JPS63100499A - speech synthesizer - Google Patents

Abstract

(57) [Abstract] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to a speech synthesis device.

[Prior Art] FIG. 2 is a block diagram of a conventional speech rule synthesis device. In FIG. 2, an arbitrary character string is input to the character string analysis unit 1, and the character string analysis unit l performs syntactical interpretation of the input character string, determines the intonation pattern of the entire character string, and also determines the intonation pattern of the entire character string. The words included in the character string are searched by referring to the character string, the accent and phonetic symbol sequence of each word in the character string are determined, and the phonetic symbol sequence and accent pattern of the character string are determined. Next, the phonetic symbol sequence, accent pattern, and intonation pattern of the character string determined by the character string analysis unit 1 are output to the phonological rule control unit 3, and the phonological rule control unit 3 stores them in the rule file 4. With reference to the phoneme control rules for combining each phoneme and the prosody control rules for controlling each prosody, the phoneme feature parameters of the character string, the pitch value between each phoneme, the power and duration length of each phoneme, etc. The parameters necessary for speech synthesis are generated, and the generated parameters are output to the speech synthesizer 5. Further, the speech synthesizer 5 performs speech synthesis as is known in the art based on the input parameters and outputs rule-synthesized speech corresponding to the input character string.

In addition, DecTalk (DECTa1k) is known as an English speech rule synthesis device that can generate voices with individuality. It is possible to synthesize six different phonemes.

[Problems to be Solved by the Invention] However, in the above-mentioned two types of conventional speech standard al1 synthesis devices, rule synthesized speech is uniformly generated for input character strings, and therefore, it is difficult for individuals to There is a problem in that it is not possible to automatically reproduce rule-based synthesized speech that has gender characteristics.

The purpose of the present invention is to solve the above-mentioned problems and generate synthesized speech of arbitrary characters or character strings having individual characteristics by inputting a voice signal to a person and analyzing the input voice signal. An object of the present invention is to provide a speech synthesis device that can generate speech.

Means for Solving Problem E] The present invention provides means for generating speech synthesis parameters from characters or character strings based on predetermined phoneme control rules and prosody control rules, and a means for generating feature parameters representing individuality from speech signals. The present invention is characterized in that it includes means for extracting, and means for generating synthesized speech from the speech synthesis parameters based on the feature parameters.

[Operation] With the above configuration, speech synthesis parameters are generated from characters or character strings based on predetermined phonological control rules and prosody control rules, and feature parameters representing individuality are extracted from speech signals. Thereafter, synthesized speech is generated from the speech synthesis parameters based on the feature parameters. Thereby, it is possible to generate synthesized speech of the above-mentioned characters or character strings having the characteristics of the individuality of the input speech signal.

[Embodiment] FIG. 1 is a block diagram of a speech rule synthesis device that is an embodiment of the present invention. In FIG. 1, the same parts as in FIG. 2 are given the same reference numerals. The difference between this speech rule synthesis device and the conventional device shown in Fig. 2 is that it uses natural speech as input and analyzes its individuality, and adds information on the analyzed personality characteristic parameters to create individuality characteristics. It is designed to obtain rule-based synthesized speech, and compared to the conventional device, a personality analysis section 6, a personality rule control section 7, and a rule file 8 having personality control rules are added. It is said that

The configuration and operation of the speech rule synthesis device according to the present invention will be explained below with reference to FIG. In Figure 1, any character string is connected to a microphone or keyboard (not shown).
is input to the character string analysis unit 1 via searches for words contained in the string, determines the accent and phonetic symbol sequence of each word in the string,
The phonetic symbol sequence and accent pattern of the character string are determined. The character string may be any signal that can identify the character string information, and may be, for example, a coded electrical signal or an audio signal.

The phonetic symbol sequence, accent pattern, and intonation pattern of the character string determined in the character string analysis section 1 are output to the phonological rule control section 3, and the phonological rule control section 3 analyzes each phoneme stored in the rule file 4. With reference to the phonological control rule for combining and the phonological control rule for controlling each prosody, the phonological feature parameter of the character string,
Parameters required for speech synthesis, such as the pitch value between each phoneme, the power and duration of each phoneme, are generated, and the generated parameters are output to the individuality rule control section 7.

On the other hand, natural voices to be analyzed for individuality are input to the individuality analysis unit 6 via a microphone (not shown), and the individuality analysis unit 6 analyzes the individuality of the input natural voices, Personal characteristic parameters, which will be described in detail later, are generated and output to the personal rule control section 7. Further, the individuality rule control unit 7 retrieves the corresponding individuality control rule from the rule file 8 based on the individuality feature parameters input from the individuality analysis unit 6, and then first extracts the individuality control rule from the phonological rule control unit 3. The input phonological feature parameters, pitch values, power values, and duration lengths are corrected based on the personality control rules extracted above, and the modified phonological feature parameters, pitch values, power values, and duration lengths are used for speech synthesis. Output to device 5. Furthermore, the speech synthesizer 5 performs speech synthesis as is known in the art based on the input parameters, and outputs a rule-synthesized speech that corresponds to the input character string and has been modified based on the information on the personal characteristic parameters.

With the above configuration, the input character string input to the character string analysis unit 1 is corrected by the information on the personality characteristic parameters of the natural voice input to the personality analysis unit 6, and the character string is modified to improve the quality of the natural voice. The speech synthesizer 5 outputs a rule-synthesized speech having individuality.

Next, regarding the specific analysis method of the individuality characteristic parameters in the individuality analysis unit 6, individuality rule control unit 7, and rule file 8 shown in FIG. ) Classify and explain the cases in which individuality is achieved through temporal variation.

(1) When realizing individuality from the aspect of sound quality In general, it is thought that the resonant frequency of the vocal tract differs depending on individual differences in vocal tract length, and individuality is expressed. By setting a resonant frequency and determining the difference between the set standard resonant frequency of the vocal tract and the resonant frequency of the input natural voice, individuality can be expressed based on information on the difference in resonant frequency. That is, first, from the natural speech input to the individuality analysis unit 6, only voiced sounds that regularly and significantly change depending on the vocal tract length are targeted, and for example, the partial correlation coefficient (
PARCOR coefficient), linear prediction coefficient (LPG), and line spectrum pair (LSP), which are individual characteristic parameters that express phonology, are converted into parameters that can be physically correlated, such as formant frequency and vocal tract cross-sectional area ratio. Convert. Here, if the individuality analysis unit 6 can directly extract the formant frequency and vocal tract cross-sectional area ratio from the input natural speech, there is no need to perform the above conversion.

Next, the phonological feature parameters or rule file 4 of the formant frequency and vocal tract cross-sectional area ratio that are generated in advance as a reference.
The information on the difference is stored as a personalization control rule in the rule file 8, and the personalization rule control section 7 adds personalization using the above-mentioned procedure, and adds the personalization to the phoneme control rules necessary for speech synthesis. obtain the above-mentioned predetermined parameters.

The following method may be used as a simple method for the above operation. That is, first, six types of audio signals, including the five vowels of "Aiueo" and the broken sound of "N", which have individuality to be added, are inputted to the individuality analysis section 6, and the input signals are input to the individuality analysis section 6. After the formant frequency and vocal tract cross-sectional area ratio of natural speech are calculated, the ratio of the formant frequency and vocal tract cross-sectional area ratio generated in advance as a reference and the above calculated value is calculated, and this ratio is calculated based on individuality. It is stored in the rule file 8 as a control rule. Next, in the individuality rule control unit 7, the formant frequency and the vocal tract cross-sectional area ratio, which are the phonetic feature parameters input from the phonology rule control unit 3, are combined with the individuality control stored in the rule phi) and L/8. By multiplying the ratios in the rules, it is possible to obtain phoneme feature parameters to which individuality information is added, and by outputting this to the speech synthesizer 5, it is possible to obtain rule-based synthesized speech having individuality.

Further, after extracting an excitation source from the input natural voice in the individuality analysis unit 6, individuality is extracted from this excitation source,
The individuality rule control unit 7 may add the individuality of this excitation source.

In other words, as a simple method, excitation@
When the source contains a considerable amount of personality information, the personality analysis unit 6 performs linear prediction analysis, and compares the linear prediction residual obtained as a result with the linear prediction residual of the standard generated in advance. The individuality information is extracted, the extracted individuality linear prediction residual is stored in the rule file 8 as the individuality control rule, and the individuality rule control section 7 (
・Add personal information. Here, as an even simpler method, the linear prediction residual of the natural speech obtained by the individuality analysis unit 6 is used as the individuality control rule in the rule file 8.
It is also possible to store and use it.

There is a pitch frequency as the personal information included in the above-mentioned excitation source, and a specific case in which personal information is added using this pinch frequency as a parameter will be described below.

First, the individuality analysis unit 6 extracts the pitch frequency from the input natural speech and calculates the average pitch frequency. On the other hand, a reference pitch frequency is stored in advance in the rule file 4 of the phonological control rules, a ratio is calculated from this reference pitch frequency and the average pitch frequency calculated above, and this pitch frequency ratio is set as the individuality control rule. It is stored in the rule file 8 as . Furthermore, the individuality rule control unit 7
Personality information can be added by multiplying the pitch frequency, which is the phonetic feature parameter input from the phonetic rule control unit 3, by the pitch frequency ratio stored in the rule file 8.

(2) When realizing individuality through temporal variation, individuality information is added by storing prosodic control rules for word accent patterns and power patterns unique to each individual in advance as individuality control rules in the rule file 8. can do. That is, first, the pitch frequency and power are calculated from the input natural speech in the individuality analysis unit 6, and the pitch frequency and power of the prosody control rule stored in the rule file 4 are combined with the pitch frequency and power calculated above. The difference is added to the rule file 8 as a personality control rule. This allows personalization information to be added in the same manner as described above.

Furthermore, if individual-dependent features such as devoicing and nasalization can be extracted, personality information can be added to the rule-based synthesized speech as described below, similar to the case of extracting the personality of word accents.

First, in the case of unvoiced individuality extraction, the individuality analysis unit 6 calculates the pitch frequency from the input natural speech and determines whether it is a voiced sound or an unvoiced sound.

Next, it is checked whether the pitch frequency calculated above is a voiced sound or an unvoiced sound from the reference phoneme control rule fi11 generated in advance. The information on the difference can be stored in the rule file 8 as a personalization control rule, and the personalization information can be added in the personalization rule control section 7 in the same manner as described above.

Next, in the case of extracting the individuality of nasalization, first, the individuality analysis section 6 extracts the nasality from the input natural speech. Nasality can be extracted based on the ratio of high frequency components and low frequency components of natural speech, or the presence of a zero point with a sharp peak due to ARMA 9 fold. Next, the nasality of the standard phonetic control rule generated in advance is compared with the nasality extracted above, and information on whether the speech is nasalized or not, and if it is nasalized, the difference in the degree of nasalization is obtained. The information can be stored in the rule file 8 as a personalization control rule, and the personalization information can be added in the personalization rule control section 7 in the same manner as described above.

The method described above uses the speech rule synthesis device shown in Figure 1 to automatically obtain rule-based synthesized speech to which personal characteristics are added. If it is difficult to automatically regularize individuality feature parameters, or if you require sound quality with better individuality than speech synthesized based on automatically generated rules, you can manually By repeating the process of calculating feature parameters representing gender, creating individuality control rules, generating synthesized speech based on the rules, and recreating individuality control rules based on synthetic speech listening experiments, it is possible to create synthetic speech. It is also possible to improve the quality and optimize the rules.

In addition, the voices of an unspecified number of speakers are analyzed in advance, and many extracted individuality control rules are stored in the rule file 8, which is a storage medium, and by using these individuality control rules, high-quality individuality can be achieved. It is possible to obtain rule-synthesized speech with added . In this case, the individuality analysis section 6 shown in FIG. 1 is unnecessary.

In the above embodiment, a character string is input to the character string analysis section l, but it may be input to the character string analysis section l character by character and processed.

In the above embodiment, even if the amount of information of the phonological control rules and prosody control rules stored in the rule file 4 is small, the individuality rule control is performed based on the individuality characteristic parameters generated by the individuality analysis section 6. Since the parameters necessary for speech synthesis are modified in the section 7 and the modified parameters are output to the speech synthesizer 5, it is possible to generate synthesized speech that is faithful to the natural speech input to the personality analysis section 6. can. Another advantage is that it is possible to easily generate audio with a quality that suits the user's preference.

[Effects of the Invention] As detailed above, according to the present invention, speech synthesis parameters are generated from characters or character strings based on the predetermined phoneme control rule 1'l and prosody control rule 1111, and on the other hand, speech synthesis parameters are generated from the speech signal. After extracting the characteristic parameters representing individuality, synthesized speech is generated from the voice synthesis parameters based on the characteristic parameters, so that the character or character string having the characteristics of the individuality of the input voice signal It is possible to generate synthetic speech. Therefore, even if the amount of information on the rules required in the means for generating the voice synthesis parameters is small, there is an advantage that synthesized voice faithful to the input voice signal can be generated.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a speech rule synthesis device that is an embodiment of the present invention, and FIG. 2 is a block diagram of a conventional speech rule synthesis device. 1... String solution V'r part, 2... Word dictionary, 3
... Phonological rule control unit, 4.8... Rule file,
5...Speech synthesizer, 6...Personality analysis section, 7
...Individuality rule control department.

Claims (1)

[Claims] (1) means for generating speech synthesis parameters from characters or character strings based on predetermined phonological control rules and prosody control rules; means for extracting feature parameters representing individuality from speech signals; A speech synthesis device comprising means for generating synthesized speech from the speech synthesis parameters.