JPS5913040B2 - speech synthesizer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a control information creation device for a speech synthesis device, and particularly to an accent information creation device.

10 As voice responses have been put into practical use, it has begun to be pointed out that a shortcoming of conventional recording and editing methods is that the number of words (types of words, usually referred to as words) that can be responded to is small.

Analyzing this required example, we find that in the vast majority of cases, it is not required to be able to respond with 15 completely arbitrary contents, but rather that there are only a relatively small number of possible responses. It is limited to the sentence form, and words that appear in a part of the group, especially words that convey information, such as individual nouns such as person's name, place name, company name, date, and amount, can be arbitrarily selected. I found out that all I had to do was switch them and respond. In other words, in order to prevent practical problems with voice services, it is necessary to be able to synthesize arbitrary words, primarily noun (proper noun) phrases such as people's names, place names, and company names.

By the way, in order to be able to compose arbitrary words,
Since the number of words is almost infinite, it is not possible to compress and store the control information for synthesis, and it is not possible to compress and store the control information for synthesis. It is necessary to create control information depending on the

When this device is broken down in more detail, it can generally be divided into the following means (1) to (4) in the case of Japanese.

(1) Literal representation of the word given as input 15 (
A means of determining the accent type of a word from its kana or romaji (for example, kana or romanization). (2) Divide the spelling of the word given as input into phonetic units (e.g., phonemes, monosyllables, phonological chains, etc.) that are used as units of combination and formation, and based on the determined accent type, A means of determining the length of time (duration) for each unit.

(3) A means for determining a pitch periodic pattern that physically determines the accent of a word from the determined duration and accent type.

(4) Means for converting the above audio information into complete control information in a form suitable for each synthesis method.

The present invention provides an accent information creation device that determines the accent type of a word in 1) above, particularly the accent type of a Tokyo dialect. The principle used in this device will be explained below.

However, as a unit of synthesis, enumerate from the smallest to
(a) Phonemes (each seed sound and various vowels) (b) Monosyllabic (consonant + vowel or vowel only structure (usually called C and V)
(c) Phonological chains (those with the structure of vowel ten consonants + vowel or vowel + vowel (usually written as VlCV2 and VlV2)), etc. However, in the following explanation, based on the balance between the quality of the synthesized speech and the required storage capacity of speech segments, and the characteristics of Japanese,
We will provide a detailed explanation below using monosyllables as units, which are most useful in practice. Now, in the Tokyo dialect, N
Syllables (to be exact, they are called N beats, but below, according to common usage, they are called N beats)
A word (called a syllable) can have N+1 types of accent types.

Japanese accents are defined by changes in the pitch of the voice.
Accent type M is the position where the voice subjectively changes from high to low (called the accent core). There are two ways to express the position: one is the number from the beginning of the syllable just before the nucleus, and the other is the number from the end of the syllable. In the following, the latter will be used. M=1 means that after the word there is a particle ``
The type in which the nucleus comes before the particle when ``ga'' and ``ノ'' are added, M
=0 is a type that does not have a nuclear hail even if it has "ga" or "no". Also, M-0 is called a flat plate type, M\0 is called an undulating type, and the undulating type with M=N is called a high-head type. In the Tokyo dialect, the pitch of the voice always changes between the first and second words, and there is only one change from high to low within a word. Therefore, in the high-initial type, all syllables after the second syllable from the beginning of the word are low. In addition, except for the high-initial type, the first syllable of a word is always low, everything from the second syllable to the syllable just before the nucleus is high, and everything after the nucleus is low. Therefore, if the position of the accent nucleus is known, the accent level of each syllable can be automatically specified. The primary effect of the accent and the corresponding pattern of changes in the pitch period is to make it easier to hear the speech as a single unified pattern, and in the Tokyo dialect, the distinction of meaning has become secondary.

Synthesized speech that is not controlled by accents and corresponding pitches requires ``wide listening'' for each syllable, and is extremely difficult to hear. Therefore, in order to synthesize an arbitrary word with good quality, good pitch pattern information is required, and in order to create pitch pattern information, it is necessary to determine the accent type from the spelling of the word. Therefore, in order to know the procedure for deciding the accent type from a practical point of view,
Analytical studies were conducted on several thousand samples.

In the following explanation, we will explain the analysis procedure and results using examples of relatively frequently used proper nouns, especially company names and personal names. First, in order to analyze the actual state of accent patterns in company names, for example, several thousand company names are randomly selected from an occupational telephone directory. Next, divide this company name into its component words (for example, "Hitachi" and "Hitachi" and "
Seisakusho''), the entire word (hereinafter referred to as a compound word; in the example, ``Hitachi Seisakusho''), and the accent type of each component word are judged subjectively, and the syllable-based notation (a notation that roughly corresponds to kana characters) is used. The accent type and number of syllables are added to each compound word and constituent word written in Roman letters (such as ``gear'' is regarded as one unit), and the data is created. The same is true for personal names, but there are two types of constituent words: surname and first name. Next, from these data, we divided compound words and constituent words for company names and personal names (for personal names, surnames and first names are separated).
Examine the relationship between syllable number and accent type.

From this table, compound words in company names are type 3 (M-3) and type 4.
It can be seen that the nature of the accent type differs depending on the type of noun, as in the case of type (M-4), which is more common, but with personal names, there is more variation. In other words, in the case of company name type nouns, the bond between the constituent words is strong and the accent type of the constituent words tends to disappear, and the accent type as a compound word tends to be determined regardless of the accent type of the constituent words. For nouns, it is expected that there will be weak combinations in which the accent type of the compound word differs depending on the accent type of the constituent words. Therefore, for personal name-type nouns, we can discover the rules of combination by tabulating and observing the relationship between the accent types of constituent words and compound words. In the case of personal names, (a) if the surname is a head-high type (with a nucleus after the first syllable), the compound word is a head-high type; (Accent) The position of the nucleus is maintained, and (c) If the surname is an undulating type other than the head-height type, and the given name is also an undulating type (those with an accent nucleus), the position of the nucleus of the given name is maintained. Served. Now, it is known that in the Tokyo dialect, the accent nucleus does not come immediately after the ``n'' or the consonant ``tsu''. In order to know the phonological structure and the position of the accent nucleus, we created a list of the relationships between the combinations of one syllable before the accent nucleus and two syllables after the accent nucleus and the position of the nucleus, and identified impossible combinations and exceptionally low proportions. Search for combinations and determine the influence law of phonological structure. In other words, it is a transformation rule based on phonological structure. As a result, in addition to the above-mentioned positive sound ``n'' and consonant sound ``tsu'', long vowels (such as ``aichi'') and diphthongs (such as ``ai'')
It was found that the nucleus does not come even after (such as) (the position of the nucleus shifts forward by one syllable). Applying this rule in reverse,
The accent types of the original data that are thought to have been transformed by these rules are returned to their original forms, and the distribution table of syllable counts and accent types is again created. From this table, it can be seen that type 3 is overwhelmingly common among corporate type nouns. Therefore, in the case of corporate type nouns, it is necessary to estimate the type determination procedure for types other than type 3, and in the case of personal name type nouns, it is necessary to estimate the determination procedure for the types of constituent words. Therefore, next we will analyze the elements of "meaning". In order to process the meaning as formally as possible from a practical standpoint, we first divide each constituent word into smaller parts. For example, [manufacturing place] means "manufacturing" and "dokoro".
In other words, ``Hanako'' is divided into ``Hana'' and ``Child''. These will be called element words. All kinds of element words are examined, and the distribution of each accent type is examined in descending order of the number of occurrences: when it appears at the beginning of a word, when it appears in the middle of a word, and when it appears at the end of a word. As a result, it was found that some specific element words tend to have a specific accent type when they come at the end of words. An element word that comes at the end of a word is called a word ending. For example, a company name that ends with "gumi" is type "0", and "
Company names that end with ``sha'' are in the ``2'' type. In the example of a person's name, the accent type of the constituent word is determined from the number of syllables and the ending of the constituent word. For example, names with the endings of ``hey'' or ``pay'' have three syllables with a head-height shape, but names with four or more syllables have an "O" shape. Based on the analysis procedure and results described above, the following procedure can be practically used to estimate the accent type of a proper noun. (1) Determine the degree of bonding between constituent words depending on the type of word (for example, person's name or company name).

(2) If the combination is strong, check whether the word ending is an exception, and if it is an exception, estimate the type by table lookup,
In other cases, it is assumed to be type 3.

(3) If the combination is weak, the type of the compound word is estimated by estimating the type of each constituent word from the ending, and then applying the matching conditions among the combination rules (A) to (C). (4) Apply transformation rules based on phonological structure. By providing the control unit of the speech synthesis device with a means for performing such procedures,
It is possible to obtain easy-to-listen speech that cannot be obtained with a speech synthesizer that does not have such a means. In other words, since the speech produced by a speech synthesizer without such a means does not have information that it forms words that are a group of words with one meaning, it is difficult to tell from where to where the word is a group of words. It's very difficult to listen to because you have to make decisions while listening to each syllable. The situation is similar to reading a sentence written in katakana with no punctuation or spaces. On the other hand, a device that has a means for estimating the comparison set type produces an easy-to-listen sound that corresponds to reading a sentence with punctuation marks and spaces mixed with Kanji and Kana. It is an object of the present invention to provide such a speech synthesis device. A schematic diagram of an apparatus according to an embodiment of the present invention is shown in FIG.

Input command 4 consisting of a code indicating the noun to be synthesized and its type, which is composed of a code string corresponding to a syllable.
is added to the control information creation section 11 of the speech synthesizer. The memory 12 contains a table that describes the endings of each type of noun and the accent type characteristics of the endings. The control information creation unit 11 checks the ending of the spelling of the input command 4 and stores it in the memory 1.
If there is a word ending that matches the word ending in 2, the word accent type that the input command 4 is to be synthesized is estimated based on the accent type characteristics of that ending and the input command, and
The accent type is determined by applying transformation rules depending on the phoneme. Furthermore, the control information creation unit 11 determines the time length of each syllable of the word from this accent type and the input spelling, creates pitch control information from this time length and accent type, and creates pitch control information using the spelling, time length of each syllable, and pitch control information. The speech synthesis section 3 is controlled using the three types of control information as the control information, and the speech 5 is output. Second
The figure is a diagram for explaining the present invention in more detail, and is a block diagram of an apparatus for synthesizing speech using pitch information 6, time information 7, and phoneme information address 8, which are created based on accent types.

Input commands 4 to be synthesized are input to the system control unit 1,
Pitch information 6 and time information 7 respectively created by control information creation units 11-b to 11-d based on accent information 9 created by accent information creation unit 11-a of control information creation unit 11 in system control unit 1 and the phoneme information address 8 are given to the speech synthesis section 3. The speech synthesis section 3 synthesizes speech by connecting the speech segments recorded in the phoneme information memory 2 according to the various information groups provided. In the phoneme information memory 2, speech segments are stored grouped for each syllable, and when a starting address is designated, the segments belonging to that syllable are read out sequentially from the beginning. All fragments have a fixed length, and adding one address at a time determines the address of the next fragment. Pitch information 6 is information specifying the length of a speech segment to be synthesized and is given for each segment. The time information 7 is information specifying the duration of a syllable, and when the duration of that syllable ends, the next time information and phoneme information address are requested and the next syllable is synthesized. Next, the procedure for creating accent information 9 and pitch information 6, which is the main subject of the present invention, will be described in more detail.

In addition, although in FIG. 2, the accent information 9 is shown as being sent from the control information creation section 11-a to the control information creation section 11-b in order to make the explanation easier to understand, it is actually sent via the storage section 12. This is as described in the following explanation. FIG. 3 shows the system control unit 1 shown in FIG. 2 in more detail.

In FIG. 3, the various processing sections 16 to 26 and the control section 10 are composed of microprocessors, and are connected to each other by a data bus 14 and two interrupt lines 13-1 and 13-2.
Mutual calls are made by interrupts and device numbers on the data bus determined for each processing section, and each processing section has its own dedicated memory and processing procedures. Furthermore, the memory 12 is a common memory. The input processing unit 16 inquires of the control unit 10 when there is a request from the center to synthesize and output audio, and if it is possible to process it, puts the input/output control unit 16 into a state where it can accept the content, and receives the content to be synthesized and sends it to the memory 12.
- Record at a predetermined position in a. The content is expressed as a code string L, ......, LN corresponding to the type S and kana characters, and by counting the number of input codes, the number N of syllables of the word to be synthesized can be calculated. The result is also recorded in the memory 12-a. When the input is completed, the input processing section 16 notifies the control section 10 of the end of reception. The control unit 10 activates the person name main processing unit 17 or the company name main processing unit 18 depending on whether the type S of the word to be synthesized indicates a person's name or a company name. If the type S indicates a company name, the information of that S is sent to the ending detection processing unit 19.
Start. The word ending detection processing unit 19 searches the company terminology ending table, and if there is a match, it learns the accent type M required by that ending from the table and outputs it to the memory 12-a, and if there is no matching ending, then control is again passed to the company name processing unit 18 with type 3 as accent type M. The company name processing unit 18 then activates the phoneme transformation processing unit 20. The phoneme modification processing unit 20 stores the constellation in the memory 12-a. Ll,・
・・・・・・・・・ From LN and accent type, judge whether the syllable immediately before the accent nucleus is a consonant, a consonant, or the second vowel of a long vowel, and in that case, increase M by 1 and create a new M+1. is recorded in the memory 12-a as accent type M. If the type S indicates a person's name, the person name main processing unit 17 separates the surname and first name from the code strings Ll, . The word ending detection processing section 19 is activated. The word ending detection processing unit 19 searches the surname and each term ending table, and if there is a match, it determines the accent type according to the table, and if there is no match, it passes control to the person name main processing unit 17 as type 3. . When the person name main processing section 17 finishes estimating the accent type by detecting the end of the last name or given name, it activates the phoneme modification processing section 20 to perform modification based on phoneme information. Person name processing unit 17
After estimating the accent type of the surname and first name, the system estimates the accent type when the entire person's name is considered as one word according to the above-mentioned combination rules (a) to (c). When the company name processing section 18 or the person name processing section 17 finishes estimating the accent type, it passes control to the control section 10. At this time, the accent type M finally estimated is recorded in the memory 12-a together with the code string Ll, . . . , LN and the number N of codes. Pitch information 6 is stored in the memory 12b in advance in a table format for each combination of chord number N and accent type M.
The control unit 10 is recorded in the pitch information processing unit 21.
, the pitch information processing unit 21 learns the position P of necessary pitch information from the accent type M and the number N of chords by looking up the table, and stores the result in a predetermined position in the memory 12-b.
Control is passed to the control section 10. Pitch information transmission processing section 22
When receiving a request to send pitch information from the synthesizer 3, it inquires of the control unit 10 whether it is possible to send the pitch information, and if possible, stores the pitch information in the memory using the position P of the pitch information in the memo I month 2-b as a clue. 12-b and sends it out. On the other hand, the time information processing section 23 and the phonetic information address processing section 25 are activated at the same time as the pitch information processing section 21 in the memo I month 2-a when the accent type determination in the person name processing section 17 or the company name processing section 18 is completed. It is activated and the code string Ll,...
......, memorize the syllable duration corresponding to LN l month 2
-c is determined by a lookup table and stored in a predetermined position in the memory 12-c. Phonological information address processing section 25
is the code string Ll in the memory 12-a,...
・The phoneme information address corresponding to LN is stored in the memory 12-d.
It is determined by a lookup table from the table inside, and is also stored at a predetermined position in the memo 2-d.

The control unit 10 includes a pitch information processing unit 21 and a time information processing unit 2.
Synthesizing unit 3 only when it is confirmed that all processes in 3 and phoneme information address processing unit 25 have been completed.
Respond to requests for information from each person. In this case, according to the request from the synthesis section 3, pitch information is sent out from the pitch information sending section 22, time information is sent out from the time information sending section 24, and phonological information address is sent out from the phonological information address sending section 26. . Figure 4 shows the values of phoneme duration length.

As explained in the above embodiments, the present invention makes it possible to synthesize the speech of any proper noun based on the spelling of the noun.

According to our experiments, approximately 90% of nouns can be synthesized into speech with the correct accent using the present invention. In addition, the remaining 10% was given a sufficient accent to withstand use, and it became possible to respond with a voice that included any proper noun that was easy to hear. In the example described above, for convenience of explanation, a method of controlling one synthesis section has been described, but each processing section does not always work.

Therefore, while information for one synthesis section is not being processed, it is possible to process information for other synthesis sections, and it goes without saying that multiple control for controlling a large number of synthesis sections is also possible.

[Brief explanation of drawings]

Fig. 1 is a block diagram of one embodiment of the present invention, and Fig. 2 is a block diagram of an embodiment of the present invention.
A block diagram of a speech synthesizer that explains the diagram in more detail,
Figure 3 is a diagram for explaining a part of Figure 2 in more detail;
FIG. 4 is a diagram showing the value of the duration length of a phoneme. 1...System control unit, 2...Phonological information memory, 3...Speech synthesis unit, 4...
Input command, 5... Output voice, 6... Pitch information, 7... Time information, 8... Phonological information address, 9...・Accent information, 10
...Control unit, 11...Control report creation unit, 11a...Accent information creation unit, 11-b
...Pitch information creation unit, 11-c...Time information creation unit, 11-d...Phonological information address creation unit, 12...Memory, 12- a...
・Memory for creating accent information, 12-a-2...
... Word ending table memory, 12-b... Memory for creating pitch information, 12-c... Memory for creating time information, 12-d... Memory for creating phonological information addresses. Memory, 13...Interrupt line, 13-1...Interrupt line to the control unit, 13-1...Interrupt line to each processing unit, 14. ...Data bus, 15...
Input/output control section, 16... Input processing section, 17...
...Person name head processing unit, 18...Company name head processing unit, 19...Word ending detection processing unit, 20...
- Phonological transformation processing unit, 21... Pitch information processing unit, 22... Pitch information transmission processing unit, 23...
... Time information processing section, 24 ... Time information sending section, 25 ... Phonological information address processing section, 26.
...Phonological information address sending unit.

Claims (1)

[Claims] 1. Means for inputting words converted into character code strings, voice information creation means for creating information necessary for speech synthesis from the input words converted into character code strings, and voice information creation means based on the output of the voice information creation means. In the speech synthesis device having means for synthesizing (1), the speech information generating means (1)
) a storage means for storing an accent type corresponding to the ending of the word converted into a string of character codes, and (2) the ending of the word inputted into the string of character codes and the character code string stored in the storage means. If the ending of the input character code string matches the ending of the word that is stored as a character code string, the corresponding accent type is specified and a match is made. (3) an accent type specifying means for outputting a signal specifying a predetermined accent type; and (3) a character code immediately before the position of the accent nucleus of the word inputted as a character code string, which is determined by the output of the accent type specifying means. is characterized in that it has an accent type correction means that converts and outputs the specified accent type to a predetermined value if it is a predetermined value, and outputs the specified accent type as it is if it is not a predetermined value. Speech synthesis device.