively, and ICRLB The total number of beats is (24). The table shown in FIG. 16 shows the positional relationship of the phrase command that minimizes the cost for each division number.

The cost function F calculates a value for each term, and obtains the sum of these values. As a result of this cost calculation,
When the number of divisions was 2 and it was divided into (4 + 8 + 3,4 + 5), the cost for the input sentence was the smallest at 3.6. If the setting of the pause command phrase command in the command setting section 46a in accordance with this division, data generated with respect to the input sentence, "P ₁ Denwaga, Tsunagatteshimau, Koto] moth, P _3, A} Rukamo, Shiremasedin ”. This is because, in particular, by considering the prosody enhancement information and the length of the prosodic phrase after division in the second and third terms of the equation (1), data having a natural prosody is generated. became.

On the other hand, when the division candidates are evaluated using only the first term of the cost function F, the division number is 2 and (4 + 8,3 + 4 + 5)
When divided, the error became 0 and the minimum. At this time,
Data generated with respect to the input sentence, "P ₁ Denwaga, Tsunagatteshimau, P ₃ things] moth, A} Rukamo, Shiremase}
". As indicated by this data, a phrase command is set immediately before “KOTO”. However,
It is known that formal nouns such as are usually suppressed and not pronounced strongly. Even if the minimum value is found by summing the errors due to only the first term of the equation (1), if a candidate for the division in which the suppressed word comes at the beginning of the prosodic phrase is selected, the synthesized speech is eventually generated with an unnatural prosody. Will be lost. Therefore, by evaluating the data of the input sentence in all the terms of the cost function, it becomes possible to select an accurate division candidate.

For example, when the input sentence “Basic policy to establish the same resale rules as applied to the disposal of state-owned land has been decided” is input to the speech synthesizer 10, the processing of the prosodic words, By the processing of the accent command, the input sentence is
It is processed as "Kokuyu] Chino, Show ↓ Bunni, Tekiyosare ↓ Tailnot, Onaji, Tenbaikiso] ヲ, Moke ↓ Rutouyu, Kihonho ↓ Shinshin, Kimemamashita. The number of beats for each prosodic word is (6), (4), (11), (3), (8), (7),
In (8) and (5), the total number of beats was 52. When only the first term of the cost function F is evaluated, as shown in FIG.
In addition, it can be seen that the sum of the errors is as small as 6 in the case of the (6 + 4,11 + 3,8 + 7,8 + 5) division candidates. In addition, even when comparing the total cost of all terms of the cost function with each combination of division candidates, the cost is 3.4 when the number of divisions is 4 and (6 + 4,11 + 3,8 + 7,8 + 5). , Indicating the minimum. Thus, the same result may be obtained in some cases.

Next, a modification of the speech synthesizer of the present invention will be described with reference to FIGS. This speech synthesizer 10 has basically the same configuration as that of the above-described embodiment, except that a pause / phrase command generation unit 46 of the phoneme processing unit 40 is provided.
Has been modified. The pause / phrase command generation unit is a main part of the phoneme processing unit 40 with this change.
The configuration of 46 is shown in FIG.

The pause / phrase command generation unit 46 includes a pause / phrase command setting unit 46A, as in the above-described embodiment.
An ICRLB dividing unit 46B is provided. In this case, the pause / phrase command setting unit 46A is supplied with the output from the language analysis unit 30 to the command setting unit 46a. The phrase command setting unit 46A includes a command setting unit 46a and a standard storage unit 46b.
And The standard storage unit 46b stores a standard for giving a pause command and a phrase command to supplied data. The command setting unit 46a has a function equivalent to that of the standard division unit 460a of the above-described embodiment, and the data supplied is a beat limit L ₂ included in the prosodic phrase in this standard.
There is a function to judge whether or not to satisfy. According to this condition, the command setting section 46a selects the data output destination. The conditions will be described in detail in a later operation.

The ICRLB dividing unit 46B includes an actual case searching unit 466,
A division example storage unit 468 is provided. Example search section 46
6 includes a prosodic word segmentation unit 466a for decomposing the data supplied from the command setting unit 46a into prosodic words,
A search key collating unit 466b for extracting information of the retrieval key for the output of 6a and collating it with the examples stored in the divided example storage unit 468, and a language analysis unit based on the collation result of the retrieval key collating unit 466b. An example corresponding division unit 466c that divides the output from the output unit 30 and gives a pause command and a phrase command to this output is included. The search key collating unit 466b performs collation using the same information as the collation retrieval key included in the divided actual example storage unit 468.

The division example storage unit 468 stores, in a memory, data of an example already determined to be an appropriate division as a search key for collation. The search key for collation includes actual input sentence data, emphasis information, prosodic words, the number of beats of the prosodic words, the total number of beats, the type of the prosodic words, and / or accent commands. Is used.

The operation of the pause / phrase command generator 46 will be briefly described with reference to the subroutine SUB8 of FIG. The subroutine SUB8 is a subroutine SU shown in FIG.
This is a pause / phrase command generation processing routine that replaces the processing of B3. In this process, it is preferable to store a plurality of correctly divided examples in the divided example storage unit 468 in advance. This example includes a pause command and a phrase command.

[0109] In sub-step SS80 of the subroutine SUB8 shown in FIG. 19, performs determination of whether to satisfy the prosody beats length of is defined clause limit L ₂ supplied. At this time, the command setting section 46a, the value of the standard storage beats stored from section 46b limit L ₂ is supplied. The command setting unit 46a determines whether the length of the prosodic phrase satisfies the (second) rule based on the numerical value (division length determination step). When the rule is satisfied (Yes), the process proceeds to sub-step SS81. When the rule is not satisfied (No), the process proceeds to sub-step SS82.

[0110] In sub-step SS81, the length of the prosodic phrase to satisfy the rules are in the limit L ₂ less than the number of beats, adding pause symbols phrase symbol corresponding to between prosodic phrase. The command setting unit 46a adds a pause symbol / phrase symbol in accordance with the rules of the pause command / phrase command stored in the standard storage unit 46b. After this processing, the flow shifts to return.

In the sub-step SS82, the supplied data is sent to the actual example search unit 466 of the ICRLB division unit 46B, and divided into prosodic words. This splitting process (ie, ICRLB splitting)
This is performed by the prosodic word segmentation unit 466a of the example search unit 466. The prosodic word element division unit 466a compares the divided data with the search key collation unit 46.
Output to 6b. After this processing, the flow advances to sub-step SS83.

In sub-step SS83, a search is made for a division example that matches the information contained in the divided data, and the degree of coincidence of the retrieval is determined. In practice, this search is performed by the search key matching unit 466b using the output from the prosodic word segmentation unit 466a as a search key and the information stored in the division example storage unit 468 as a search key for comparison. An example of this search key will be described later. When a completely matching actual example is obtained in the search (Yes), the process proceeds to sub-step SS84. Search key collation unit 46
6b is a sub-step when there is no exact match (No)
Proceed to SS85.

In the sub-step SS84, the divided actual example storage section 46
8, the matching example is output via the search key matching unit 466b and the example correspondence division unit 466c. At this time, the example corresponding division unit
466c simply passes through the information of this example and supplies it to the output destination of the command setting unit 46a when the rule is satisfied. After this supply, a return is made.

In the sub-step SS85, the division example storage section 46
8, the output from the language analysis unit 30 is divided according to the instance that matches the most, and a process of adding a pause instruction / phrase instruction is performed. Search key collation unit 466b
Supplies a high-matching example to the example-corresponding division unit 466c. In such a case, since the input sentences may not match, the example correspondence division unit 466c divides the output from the language analysis unit 30 according to the information from the search key collation unit 466b, and (Pause command, phrase command, etc.). The example corresponding division unit 466c supplies the output to the same output destination as the output destination of the command setting unit 46a in sub-step SS84. After this processing, the flow shifts to return. The processing of the subroutines SS84 and SS85 corresponds to an actual example search step.

In the return, the subroutine SUB8 for selecting a division candidate suitable for the actual example ends, and the sub-step SS22
Proceed to. In the subsequent processing, after performing the processing of the subroutine SUB2, a synthesized speech is generated by the main routine. With such a modification, the speech synthesizer 10 can be simplified.

A specific example of the information used for this search is shown below. For one input sentence, the division example storage unit 468 stores, for example, a prosodic word element, a division position (indicated by a symbol “/”), a prosodic word prime number, an overall number of beats, emphasis information, a beat number of each prosodic word element. , The type of the dependency, the position of the accent command, and / or the type of part of speech are replaced with numerical values and stored.

For example, if the input sentence is “fisherman who has gone fishing to a distant sea”, the relation between the prosodic words and the division positions: Tokno, Umimade, (/) Ryoni, Detail, Ryoshiga and the prosodic words prime number is 5 , The total number of beats is 19, the emphasis information of each prosodic phrase is 0emph, 0emph, 0emph, 0emph, 0emph, and the number of beats of each prosodic phrase is (4), (4), (3), (4), (4 ). P at this time
The symbols are (P ₁ ) Tokno, Umimade, (P ₃ ) Ryoni, Detail, Ryoshiga (P ₀ ), and the sentence example with emphasis “I arrived at the fishing island that is an island near the border” is a prosodic word element. Relation of division position: Coccyoni, Chica ↓ i, Shima ↓ Deal, (/) Waterojimani, Tsu ↓ Ita and the prosodic word prime number is 5, the total number of beats is 23, the emphasis information of each prosodic phrase is 0emp
h, 0emph, 0emph, + emph, 0emph and the number of beats in each prosodic phrase is
(5), (3), (5), (7), (3). The P symbol at this time is (P ₁ ) KOKKYONI, CHICA ↓ I, SHIMA ↓ DEAL, (P ₃ )
Uotsurijimani, have a relationship that Tsu ↓ Ita (P _0).

The sentence example including repression is also one of the strengths of having served as the chairman. The relation between prosodic words and division positions: Kaicho ヲ, Tsuto ↓ Meta, Koto ↓ Mo, (/) Tsuyomino,
Human ↓ Tsudes, prosodic prime number is 5, the overall beat is 19,
The emphasis information for each prosodic phrase is 0emph, 0emph, -emph, 0emph, 0emp
In h, the number of beats of each prosodic phrase is (4), (4), (3), (4), (4). At this time, the P symbol is (P ₁ ) Kaicho ヲ, ↓ ↓ meta, ト ↓ モ, (P ₃ ) tsuyomino, and human ↓ tsudes (P ₀ ).

Finally, as an example of using the type of dependency as information, the above-mentioned input sentence “It is an indispensable tool in our lives.” Is “Watashi ↓ Tachino, Sekatsukara, (/) Kakasenai, Dog ↓ G, Na ↓ Meiteis, prosodic word prime number 5, the total number of beats is 27, the dependency type is continuous, continuous, continuous, continuous, none for each prosodic phrase, and the emphasis information of each prosodic phrase is 0emph, 0emph, 0emph, 0emph,
At 0 emph, the number of beats in each prosodic phrase is (6), (6), (5), (4), (6). At this time, the P symbol is (P ₁ ) I ↓ Tachino, Sekatsukara, (P ₃ ) Kakasenai, Dog ↓ G, N ↓ Itomies (P ₀ ). By digitizing such information and storing it in the division example storage unit 468 and using it as a search key for collation at the time of search, it is possible to set parameters without performing
Example retrieval and retrieval accuracy can be improved.

With the above configuration, a reasonable
Since ICRLB division can be performed, phonological processing for generating an unnatural synthesized voice can be avoided, so that a synthesized voice that is more natural and easy to hear can be output. This speech synthesizer can semi-automatically cope with parameter value changes accompanying rule changes. The operability of this device can be facilitated, and the labor required by humans can be reduced.

Further, by using an example obtained by searching and selecting an example, processing such as parameter setting and calculation processing can be eliminated, and cost can be reduced by simplifying the apparatus configuration.

The present invention is not limited to the above-described embodiment. For example, the present invention may be configured by combining the above-described embodiment with this modification. At the initial stage when the speech synthesizer 10 is started to operate, an optimal division candidate is determined by the configuration of FIG. 4 to accumulate information, an intermediate language is created, and the information of the determined division candidate is stored in the division example storage unit 46.
8 for learning various examples in advance. After performing this processing and accumulating the information in the division example storage unit 468, the processing for the supplied data is switched to the above-described modified example (the configuration shown in FIG. 18) to search for the information, and to execute the pause command / phrase. A command may be added. As a result, first, data for reliable speech synthesis used by the user is stored in the speech synthesizer 10, so that useless data storage can be avoided. By performing a search by combining search keys without performing the search, a desired division candidate and information to be added thereto can be easily obtained.

In the above-described embodiment, the number of beats is used in the equation (1) for selecting the division position. However, the calculation may be performed using the central word of the phrase, the type of inter-phrase dependency, or the like as a variable. it can.

[0124]

As described above, according to the speech synthesizer of the present invention, the output of the language analysis means is divided by the division candidate generation means based on the first rule, and the divided prosodic phrases are divided by the second rule. When the prosodic phrase is re-divided as determined by the above, the combination of the prosodic phrases generated by the positions separating the prosodic phrases is used as a division candidate, and information and parameters of speech expressions included in the division candidate obtained by the division candidate selection means are stored. An evaluation value indicating the validity of the division is calculated using the parameters supplied from the means, and the division candidate selected by selecting the division candidate satisfying the third rule by the division candidate selection means using the evaluation value. Is converted into a prosodic phrase divided into optimal lengths, and the phonological analysis means performs intermediate processing in the speech synthesizer that generates an intermediate language according to this division, so that phonological processing that produces unnatural synthesized speech is avoided. More It can be output easily synthesized speech heard by natural changes in the parameter values due to changes in rules can also be semi-automatically correspond. As a result, the operability of the device can be further facilitated, and the labor required by humans can be reduced.

Further, the speech synthesizing apparatus stores an example in the divided example storage means, searches the stored information that matches the information from the outside by the divided example search means, and uses the search result to match the stored information. When a command or the like is added to a division candidate, high-quality synthesized speech can be obtained with a simple configuration without performing processing for parameters, calculating the cost of the division candidate, and the like. The cost of the apparatus can be reduced.

According to the text analysis method of the speech synthesizing apparatus of the present invention, the analysis result of a sentence is processed and determined in order according to the first rule and the second rule, and information is re-divided according to the result. The combination obtained at the time of
On the other hand, the evaluation value of the division validity is calculated using the stored parameters and the information of the speech expression included in the division candidate, and the third value is calculated.
By performing a phonological analysis to select a division candidate that satisfies the rule of and generating an intermediate language corresponding to the division candidate,
It is possible to improve the accuracy of outputting a synthesized voice that is more natural and easy to hear.

Further, the division candidate stores in advance the information to be supplied and / or the information to be added to an appropriately divided example as a search key, and determines the output destination according to the judgment result for the supplied prosodic phrase. After the selection, searching for an example corresponding to the division of the prosodic phrase supplied at the selected output destination, and by giving various instructions according to the search result, the useless operation is greatly reduced. It is also possible to generate a more natural and high-quality synthesized speech by an easy operation.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a schematic configuration of a speech synthesizer of the present invention.

FIG. 2 is a block diagram illustrating an example of a schematic configuration of a language analysis unit in the speech synthesis device in FIG. 1;

FIG. 3 is a diagram illustrating a relationship between processing units of a language handled by a language analysis unit in FIG. 2 and a classification example thereof.

FIG. 4 is a block diagram illustrating an example of a schematic configuration of a phoneme processing unit in the speech synthesis device in FIG. 1;

FIG. 5 is a main flowchart for explaining the operation of the speech synthesizer in FIG. 1;

FIG. 6 is a flowchart illustrating an operation of a language analysis process (subroutine SUB1) shown in the main flowchart of FIG. 5;

FIG. 7 is a diagram showing a specific example of a syntax tree and a dependency structure as a syntax analysis process in the processing routine shown in FIG. 6;

FIG. 8 is a flowchart illustrating the operation of the phoneme analysis process (subroutine SUB2) shown in the main flowchart of FIG.

FIG. 9 is a flowchart illustrating an operation of a pause / phrase command setting (subroutine SUB3) shown in a subroutine SUB2 of FIG. 8;

FIG. 10 is a flowchart illustrating an operation of an ICRLB division process (subroutine SUB4) shown in a subroutine SUB3 of FIG.

11 is a flowchart illustrating an operation of a division candidate generation process (subroutine SUB5) shown in a subroutine SUB4 of FIG.

FIG. 12 is a flowchart illustrating an operation of a parameter storage process (subroutine SUB6) shown in a subroutine SUB4 of FIG.

FIG. 13 is a flowchart illustrating the operation of a division selection process (subroutine SUB7) shown in subroutine SUB4 of FIG.

14 is a diagram illustrating numerical values of variables obtained for a division candidate obtained by dividing the input sentence supplied to the configuration of FIG. 1 and selection of the division candidate.

FIG. 15 is a diagram illustrating numerical values of variables obtained for the division candidates obtained by dividing the input sentence supplied to the configuration of FIG. 1 and selection of the division candidates.

16 is a diagram illustrating numerical values of variables obtained for the division candidates obtained by dividing the input sentence supplied to the configuration of FIG. 1 and selection of the division candidates.

17 is a diagram illustrating numerical values of variables obtained for division candidates obtained by dividing the input sentence supplied to the configuration of FIG. 1 and selection of the division candidates.

FIG. 18 is a block diagram showing a schematic pause / phrase command generation unit which is a main part of another configuration in the phoneme processing unit of FIG. 4;

FIG. 19 shows a pause / phrase instruction process (subroutine SUB8) in accordance with an example of the pause / phrase instruction generation unit in FIG.
5 is a flowchart for explaining the operation of FIG.

[Explanation of symbols]

 10 voice synthesizer 20 data input unit 30 language analysis unit 40 phoneme processing unit 42 prosodic word generation unit 44 accent command generation unit 46 pause / phrase command generation unit 50 control parameter generation unit 60 voice signal generation unit 460 division candidate generation unit 462 division Candidate selection section 464 Parameter storage section

Claims (16)

[Claims] 1. A language analyzing means for analyzing a feature at a language level based on a morpheme included in a sentence supplied as information and a language used for processing a syntax of the sentence, and an output of the language analyzing means. Performs analysis based on the features of the speech language level, and generates phoneme analysis means for generating an intermediate language serving as an instruction for speech synthesis based on the obtained analysis results, and generates control parameters corresponding to the output of the phoneme analysis means. A voice synthesizing device that synthesizes a voice signal based on an output of the control parameter generating device, and synthesizes the information into a voice artificially. A first rule for dividing the output of the linguistic analysis means according to a modification relation of a prosodic phrase in which a plurality of prosodic words which are a chain of morphemes are collected, and a prosody obtained by the division. Second to determine but less below the size that is set in advance
A division candidate generating means for generating, as a division candidate, a combination of prosodic phrases obtained by positions at which the prosody phrase is divided when the prosody phrase is subdivided in accordance with the rule of: A third evaluation step of calculating an evaluation value for evaluating the validity of the division using the information of the speech expression including the information and selecting a division candidate having the minimum evaluation value;
A speech synthesis apparatus comprising: in the phonological analysis means, division candidate selection means for selecting a division candidate according to the following rule: and parameter storage means for storing a parameter used for calculating an evaluation value in the division candidate selection means. . 2. The apparatus according to claim 1, wherein the division candidate generation unit divides an output of the language analysis unit using the first rule, and A second rule processing means for judging the divided length in the output of the first rule processing means by the second rule; and a second rule processing means for determining the divided length in accordance with the judgment of the second rule processing means. And a subdivision unit configured to subdivide the prosodic phrase to be output to generate a plurality of division candidates. 3. The apparatus according to claim 1, wherein the parameter storage means includes: a first weighting factor based on information of a speech expression in the division candidate; and an average of prosodic words according to the number of divisions of the division candidate. And a second weighting factor for multiplying the number of beats. 4. The apparatus according to claim 1, wherein the information of the speech expression is emphasis information for designating emphasis and suppression of the prosodic word element and excluding both. Synthesizer. 5. The apparatus according to claim 1, wherein the phonological analysis unit stores an example of a division setting position of a prosodic phrase satisfying the rule, and the phonological analysis unit stores the example. And a divided example search unit for searching an example in which the supplied sentence is optimally divided from the examples. 6. The apparatus according to claim 5, wherein the divided example search means includes a search for the example stored by the divided example storage means, the emphasis information, the prosodic word element, and a beat rate of the prosodic word element. , A total number of beats, a type of dependency of the prosodic word, an accent command and / or a type of part of speech. 7. The apparatus according to claim 1, wherein the language analysis unit sets an emphasis information rule and includes an emphasis information setting unit that stores the set emphasis information. Speech synthesizer. 8. Analyzing features of the language level based on a morpheme included in a sentence supplied as information and a language used for processing the syntax of the sentence, and based on the analysis result, a speech language level for a prosody of the sentence. After performing an analysis based on the characteristics of the above, an intermediate language serving as a command for speech synthesis is generated based on the obtained analysis result, and a control parameter corresponding to the generated intermediate language is generated. In a text analysis method of a speech synthesizer for artificially synthesizing a corresponding speech, the method includes dividing the information according to a modification relation of a prosodic phrase in which a plurality of prosodic words that are a chain of the morphemes are combined. A rule dividing step of dividing the analysis result of the sentence using the first rule, and determining whether the size of the prosodic phrase is smaller than a predetermined size as a second rule, A division length determining step of determining the result of the rule division step by the second rule; and a prosody phrase delimiter setting position obtained by subdividing the information according to the determination result of the division length determining step. A division candidate generating step of generating a division candidate as a division candidate; a parameter storage step of storing a parameter used for evaluating division validity based on information of a speech expression included in the division candidate obtained in the division candidate generation step; An evaluation value calculating step of calculating an evaluation value from the information of the speech expression included in the division candidate generated in the step and the parameter; and selecting a division candidate having the smallest evaluation value among the evaluation values obtained in the evaluation value calculation step. And a division candidate selection step of selecting the division candidate based on the third rule. 9. The method according to claim 8, wherein the division candidate redivides the prosodic phrase when a beat length included in the prosodic phrase is larger than a preset subdivision beat length indicating re-division. A text analysis method for a speech synthesis apparatus, characterized in that a symbol for speech tone command, which is one of the features of the speech language level, is inserted into the boundary of the division candidate. 10. The method according to claim 8, wherein the evaluation value is a difference between a beat length of a section obtained by dividing the prosodic phrase and a beat length obtained by equally dividing the beat length of the entire prosodic phrase. An error sum calculating step of calculating a sum of absolute values; and a section obtained by dividing the prosodic phrase according to the presence of a feature amount represented by information of a speech expression included in the feature of the speech language level in the prosodic phrase. A feature amount sum calculating step of calculating a sum of beat lengths; a weight calculating step of calculating a weight coefficient based on information of a speech expression included in the prosodic word of the division candidate; a feature amount sum calculating step Is multiplied by the weighting factor obtained in the weight calculation step, and a weighted feature amount sum calculation step of calculating the sum of the multiplication results; and the average beat length after dividing the prosodic phrase and the prosodic phrase The product of the beat lengths obtained by dividing the entire beat length equally A product calculation step to be issued, and a result of the error sum calculation step and a result of the weighted feature sum calculation step, and a result of the product calculation step is subtracted from a result obtained by the addition. A text analysis method for the speech synthesis apparatus, wherein the text analysis method is calculated using an evaluation value calculation step of calculating an evaluation value of a division candidate. 11. The method according to claim 8, wherein the information of the phonetic expression is set in such a manner that the information is classified into emphasis and suppression of the prosodic words and a case where both are excluded. A text analysis method for a speech synthesizer to be executed. 12. The method according to claim 11, wherein the emphasis is classified by including a proper noun or a number in a prosodic clause of the division candidate, and the suppression is performed by using a formal noun, a verb in a prosodic clause of the division candidate. Or the beginning of the sentence, and classifying the prosodic phrase by including a particle at the end of the prosodic phrase, and assigning a predetermined value to the emphasis and the suppression. Text parsing method. 13. The method according to claim 8, wherein the division candidate stores in advance information to be supplied and / or information added to an appropriately divided example as a search key, and determines the division length. In the process, an output destination is selected in accordance with a result of the judgment on the supplied prosodic phrase, and thereafter, an example corresponding to the division of the prosodic phrase supplied by the selected output destination is searched, and according to the search result, A text analysis method for a speech synthesizer, comprising an example search step of dividing the prosodic phrase and giving various instructions. 14. The text analysis method according to claim 13, wherein in the example storing step, various examples are stored in a learning manner in advance. 15. The method according to claim 8, wherein said parameter is obtained by statistical analysis or multivariate analysis. 16. The method according to claim 8, wherein the language is Japanese.