JP5474713B2 - Speech synthesis apparatus, speech synthesis method, and speech synthesis program - Google Patents

Description

  The present invention relates to a speech synthesizer, a speech synthesis method, and a speech synthesis program that generate a synthesized speech waveform from speech synthesis information configured as a set of phonemes.

  Text-to-speech is a typical method of using speech synthesis technology. Hereinafter, a device that generates a speech waveform using a symbol representing the type of phoneme and prosodic features obtained as a result of text analysis or the like as an input is called a speech synthesizer. A speech synthesizer is a component of a text-to-speech conversion system.

  The symbols input to this speech synthesizer are hereinafter referred to as speech synthesis symbols. There are various forms of the symbols for speech synthesis. Here, let us consider a case in which phonological information constituting a series of speech and prosodic information mainly expressed as a pose or a voice pitch are simultaneously described. An example of such a symbol for speech synthesis is JEITA (Electronic Information Technology Industries Association) standard IT-4002 “symbol for Japanese text speech synthesis” (see Non-Patent Document 1). The speech synthesizer generates a speech waveform corresponding to such a speech synthesis symbol. However, since the speech waveform is generally strongly influenced by not only the phonemes to be synthesized but also the types of phonemes before and after and the prosodic features, the correspondence between symbols and speech waveforms is generally complicated.

  There are various methods for generating a speech waveform by a speech synthesizer. The speech waveform is generated based on the short-time spectral features of voice, voiced / unvoiced information, and the fundamental frequency (F0) as direct parameters. The method is the main background art. A typical speech waveform generation method is speech synthesis based on a sound source / filter model. In the sound source / filter model, a sound waveform is synthesized in a signal processing manner by driving an articulation filter that generates sound of sound with an appropriate sound source.

  When using a relatively simple sound source such as an impulse train or white noise source, switching between the impulse train and the white noise source is based on voiced / unvoiced information, and the fundamental frequency of the impulse train can be controlled based on the F0 parameter. it can. On the other hand, MFCC (Mel-Frequency Cepstral Coefficient) or a linear prediction coefficient is used as a parameter representing the characteristics of the spectrum, and an AR (autoregressive) type filter is used as an articulation filter, or in particular, when MFCC is used as a parameter. An MLSA (Mel logarithmic spectrum approximation) filter (see Non-Patent Document 2), which directly uses MFCC as its parameter, is used.

  For example, in order to synthesize speech such as consonants, it is necessary to change the speech synthesis parameters over time. Therefore, in this method, for example, it is common to synthesize speech while changing the characteristics by updating the speech synthesis parameters at a constant period of about 5 ms, for example. One period of this fixed period is generally called one frame. Therefore, in general, in order to synthesize speech, it is necessary to create time-series data of frame periods for speech synthesis parameters from speech synthesis symbols.

  The simplest method is to prepare time-series data of the frame period for a certain phoneme length in advance for each required phoneme, and synthesizing them according to the phoneme sequence of the speech to be generated A method of concatenating the parameter time series to form a speech synthesis parameter time series for one utterance is conceivable. However, as described above, the characteristics of the same phoneme may vary greatly depending on the type of phonemes before and after, the speed of speech, the pitch of the voice, and the temporal distance from the immediately preceding or immediately following pose. In order to deal with such cases, it is necessary to use complex phoneme classifications that take into account the preceding and following phonemes and prosodic features, but with such complex phoneme classifications, the number of phoneme types is enormous. Therefore, it is difficult to create and store in advance all necessary speech synthesis parameter time series sets.

  Therefore, in practice, the time change of the speech synthesis parameter time series is modeled based on an appropriate model, and the model parameters are generated by first predicting from the speech synthesis symbols, and the speech synthesis parameter time series is obtained from the obtained model. Is used to generate an arbitrary speech. Hereinafter, this model is referred to as a speech generation model.

  For example, the features of a speech synthesis parameter of a phoneme are divided into three states in terms of time, and for the number of frames in each state, their statistical distribution parameter vectors are d1, d2, and d3 in order from the first state. The speech synthesis for synthesizing the phoneme is made by concatenating the elements of, making a vector d, and if the statistical distribution parameter vector of each state of the speech synthesis parameter is v1, v2, v3 in order from the first state The characteristics of the parameters can be expressed by four vectors d, v1, v2, and v3 that constitute the parameters of the speech generation model. Further, by constructing a predictor that generates these parameter vectors from speech synthesis symbols in advance and using the predictor during speech synthesis, it is possible to synthesize speech from a relatively small amount of data.

  A typical example based on this method is an HMM speech synthesis method. The HMM speech synthesis method assumes a model based on HMM (Hidden Markov Model) as a speech generation model. The plurality of vectors constituting the parameters of the speech generation model are each determined based on a decision tree from speech synthesis symbols, as in the method used in the state sharing HMM in speech recognition technology (see Non-Patent Document 3). . Here, the decision tree is constructed (learned) using a prepared learning speech and a corresponding speech synthesis symbol.

  When synthesizing one utterance voice, first, a voice generation model for one utterance is constructed by connecting the voice generation models for each unit voice. Then, a speech synthesis parameter time series having the maximum likelihood is obtained for the constructed speech generation model, and this is used for speech waveform generation. The likelihood of the speech generation model with respect to the speech synthesis parameter time series is expressed as follows in the speech generation model, for example.

That is, according independently a and normally distributed statistical distribution of values xi are to other types of speech synthesis parameters of the speech synthesis parameters x at frame i, mean μi of the distribution, when the variance is .sigma.i ^2, the voice When the total length is n frames, the log likelihood of the speech generation model for the time series xi of the speech synthesis parameter x of one utterance is given by the following equation.

However, when the speech synthesis parameters of the frame period are directly modeled with several normal distributions, the most likely parameter series is the one in which the average value of the normal distribution is continuously output within the state, and when the state switches , The value becomes discontinuous. That is, it becomes a stepwise parameter time series. Since this is different from actual speech features, not only speech synthesis parameters themselves (hereinafter referred to as static features), but also dynamic features of speech synthesis parameters, first-order differences of speech synthesis parameter time series data ( Delta), second-order difference (delta delta), and the like are used as feature vectors to perform modeling in consideration of continuous changes in speech synthesis parameters (see Non-Patent Document 4). The delta Δxi and delta delta Δ ² xi of the value xi in the i-th frame of a certain speech synthesis parameter x are given by, for example, Expression (2) and Expression (3), respectively.

Hereinafter, a method for calculating time-series data of speech synthesis parameters will be described. First, for the sake of explanation, the feature vector in frame i is assumed to be o _i . Uppercase letters and lowercase letters in bold in the formula mean vectors (the same applies hereinafter).

The length of the voice is n frames. In addition, the following matrix is defined. However, the superscript T represents a transposed matrix, and the superscript -1 represents an inverse matrix (the same applies hereinafter).

Furthermore, a transformation matrix for obtaining a feature vector time series O including a dynamic feature from a static feature time series X defined by Equations (2) and (3) is W here. That is, the following relationship holds. Here, W is a matrix of 3n rows × n columns.

When the parameter distribution follows a normal distribution, the logarithmic likelihood p (X) of X is given by the following equation. Here, μ is an average vector of O distribution, and U is a variance-covariance matrix of O distribution. Each element of μ and U is determined from a speech synthesis symbol by a decision tree learned in advance.

X that maximizes the log likelihood p (X) satisfies the following relationship.

すなわち、数式（１０）を計算することで、最ゆう基準に基づく、動的特徴を考慮したパラメータ時系列が得られる。音声合成パラメータxを多次元のベクトルに拡張した場合も同様である。 Solving Equation (8) and Equation (9) with respect to X yields the following equation:

That is, by calculating Equation (10), it is possible to obtain a parameter time series in consideration of dynamic features based on the maximum likelihood criterion. The same applies when the speech synthesis parameter x is extended to a multidimensional vector.

"Symbols for Japanese text-to-speech synthesis" JEITA standard IT-4002, March 2005 Sei Imai, Kazuo Sumita, Chieko Furuichi, "Mel Log Spectrum Approximation (MLSA) Filter for Speech Synthesis", IEICE Transactions (A), J66-A, 2, Feb.1983, pp.122-129 Takamura Yoshimura, Keiichi Tokuda, Takashi Masuko, Takao Kobayashi, Tadashi Kitamura, “Simultaneous Modeling of Spectrum, Pitch, and Duration in HMM-Based Speech Synthesis”, IEICE Transactions (D-II), J83-D -II, 11, Nov.2000, pp.2099-2107 Masashi Takashi, Tokuda Keiichi, Kobayashi Takao, Imai Kiyoshi, "HMM-based speech synthesis using dynamic features", IEICE Transactions (D-II), J79-D-II, 12, Dec. 1996, pp.2184-2190

  When learning a decision tree for predicting the parameter vector of the speech synthesis parameter distribution, including not only the phonemic phoneme type but also the linguistic prosodic features such as accent type and accent phrase boundary in its explanatory variables In the distribution of feature vectors predicted by a learned decision tree, the variance of element values associated with delta features and delta-delta features often tends to be smaller than the variance of element values associated with static features. appear. This is considered to be due to the fact that the linguistic prosodic feature has a stronger correlation with the short-time change than the absolute value of the speech synthesis parameter.

  According to the logarithmic likelihood equation (8) of the parameter time series to be generated, in the calculation of the logarithmic likelihood, the distribution average is always weighted by the reciprocal of the distribution variance. Therefore, from the above tendency, in the calculation process, the value including the delta feature and the information including the delta-delta feature information is often relatively larger than the value including the static feature information.

  In a calculation by a device that has limited calculation resources and requires a fixed-point number operation such as a portable terminal, it is necessary to set the value of the calculation result to a certain value or less in order to prevent overflow in the calculation. For this reason, when the range of values that can be processed (for example, the ratio between the maximum value and the minimum value) cannot be taken sufficiently, the information represented by a small value is rounded off due to a digit loss, and an error is likely to occur. That is, in the process of generating the speech synthesis parameter time series, accurate information regarding the static feature distribution is likely to be lost. Since static features are information that determines the absolute position of feature parameters, this rounding error can generate a speech synthesis parameter time series whose position is shifted up and down along the feature axis. Causes unnaturalness.

  The present invention has been made in view of such circumstances, and provides a speech synthesizer, a speech synthesis method, and a speech synthesis program based on an accurate speech synthesis parameter time series even when high-precision calculation is difficult. The purpose is to do.

  (1) In order to achieve the above object, a speech synthesizer according to the present invention is a speech synthesizer that generates a synthesized speech waveform from speech synthesis information that describes the type of unit speech included in a series of unit speech sequences. A first speech synthesis parameter generation unit that generates first speech synthesis parameter time-series data having a large numerical range using feature vector distribution information based on the given speech synthesis information; A second feature of correcting the feature vector distribution information based on the speech synthesis information and having a numerical range smaller than the first speech synthesis parameter time-series data as the time-series data of the difference from the first speech synthesis parameter A second speech synthesis parameter generation unit for generating speech synthesis parameter time-series data; the first speech synthesis parameter time-series data; and the second speech synthesis parameter time A speech synthesis parameter addition unit that adds column data and generates third speech synthesis parameter time-series data is provided, and a synthesized speech waveform based on the third speech synthesis parameter time-series data is generated. .

  As described above, the first speech synthesis parameter generation unit generates first speech synthesis parameter time-series data having a large numerical range, and the second speech synthesis parameter generation unit generates first speech synthesis parameter time-series data. The second speech synthesis parameter time series data having a smaller numerical value range is generated.

  As a result, the first speech synthesis parameter generation unit sets feature information, which has been a problem of rounding error in the conventional method, as a main processing target, and excludes other features from the processing, so that all feature information is also collected. Therefore, it is possible to suppress a rounding error during the processing as compared with the case where the processing is performed. Thereby, the calculation error of the final speech synthesis parameter time-series data can be reduced as a whole.

  (2) In the speech synthesizer according to the present invention, the second speech synthesis parameter generation unit may calculate the average parameter in the distribution information of the feature vector in the process of calculating a general parameter time series based on the maximum likelihood criterion. The distribution information of the feature vector is corrected by replacing the first speech synthesis parameter time series data with the feature vector for calculation.

Thereby, the correction of the feature distribution parameter for obtaining the second speech synthesis parameter includes the feature μ including the dynamic feature corresponding to the first speech synthesis parameter time-series data X ₀ and the parameter μ relating to the distribution average in the conventional method. This can be realized by replacing the difference with the vector.

  At this time, since the sum of the first speech synthesis parameter time-series data and the second speech synthesis parameter time-series data completely coincides with the speech synthesis parameter time-series data according to the conventional method, the approximate speech synthesis is performed. It is possible to generate a speech synthesis parameter time series that is more accurate than a technique including parameter time series data generation processing.

  (3) In the speech synthesizer according to the present invention, the first speech synthesis parameter generation unit generates speech synthesis parameter time-series data from static feature distribution information that does not directly represent temporal changes. It is characterized by that. As a result, it is possible to reduce the influence of the rounding error of the value related to the static feature, which has been a problem in the conventional method, and to finally generate an accurate speech synthesis parameter time series.

(4) Further, in the speech synthesizer according to the present invention, the speech synthesis parameter time series data generated from the feature vector distribution information of the static features is a time series of static feature distribution average parameters. It is said. At this time, in the vector (μ−WX ₀ ), which is a correction result of the feature distribution parameter for obtaining the second speech synthesis parameter, the values of the elements corresponding to the static feature distribution average parameter are all zero. Since no rounding error due to calculation occurs with respect to the value 0, the influence of the rounding error of the value related to the static feature, which has been a problem in the conventional method, is reduced, and finally the speech synthesis parameter time series more accurate than the conventional method Can be generated.

  (5) In the speech synthesizer according to the present invention, the first speech synthesis parameter generation unit may finally generate the first speech synthesis parameter by generating the first speech synthesis parameter. Numerical value range information for each divided time of the series data is stored, and the second speech synthesis parameter generation unit generates the second speech synthesis parameter to generate the second speech synthesis parameter time series data. A numerical change for each divided time is calculated, and the speech synthesis parameter addition unit reflects the stored numerical range information in the calculated numerical change by the addition.

  Thus, it is only necessary to temporarily store the divided time-dependent values of the speech synthesis parameter time series data to be finally generated and reflect the value in the second speech synthesis parameter. Since no substantial calculation occurs for, no error occurs.

  (6) A speech synthesis method according to the present invention is a speech synthesis method for generating a synthesized speech waveform from speech synthesis information that describes the type of unit speech included in a series of unit speech sequences. Using the feature vector distribution information based on the synthesis information, generating a first speech synthesis parameter time series data having a large numerical range, and correcting the feature vector distribution information based on the given speech synthesis information Generating second speech synthesis parameter time-series data having a numerical range smaller than that of the first speech synthesis parameter time-series data as time-series data of the difference from the first speech synthesis parameter; First speech synthesis parameter time-series data and the second speech synthesis parameter time-series data are added to generate third speech synthesis parameter time-series data That includes a step of, is characterized by generating a synthesized speech waveform based on the time-series data and the third speech synthesis parameters.

  In this way, feature information that has been subject to rounding errors in the conventional method is the main processing target, and other features are excluded from the processing, so that all the feature information is processed as compared to the case of processing all at once. Rounding error can be suppressed.

  (7) The speech synthesis program of the present invention is a speech synthesis program that is executed by a computer to generate a synthesized speech waveform from speech synthesis information that describes the type of unit speech included in a series of unit speech sequences. Then, using the distribution information of the feature vector based on the given speech synthesis information, processing for generating the first speech synthesis parameter time series data having a large numerical range, and the feature based on the given speech synthesis information Correcting the vector distribution information, the second speech synthesis parameter time series data having a numerical range smaller than that of the first speech synthesis parameter time series data is obtained as time series data of the difference from the first speech synthesis parameter. Adding the first speech synthesis parameter time-series data and the second speech synthesis parameter time-series data to a third sound It includes a process for generating a composite parameter time series data, and is characterized by generating a synthesized speech waveform based on the time-series data and the third speech synthesis parameters.

  In this way, feature information that has been subject to rounding errors in the conventional method is the main processing target, and other features are excluded from the processing, so that all the feature information is processed as compared to the case of processing all at once. Rounding error can be suppressed.

  The first speech synthesis parameter generation unit generates first speech synthesis parameter time-series data mainly from information in which calculation error has been a problem in the conventional method, and the second speech synthesis parameter time-series generation unit Then, a difference between the final speech synthesis parameter time series data and the first speech synthesis parameter time series data is generated as second speech synthesis parameter time series data.

  In the first speech synthesis parameter generation unit, all feature information is processed in a lump by excluding the other features from the processing, with feature information for which rounding error has been a problem in the conventional method as a main processing target. Compared with the conventional method, the rounding error during the processing can be suppressed. Thereby, the calculation error of the final speech synthesis parameter time-series data can be reduced as a whole.

It is a block diagram which shows the speech synthesizer of this invention. It is a flowchart which shows operation | movement of the speech synthesizer of this invention. (A)-(c) It is a figure which shows an example of each speech synthesis parameter time series data. (A)-(c) It is a figure which shows an example of each speech synthesis parameter time series data.

  In the following description, “unit speech” is a minimum speech processing unit in the speech synthesizer. Specific examples of unit speech include phonemes, syllables, and words. Unit speech is classified in consideration of differences in phonological environment such as the types of phonemes before and after, and differences in prosodic features such as accent, intonation, and speech speed. “Unit utterance” refers to a series of unit speech strings having continuous features, and corresponds to a single sentence utterance or an exhalation paragraph (unit read in one breath). The “speech synthesis symbols” are a series of symbols for describing each type of unit speech included in speech of one unit utterance.

  The speech synthesizer 100 finally generates a speech waveform from the speech synthesis parameter time series. However, the present invention is not limited to a system that generates a speech waveform in a signal processing from speech synthesis parameter time-series data using a sound source / articulation filter. For example, a speech segment database is constructed from pre-recorded speech data, speech segment sequences corresponding to speech synthesis parameter time series data are selected, and speech is synthesized by connecting them. Systems are also included. The speech synthesis parameter may be a multidimensional vector.

(Configuration of speech synthesizer)
FIG. 1 is a block diagram showing the speech synthesizer 100. The speech synthesizer 100 outputs a synthesized speech waveform in response to input of a speech synthesis symbol. As shown in FIG. 1, the speech synthesizer 100 includes a speech feature distribution parameter generation unit 105, a first speech synthesis parameter generation unit 110, a second speech synthesis parameter generation unit 120, a speech synthesis parameter addition unit 130, a speech waveform. The generator 140 is configured. The second speech synthesis parameter generation unit includes a speech feature distribution parameter correction unit 121 and a speech synthesis parameter time series calculation unit 122.

  Hereinafter, each unit will be described along the flow of processing for generating a synthesized speech waveform from a speech synthesis symbol. The speech feature distribution parameter generation unit 105 generates speech feature distribution parameters from the speech synthesis symbol string. Here, the audio feature includes not only a static feature but also a dynamic feature such as a delta feature or a delta-delta feature. The speech feature distribution parameter generation unit includes a predictor that predicts a speech feature distribution parameter using the learning speech. All the above features assume a normal distribution, and the distribution parameters are composed of the mean vector and the variance-covariance matrix. Each of the above parameters can be generated using a decision tree. In the decision tree used here, the relationship between the speech synthesis symbol and the corresponding feature is learned in advance using the learning speech.

The first voice synthesis parameter generation unit 110 generates a first speech synthesis parameter time series data X ₀ from the speech feature distribution parameter. However, the first speech synthesis parameter time-series data may not be the same as the speech synthesis parameter time-series data finally used by the speech waveform generation unit 140.

The second speech synthesis parameter generation unit 120 receives the speech feature distribution parameter and the first speech synthesis parameter time series data X ₀ as inputs, and the speech synthesis parameter time series data X finally used by the speech waveform generation unit 140. If, to generate the time-series data of the first difference of the speech synthesis parameter time series data X ₀ as time-series data X ₁ second speech synthesis parameters. That is, the following relationship is established.

The speech synthesis parameter adding unit 130 receives the first speech synthesis parameter time-series data X ₀ and the second speech synthesis parameter time-series data X ₁ as inputs, and uses the sum sequence at each time as the final speech synthesis parameter. Output as time series data, that is, the third speech synthesis parameter time series X. Finally, the speech waveform generation unit 140 synthesizes and outputs a speech waveform corresponding to the speech synthesis parameter time series data X.

The second speech synthesis parameter generation unit 120, the audio feature distribution parameter modifying section 121, audio feature distribution parameter input mu, a U, is corrected using the X _0. Then calculated time-series data X ₁ second speech synthesis parameters in speech synthesis parameter time series calculator.

From equation (10) and Equation (11), _{X 1} can be obtained by the following calculation.

Expression (12) includes a dynamic feature corresponding to X ₀ , with the parameter μ relating to the distribution average in the calculation process (expression (10)) in the case where correction of the feature distribution parameter for obtaining X ₁ is general. shows that can be realized by replacing the difference between the feature vector WX _0.

First speech synthesis parameter time series data X ₀ of the first speech synthesis parameter generation unit 110 outputs, it is possible to set an arbitrary time-series data. In that case, it is more preferable to set the X ₀ as the final calculation error is reduced. Such X _0, there is a sequence which is constituted by the distribution mean of the static feature at each time. At this time, all the values of the elements corresponding to the static feature distribution average parameter in the vector (μ−WX ₀ ) are zero. Because not occur rounding error by calculation for the value 0, in calculating X ₁ in second speech synthesis parameter generating unit, the influence of rounding error values for the static characteristic which is a problem in the conventional method Finally, it is possible to generate speech synthesis parameter time-series data that is smaller and more accurate than the conventional method.

Alternatively, the value of X ₁ generated by setting a series that is considered to have a small difference from X to X ₀ , such as a series obtained by temporally smoothing the distribution average parameter time series of static features using a low-pass filter Can be made narrower than the range of the value of X according to the conventional method. As a result, the number of digits after the decimal point in fixed-point arithmetic can be further increased, and rounding errors during calculation can be reduced.

(Operation of speech synthesizer)
The operation of the speech synthesizer 100 configured as described above will be described. FIG. 2 is a flowchart showing the operation of the speech synthesizer 100. First, a speech synthesis feature distribution parameter is generated based on the speech synthesis symbol string (step S1). Next, the speech synthesis feature distribution parameter, for generating a first speech synthesis parameter time series data X ₀ by a preset reference (step S2). The reference set in advance is, for example, a reference for separating the calculation result into a large numerical value range and a small numerical value range.

Next, the speech feature distribution parameter for generating the second speech synthesis parameter time-series data is corrected based on the above setting criteria (step S3). Then, to produce a second speech synthesis parameter time series data X ₁ from the speech feature distribution parameter that has been modified (step S4). Then, the two speech synthesis parameter time series data X ₀ and X ₁ obtained as described above are added to generate the third speech synthesis parameter time series data X (step S5). Then, a speech waveform is generated using the third speech synthesis parameter time-series data X (step S6). The series of processes described above can be performed by executing a program installed in a mobile terminal or the like. In addition, the method as described above temporarily stores the numerical range information of the time series data to be finally generated, calculates other numerical changes, and reflects the stored numerical range information in the calculated numerical changes. It can be said that

(Example of speech synthesis parameter time-series data)
An example of the speech synthesis parameter time series data obtained by the above embodiment will be described. FIGS. 3A to 3C are diagrams illustrating an example of each speech synthesis parameter time series data. In the figure, the horizontal axis represents time, and the vertical axis represents a certain dimension value of the speech feature vector. FIGS. 3 (a) shows a first speech synthesis parameter time series data X ₀ obtained as an average value for each is segmented time. This corresponds to a speech generation model assuming that the feature distribution parameter is constant within the divided time. Although the first speech synthesis parameter covers a wide numerical range, an average value can be calculated independently for each divided time, so that an error hardly occurs during the calculation. FIG. 3B shows second speech synthesis parameter time series data X ₁ obtained as a difference between this and the third speech synthesis parameter time series data X to be finally obtained. The second speech synthesis parameter has a complicated time series change, but is limited to a narrow numerical range. FIG. 3C shows third speech synthesis parameter time-series data X obtained by adding the first speech synthesis parameter time-series data X ₀ and the second speech synthesis parameter time-series data X _1. Yes.

FIGS. 4A to 4C are also diagrams illustrating an example of each voice synthesis parameter time series data. In the figure, the horizontal axis represents time, and the vertical axis represents a certain dimension value of the speech feature vector. 4 (a) is to the average value series for each partitioned time, shows a first speech synthesis parameter time series data X ₀ obtained by performing smoothing by polygonal line approximation. First speech synthesis parameter time series data X ₀ is that over a wide range, errors in the time change of the series a simple calculation hardly occurs. In this case, the first speech synthesis parameter time series data X ₀ becomes more stepped sequence by mean value calculated for each segmented section, close to the finally obtained will to speech synthesis parameters . FIG. 4B shows second speech synthesis parameter time series data X ₁ obtained as a difference between this and the third speech synthesis parameter time series data X to be finally obtained. Time series data X ₁ and the second speech synthesis parameters, the change in time series is complex, and is constrained to a narrower numerical range than that shown in Figure 3 above (b). FIG. 4C shows third speech synthesis parameter time series data X obtained by adding the first speech synthesis parameter time series data X ₀ and the second speech synthesis parameter time series data X _1. Yes.

(Modification)
In the above description, a speech waveform is generated from one vector X, but a speech synthesis vector time series is independently calculated for each type of acoustic feature of speech such as spectrum, fundamental frequency, etc., and these are generated by speech waveform generation processing. You may combine and use.

In the above embodiment, the first speech synthesis parameter generation unit and the second speech synthesis parameter generation unit generate the respective speech synthesis parameter time series from the same speech feature distribution parameter generated from the speech synthesis symbol string. However, different audio feature distribution parameters may be used. For example, in order to simplify the process of generating the X _0, using the speech feature distribution parameter generated by a more simplified speech feature distribution parameter generation processing, it is possible to generate X _0.

100 speech synthesis apparatus 105 speech feature distribution parameter generation unit 110 first speech synthesis parameter generation unit 120 second speech synthesis parameter generation unit 121 speech feature distribution parameter correction unit 122 speech synthesis parameter time series calculation unit 130 speech synthesis parameter addition unit 140 Speech waveform generator

Claims (8)

A speech synthesizer that generates a synthesized speech waveform from speech synthesis information that describes the type of unit speech included in a series of unit speech sequences,
A first speech synthesis parameter generation unit that generates first speech synthesis parameter time-series data having a large numerical range using feature vector distribution information based on given speech synthesis information;
In a general parameter time-series calculation process based on the maximum likelihood criterion, an average parameter in the feature vector distribution information, an average parameter in the feature vector distribution information, and a feature vector for the first speech synthesis parameter time-series data, The feature vector distribution information based on the given speech synthesis information is corrected by calculating the time series data of the difference from the first speech synthesis parameter. A second speech synthesis parameter generation unit that generates second speech synthesis parameter time series data having a numerical range smaller than the speech synthesis parameter time series data;
A speech synthesis parameter addition unit that adds the first speech synthesis parameter time-series data and the second speech synthesis parameter time-series data to generate third speech synthesis parameter time-series data;
A speech synthesizer for generating a synthesized speech waveform based on the third speech synthesis parameter time-series data. The first speech synthesis parameter generation unit obtains numerical range information for each divided time of the third speech synthesis parameter time-series data to be finally generated by generating the first speech synthesis parameter. Save and
The second speech synthesis parameter generation unit calculates a numerical change for each divided time of the third speech synthesis parameter time-series data by generating the second speech synthesis parameter,
The speech synthesis apparatus according to claim 1, wherein the speech synthesis parameter addition unit reflects the stored numerical range information on the calculated numerical change by the addition. A speech synthesizer that generates a synthesized speech waveform from speech synthesis information that describes the type of unit speech included in a series of unit speech sequences,
First speech synthesis parameter time series data having a large numerical range is generated from static feature distribution information that does not directly represent temporal changes among feature vector distribution information based on given speech synthesis information. A first speech synthesis parameter generation unit;
The feature vector distribution information based on the given speech synthesis information is corrected, and the numerical range as the time series data of the difference from the first speech synthesis parameter is larger than that of the first speech synthesis parameter time series data. A second speech synthesis parameter generation unit for generating small second speech synthesis parameter time-series data;
A speech synthesis parameter addition unit that adds the first speech synthesis parameter time-series data and the second speech synthesis parameter time-series data to generate third speech synthesis parameter time-series data;
A speech synthesizer for generating a synthesized speech waveform based on the third speech synthesis parameter time-series data.   4. The speech synthesizer according to claim 3, wherein the speech synthesis parameter time series data generated from the feature vector distribution information of the static features is a time series of static feature distribution average parameters. A speech synthesis method for generating a synthesized speech waveform from speech synthesis information that describes a type of unit speech included in a series of unit speech sequences,
Generating first speech synthesis parameter time-series data having a large numerical range using feature vector distribution information based on given speech synthesis information;
In a general parameter time-series calculation process based on the maximum likelihood criterion, an average parameter in the feature vector distribution information, an average parameter in the feature vector distribution information, and a feature vector for the first speech synthesis parameter time-series data, The feature vector distribution information based on the given speech synthesis information is corrected by calculating the time series data of the difference from the first speech synthesis parameter. Generating second speech synthesis parameter time series data having a numerical range smaller than the speech synthesis parameter time series data;
Adding the first speech synthesis parameter time-series data and the second speech synthesis parameter time-series data to generate third speech synthesis parameter time-series data;
A speech synthesis method, comprising: generating a synthesized speech waveform based on the third speech synthesis parameter time-series data. A speech synthesis method for generating a synthesized speech waveform from speech synthesis information that describes a type of unit speech included in a series of unit speech sequences,
First speech synthesis parameter time series data having a large numerical range is generated from static feature distribution information that does not directly represent temporal changes among feature vector distribution information based on given speech synthesis information. Steps,
The feature vector distribution information based on the given speech synthesis information is corrected, and the numerical range as the time series data of the difference from the first speech synthesis parameter is larger than that of the first speech synthesis parameter time series data. Generating small second speech synthesis parameter time-series data;
Adding the first speech synthesis parameter time-series data and the second speech synthesis parameter time-series data to generate third speech synthesis parameter time-series data;
A speech synthesis method, comprising: generating a synthesized speech waveform based on the third speech synthesis parameter time-series data. A speech synthesis program that is executed by a computer to generate a synthesized speech waveform from speech synthesis information that describes a type of unit speech included in a series of unit speech sequences,
Processing for generating first speech synthesis parameter time series data having a large numerical range using distribution information of feature vectors based on given speech synthesis information;
In a general parameter time-series calculation process based on the maximum likelihood criterion, an average parameter in the feature vector distribution information, an average parameter in the feature vector distribution information, and a feature vector for the first speech synthesis parameter time-series data, The feature vector distribution information based on the given speech synthesis information is corrected by calculating the time series data of the difference from the first speech synthesis parameter. Processing for generating second speech synthesis parameter time-series data having a numerical range smaller than the speech synthesis parameter time-series data;
Adding the first speech synthesis parameter time-series data and the second speech synthesis parameter time-series data to generate third speech synthesis parameter time-series data,
A speech synthesis program for generating a synthesized speech waveform based on the third speech synthesis parameter time-series data. A speech synthesis program that is executed by a computer to generate a synthesized speech waveform from speech synthesis information that describes a type of unit speech included in a series of unit speech sequences,
First speech synthesis parameter time series data having a large numerical range is generated from static feature distribution information that does not directly represent temporal changes among feature vector distribution information based on given speech synthesis information. Processing,
The feature vector distribution information based on the given speech synthesis information is corrected, and the numerical range as the time series data of the difference from the first speech synthesis parameter is larger than that of the first speech synthesis parameter time series data. Processing for generating small second speech synthesis parameter time-series data;
Adding the first speech synthesis parameter time-series data and the second speech synthesis parameter time-series data to generate third speech synthesis parameter time-series data,
A speech synthesis program for generating a synthesized speech waveform based on the third speech synthesis parameter time-series data.

Images