JPS6346497A - Voice synthesization system - Google Patents

Abstract

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

Description

TECHNICAL FIELD The present invention relates to a speech synthesis method, and more particularly to a pitch pattern generation method and an amplitude pattern generation method in speech synthesis.

Furthermore, natural prosodic control is cited as a factor that influences the quality of sentence-to-speech synthesis. Prosody includes elements such as phonological duration, amplitude, and pitch frequency. Conventionally, modeling using one-point pitch and accent type has been performed for pinch pattern control. Japanese accent is a pitch accent, and modeling pitch patterns based on accent types is an effective method, but it does not take into account the number of moras or phonological characteristics of words.

Furthermore, it is currently not possible to generate a pinch pattern for any word by taking these into account. Furthermore, with regard to amplitude pattern control, technological development has lagged behind pitch control rules, probably because Japanese accents are pitch accents, so it is necessary to calculate the average amplitude value for each phoneme and use this as the phoneme-specific amplitude. There are only two methods, such as a method of changing the pitch pattern and a method of multiplying the change in the pitch pattern by a coefficient to obtain a change in the amplitude pattern.

Purpose The present invention was made in view of the above-mentioned circumstances.
In particular, by using multivariate statistical analysis techniques to generate pitch patterns in speech synthesizers, it is possible to determine pitch patterns that can synthesize highly natural speech, and to improve prosody in the rule-based synthesis of speech. This was done for the purpose of generating an amplitude pattern of synthesized speech so as to enhance its naturalness.

Structure In order to achieve the above object, the present invention provides a speech rule that reads out a parameter series of speech segments prepared in advance according to an input character string or a phoneme series, connects them using a combination rule, and adds prosody using a prosody rule. In the synthesis method, qualitative parameters (word accent type) that contribute to pitch pattern generation.

By conducting quantitative type 1 analysis (categorical multiple regression analysis) using the number of moras of words, phonological sequence, etc.
Create a prediction model that minimizes the least squares error with Il'l constant data, obtain a pitch value using this model, and generate a pitch pattern. In the speech rule synthesis method that reads and connects using connection rules and adds prosody using prosody rules, it is characterized by processing various parameters using multivariate statistical analysis to obtain optimal control values and generate amplitude patterns. This is what I did. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a block diagram for explaining an example of a pitch pattern generation method that is an embodiment of the present invention.
2 is a speech analysis data section, 2 is a quantification type I analysis section, 3 is an input qualitative parameter section, 4 is a prediction model section, and 5 is a pitch interpolation section. Using multivariate statistical analysis of various parameters that contribute to pitch pattern generation, we constructed a model that generates pitch patterns that take into account phonological characteristics, etc., making it possible to synthesize highly natural synthesized speech. In this example, a model for predicting a target variable using qualitative factors through quantification 11M analysis will be considered. As the target variable Yi to predict, δ(jk) is a tick variable, and R is the number of factor items.

Cj is the number of categories in the jth item.

ajk is the coefficient (quantity) for the category.

It is.

In order to optimally predict the sample value yi using Yi, the number k (ajk) is determined so as to satisfy Σ(yi-Yi)2→minimum i=1.

Now, in order to find the number ft (ajk), prepare a large number of speech samples, and set the pitch value as the target variable (at least one pitch value for the phoneme or syllable, such as the center of gravity of the phoneme, the energy center of the center or vowel, or the boundary of the phoneme). The pitch value at a point) is observed, and the qualitative parameters that contribute to the pitch pattern, such as the phonological environment, mora position, word mora number, and accent type, are determined for that pitch point (in Figure 2, point A is the vowel (indicating the center), these can be used as factor items. In this way, by inputting the qualitative parameters at the pitch point to be determined into the configured model, the pitch (predicted value) at that point can be determined, and these points can be interpolated to form the pitch pattern of the word. It becomes possible to do so.

In addition, the pitch pattern that expresses the accent of Japanese words can be approximately expressed by linear approximation as shown in Figure 3 (point B in Figure 3 indicates the pitch change point), and the positions of these nodes and their It can also be modeled in the same way as above using the quantification method using the two pitch values as objective variables.6 Figure 4 is for explaining an example of an amplitude pattern generation method which is another embodiment of the present invention. In this block diagram, in order to generate a natural amplitude pattern, this embodiment simultaneously handles various parameters that are thought to affect the amplitude pattern, and uses multivariate statistical analysis to find the optimal predicted value for the parameters as a whole. A control model was created to obtain the
la is an actual amplitude value section, 11b is a factor item section, 12 is a quantification type analysis section, 13 is an input factor item section, 14 is a prediction 41 remodeling section, and 15 is a predicted amplitude value section. In Figure 4, at least one point for a phoneme or syllable, such as the center of gravity of a phoneme, the energy center of a vowel, or a phoneme boundary, is assumed to be an amplitude change point, and the amplitude value at this point is taken as an external reference, and the phoneme environment around 1 is calculated. A control model is set by performing quantitative type I analysis using factors such as mora position, number of moras, and accent information. With the model configured in this way, it is possible to obtain the predicted amplitude value at that point by inputting the factor item at the point of amplitude change to be obtained, and it is possible to generate an amplitude pattern by interpolating the points of Niku et al. It becomes possible. Furthermore, it is also possible to take the position of the amplitude change point (fifth measurement point) and its amplitude value as external standards, and model it in the same manner as above by quantification 1m analysis.

Effects As is clear from the above explanation, according to the present invention, a statistical analysis method (quantification type I analysis method) is used.

By constructing a model for qualitative parameters, it is possible to generate pitch patterns that can produce highly natural synthesized sounds. Furthermore, by using a multivariate statistical analysis method (quantification type 1 analysis method), a model for qualitative parameters can be constructed, and an amplitude pattern can be generated so that a highly natural synthesized sound can be obtained.

[Brief explanation of drawings]

FIG. 1 is a block diagram for explaining an example of a pitch pattern generation method according to an embodiment of the present invention, FIG. 2 is a diagram showing the relationship between pitch frequency and the center of gravity of a vowel, and FIG. , a diagram showing the relationship between pitch frequency and pitch change point, 4th
The figure is a block diagram for explaining an example of an amplitude pattern generation method which is another embodiment of the present invention, and FIG. 5 is a diagram showing amplitude change points. 1...Speech analysis data section, 2...Quantification type I analysis section. 3... Input qualitative parameter section, 4... Prefull'
l model department. 5... Pitch value section, 11a... Actual amplitude value section, 11b
...Factor item, 12...Quantification ■Type analysis section, 1
3... Input factor item section, 14... Prediction model section. 15... Predicted amplitude value section. Patent applicant: Ricoh Co., Ltd. Figure 1 Figure M2 (+Fang) (?11) Figure 3

Claims (9)

[Claims] (1) Contributes to pitch pattern generation in a speech rule synthesis method that reads a parameter sequence of speech units prepared in advance according to an input character string or phoneme sequence, connects them using a combination rule, and adds prosody using a prosodic rule. The method is characterized in that it creates a prediction model that minimizes the least squares error with measured data by performing quantitative type I analysis using qualitative parameters, and thereby obtains pitch values and generates pitch patterns. Speech synthesis method. (2) at least one for each phoneme or syllable;
A speech synthesis method according to claim (1), characterized in that a quantified Class I analysis is performed using a pitch value of a point as an objective variable. (3) The speech synthesis method according to claim (1), characterized in that a quantitative Class I analysis is performed using the position of a pitch pattern change point and its pitch value as objective variables. (4) The speech synthesis method according to claim (3), wherein the position of a pitch pattern change point is expressed as a value relative to the normalized duration length of each phoneme or syllable. (5) In the speech rule synthesis method, which reads the parameter series of speech segments prepared in advance according to the input character string, connects them using connection rules, and adds prosody using prosody rules, various parameters are processed using multivariate statistical analysis. A speech synthesis method characterized by generating an amplitude pattern by obtaining an optimal control value. (6) Quantification using qualitative parameters (accent information for each utterance, number of moras, phonological sequence, strong information, etc.) that are considered to influence the amplitude pattern to be generated.
A speech synthesis method according to claim (5), characterized in that a prediction model that minimizes a square error with an actual measurement value is created by performing Class I analysis, and thereby an amplitude pattern is generated. (7), at least one for each phoneme or syllable;
A speech synthesis method according to claim (6), characterized in that a quantification type I analysis is performed using an amplitude value of a point as an objective variable. (8) A speech synthesis method according to claim (6), characterized in that a quantification type I analysis is performed using the position of a change point of an amplitude pattern and its amplitude value as objective variables. (9) The speech synthesis method according to claim (8), wherein the position of the change point of the amplitude pattern is expressed as a value relative to the normalized duration length of each phoneme or syllable.