JP2900454B2 - Syllable data creation method for speech synthesizer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION A. Industrial Field of the Invention The present invention relates to a speech synthesis device using a rule synthesis method,
In particular, the present invention relates to a method for creating syllable data by analyzing a speech waveform.

B. Summary of the Invention The present invention analyzes voice waveforms and creates syllable data of CV data, VC data, and common V part data. A transition section is provided, and the original waveform of the common V section data and the transition section are superimposed by weighting to obtain a CV waveform, a VC waveform, and a common V section waveform. By creating syllable data for the data, parameter mismatch is eliminated when combining syllable data.

C. Conventional technology A speech synthesizer based on the rule synthesis method divides an input character string into words and phrases by syntactic analysis, determines intonation and accent, respectively, and decomposes words and phrases into syllables and even phonemes. Alternatively, the parameters of the sound source wave and the articulation filter for each phoneme are obtained, and a synthesized speech is obtained as a response output of the articulation filter to the sound source wave.

In such a speech synthesizer, in order to perform rule-based synthesis in syllable units, parameters characterizing speech in units of consonants + vowels (CV data) or vowels + consonants (VC data) are stored in a syllable parameter memory, and input characters are stored. According to the sequence, rules such as connection, duration and sound intensity (energy, pitch frequency) for each phoneme are given externally to change speech feature parameters, and these are input to the articulatory filter to obtain synthesized speech. I have to.

Here, syllable data is reduced as a syllable data unit.
A system having only 110 CV data is known, but with only this CV data, the connection between the sound source and the articulation parameters when the connection point of the CV data is switched from the V portion of the preceding syllable to the C portion of the following syllable. Mismatch occurs, and the synthesized speech waveform is greatly distorted, causing abnormal sounds in the synthesized speech.

Therefore, CV data and VC data have conventionally been used as syllable units, and between the CV data of the preceding syllable and the CV data of the subsequent syllable.
There has been proposed a method of performing a connection that inserts VC data.

D. Problems to be Solved by the Invention In a conventional speech synthesizer using CV data and VC data, the connection from the V section of the preceding syllable to the C section of the subsequent syllable is VC
Although the smoothness is caused by the interposition of the data itself, there is a problem of abnormal noise generation due to parameter mismatch in the connection from the VC data V part to the VC data V part and the connection from the VC data C part to the CV data C part. there were.

In addition to the CV data and VC data, the common V part data (six types of Ai-Wao and I) is provided and the V part of the CV data
There is also a method of using common V-part data to transfer VC data to the V-part. However, even with this method, a parameter mismatch remains in the connection, and a problem in the connection of the C-part also remains.

For example, in the case of synthesizing the voice of "Kama", it has CV data, VC data, and common V part data, and creates syllable data by connecting the following data (a) to (d) in order. Become.

(A) Data from consonant / K to vowel / A /. Normal,
It is created from the original sound "ka".

(B) Common V part data of vowel / A /. Usually, it is created using the stationary part of the original voice "A".

(C) Data from vowel / A / to consonant / M /. Normal,
Created from the original sound "Ama".

(D) Data going from vowel / A / to silence. Usually “A”
Using the latter half of the original voice.

For each piece of data used in the above synthesis, a necessary section is cut out in advance based on many original voices actually generated by a person. Therefore, a total of 50 units per unit such as CV, VC, common V part, etc.
Slightly more than 0 syllable data will be incorporated in the speech synthesizer.

However, when creating each syllable data to be used for speech synthesis, since each data is created from a separate original speech, speech synthesis using syllable data obtained by analyzing each original speech as it is, There will be gaps in the waveform before and after the connection of each data. That is,
If the original vowels and consonants have different original sounds, the sound will change subtly.Therefore, if the original sound is generated differently or the waveform is reconstructed using the extracted parameters, the difference in the waveform characteristics before and after the connection generally occurs. It will come out.

Therefore, a change in the shape and characteristics of the discontinuous waveform occurs at the connection portion of each waveform data, resulting in a synthesized voice having a large sense of noise.

An object of the present invention is to provide a method for creating syllable data in which parameter mismatch is reduced in combining syllable data.

E. Means and Action for Solving the Problems In order to achieve the above object, the present invention provides a method for analyzing vocal waveforms and creating syllable data of CV data, VC data, and common V part data. From the waveform, a common / transition section waveform having a V section common section and a transition section set on both sides of the section is obtained. From the original waveform of the CV data, the V section and the C section head have the same time width as the transition section. A CV data waveform defining a transition section is obtained, and a VC data waveform defining a transition section having the same time width as the transition section is obtained from the original waveform of the VC data in the V part and the C part thereof, and the CV data waveform or VC is obtained.
The transition section of the data waveform and the transition section of the common / transition section waveform are weighted and superimposed to obtain a CV data waveform, a VC data waveform, and a common V section data waveform. Each syllable data of the data and the common V part data is created. The CV data waveform, VC data waveform, and common V part data waveform were weighted with each other in the transition section to obtain a weighted CV data waveform, VC waveform data, and common V part waveform, and these data waveforms were analyzed to determine each parameter. Syllable data is used, and parameter mismatch is reduced in combining syllable data corresponding to the input character string.

That is, when combining each syllable data for speech synthesis, the parameters of each syllable data of the CV, VC and V parts are obtained by analyzing the weighted and weighted waveforms in the transition section of the data waveform. By doing
Similar waveforms are obtained in the transition section, that is, similar parameters are obtained, and mismatch is reduced by smoothly changing the parameters in the transition section.

F. Embodiment FIG. 1 is a flowchart of a waveform combining process according to an embodiment of the present invention, showing a combining process of a CV data waveform and a VC data waveform. In reading the common section and transition section waveform of the V section in step S1, the common / transition section waveform of the V section (vowel) of the CV data waveform and the VC data waveform is determined as waveform data for each vowel. For example, the common / transitional section waveform of the vowel "A" is obtained from the speech waveform of the vowel "A" shown in FIG.
The waveforms of the sections are stored, and the bc section is defined as a common section, and the ab section and cd section on both sides thereof are defined as transition sections.

The weighted combination in step S2 is based on the V of the CV data waveform.
A weight is assigned to the transition section determined in the section and the transition section of the V section read in step S1. For example, when the waveform of the CV data waveform “KA” shown in FIG. 3 becomes the eh section, the eg section is set as the CV data waveform, and the fg section is set as the transition section. This transition section fg has the same time width as the transition section ab in FIG. Then, as shown in FIG. 4, a-b of the common / transitional section waveform read from the voice waveform "A"
Waveform processing is performed with the waveform of the section set to zero weight at point a and weight 1 at point b, and similarly, weight processing is set to 1 at point f and zero weight at point g of transition section f-g of the CV data waveform "KA". Then, both waveforms are weighted by the transition sections ab and fg to obtain a CV data waveform “KA”.

Next, the weighted combination in step S3 is performed in step S2
In the same manner as the weighted connection by the transition section cd of the CV data waveform obtained in the above, the weighted connection by the similarly determined transition section on the VC data waveform is performed. For example, the VC data waveform
It is assumed that the audio waveform “AME” shown in the figure is cut out as a section i−1, which is divided into a common section jk and transition sections ij, k−
1 is determined, and the first half transition section ij is set to have the same time width as the transition section cd of the common / transition section waveform, and the respective sections are superimposed by weighting processing. The CV data waveform, the VC data waveform, and the common V part data waveform obtained in this manner are analyzed to extract each parameter, thereby creating syllable data.

As described above, the vowels "A", "I", "U", "E",
The common section and the transition section of the V part waveform of “O” and the nasal vowel “N” are obtained in advance, and the CV data waveform and the VC data waveform are cut out as waveform data having the transition section respectively, and the CV data waveform is obtained. When generating the VC data waveform and the common V part data waveform, waveform data weighted by the transition section is obtained, and the parameters of the syllable data are obtained by analysis using this waveform. Thereby, when combining syllable data at the time of speech synthesis, parameter mismatch in the data connection unit is eliminated, and occurrence of abnormal noise is eliminated.

Next, in the C-part weighted combination of the VC data waveform and the CV data waveform, weighting is performed by the same weighting as described above at the transition section k-1 of the VC data and the head of the CV data, and the VC data waveform and the CV data are combined. Create each waveform. At this time, since the consonant parts are weighted and connected, there is no transition such as a vowel, so that it is not necessary to provide a common section in the C section, and a sufficiently smooth waveform transition can be obtained by combining only the transition sections. Become.

Note that the common / transition interval waveform is weighted and coupled to the CV data waveform in advance as the V portion of the CV data waveform, and the VC data waveform is weighted and coupled in advance as the V portion of the VC data waveform. Parameter mismatch can be reduced while using syllable data of CV data and VC data.

G. Effects of the Invention As described above, according to the present invention, the CV data waveform and VC
In order to obtain syllable data by analyzing the data waveform and the common V part data waveform, in order to obtain each data waveform by superimposing the common / transition interval waveform and the CV data waveform and the VC data waveform weighted by the transition intervals respectively set, It is possible to eliminate a parameter mismatch between the V portion and the C portion, prevent occurrence of abnormal noise in the portion, and obtain a smooth synthesized voice.

[Brief description of the drawings]

FIG. 1 is a flowchart showing one embodiment of the present invention, and FIG.
The figure is a waveform diagram of vowel "A", FIG. 3 is a waveform diagram of CV data,
FIG. 4 is a waveform diagram of the VC data, and FIG. 5 is a weighted combined waveform diagram of the transition section.

──────────────────────────────────────────────────続 き Continued on the front page (58) Fields surveyed (Int. Cl. ⁶ , DB name) G10L 3/00-9/18 JOIS file (JICST)

Claims (1)

(57) [Claims] An utterance waveform is analyzed to generate syllable data of CV data, VC data, and common V part data. In the vowel basic waveform, a V part common section and transition sections set on both sides of the section. From the original waveform of the CV data, and from the original waveform of the VC data, a CV data waveform that defines a transition section having the same time width as the transition section at the beginning of the V section and the C section thereof is obtained from the original waveform of the VC data. A VC data waveform that defines a transition section having the same time width as the transition section is obtained in the V section and the C section, and the transition section of the CV data waveform or VC data waveform and the transition section of the common / transition section waveform are respectively determined. CV data waveform and VC data waveform and common V
A syllable data creation method for a speech synthesizer, comprising: obtaining a partial data waveform; analyzing each data waveform to create each syllable data of CV data, VC data, and common V part data.