JP5345967B2 - 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. In this method, for example, the speech synthesis parameters are updated at a fixed period of about 5 ms and the characteristics are changed. It is common to synthesize speech while 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 a certain phoneme for the length of the frame period for each required phoneme in advance, 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. By generating, a method that enables synthesis of arbitrary speech is used. 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 first 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 to other types of speech synthesis parameters values x _i of the speech synthesis parameters x at frame i, mean mu _i of the distribution, when the variance is sigma _i ² , the length of the speech is to a total of n frames, one log likelihood of the speech production model for sequence x _i when the utterance of the speech synthesis parameters x is given by the following equation.

  However, when the speech synthesis parameters of the frame period are directly modeled with several normal distributions, the maximum likelihood parameter series is the one in which the average value of the normal distribution is continuously output within the state, and the state is switched. In addition, 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 Δx _i and the delta delta Δ ² x _i of the value x _i 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 log likelihood p (X) of X is given by the following equation. Here, _mi is an average vector of the distribution of o _i , and U _i is a variance covariance matrix of the distribution of o _i . m _i and U _i are obtained from the speech synthesis symbols by a decision tree learned in advance.

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

Solving Equation (11) for X yields:

  That is, by calculating Equation (12), a parameter time series in consideration of dynamic features based on the maximum likelihood criterion can be obtained.

"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

  In order to generate the parameter time series as described above, high calculation accuracy is required. However, the decision tree for predicting the parameter vector of the parameter distribution related to the fundamental frequency includes not only phonemic phoneme types but also differences in linguistic prosodic features such as accent types and accent phrase boundaries. In the distribution of feature vectors predicted by a learned decision tree, the distribution of elements related to delta features and delta-delta features is often smaller than the distribution of elements related to static features. The tendency to become appears. 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 fundamental frequency.

  For this reason, when generating time-series data of fundamental frequency parameters based on the maximum likelihood criterion for speech synthesis, not the absolute value of the fundamental frequency of speech used for decision tree learning, but its delta and delta-delta features are greatly evaluated. There is a tendency that a parameter time series of the fundamental frequency that can be reproduced is generated. Such a method is excellent for accurately reproducing short-time changes in the fundamental frequency, but is inferior for reproducing the same fundamental frequency distribution as the learning speech over a long period of time. This causes unnaturalness of the synthesized speech.

  Such a tendency is analyzed by using the log likelihood formula of the parameter time series to be generated, because the value of the term related to the static feature becomes extremely smaller than the value of the term related to the dynamic feature. To occur. However, in practice, if the generated parameters deviate from the learned fundamental frequency distribution, the logarithmic likelihood becomes small even if very small, so the fundamental frequency of the synthesized speech deviates extremely from the learned speech distribution. It is rare to end up.

  However, when computing with a device that has limited computing resources and requires fixed-point arithmetic, such as a mobile device, and cannot handle a sufficient range of values that can be processed (for example, the ratio between the maximum and minimum values), the likelihood calculation Due to the digit loss, the value of the term related to the static parameter can be zero regardless of the fundamental frequency. When the value of the term related to the static parameter becomes zero, the restriction on the static parameter is removed. Then, it is possible to generate basic frequency time-series data that is translated extremely up or down on the frequency axis as compared with the basic frequency of the learning speech.

  The present invention has been made in view of such circumstances, and speech synthesis that can reduce the deviation of the fundamental frequency distribution that occurs with the learning speech even when the calculation accuracy of the predicted fundamental frequency parameter is insufficient. An object is to provide a device, a speech synthesis method, and a speech synthesis program.

  (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 fundamental frequency time-series data generation unit that predicts and generates first fundamental frequency time-series data using distribution information of a first feature vector based on given speech synthesis information; A second fundamental frequency time-series data generation unit that predicts and generates second fundamental frequency time-series data using distribution information of a second feature vector based on given speech synthesis information; A basic frequency time-series data correction unit that corrects the first basic frequency time-series data using two basic frequency time-series data, and synthesis based on the corrected first basic frequency time-series data voice It is characterized by generating a shape.

  As described above, since the speech synthesizer of the present invention corrects the time series data of the first feature vector using the time series data of the second feature vector, the calculation accuracy of the predicted fundamental frequency parameter is poor. Even if it is sufficient, it is possible to reduce the deviation of the fundamental frequency distribution that occurs with the learning speech. As a result, speech synthesis close to learning speech can be realized even in a device having a processor that cannot perform floating point arithmetic, such as a portable terminal.

  (2) Further, in the speech synthesizer according to the present invention, the first basic frequency time-series data generation unit uses the distribution information of the first feature vector including dynamic features representing temporal changes as elements, The second fundamental frequency time-series data generating unit is characterized by using distribution information of a second feature vector including static features that do not directly represent temporal changes as elements. In this way, time-series data generated from feature vectors containing dynamic features representing temporal changes as elements is used, and time-series data generated from feature vectors containing static features that do not express temporal changes as elements. Therefore, it is possible to reduce the deviation of the fundamental frequency distribution that is likely to occur between the learning speech and the synthesized speech.

  (3) Further, in the speech synthesizer of the present invention, the fundamental frequency time-series data correction unit calculates an average value of the first fundamental frequency time-series data for the second fundamental frequency for each predetermined time-series section. It is characterized by matching the average value of time series data. Thereby, the shift in the frequency direction of the time series data of the basic frequency parameter can be corrected using the average value. As a result, it is possible to generate fundamental frequency time-series data that strongly reflects the output result of the second fundamental frequency distribution parameter generation unit on a long-term average.

  (4) Further, in the speech synthesizer according to the present invention, the fundamental frequency time-series data correction unit distributes the dispersion of the first fundamental frequency time-series data at the second fundamental frequency for each predetermined time-series section. It is characterized by matching the distribution of series data. Thereby, it is possible to correct the deviation of the frequency distribution of the time-series data of the basic frequency parameter using the variance. As a result, it is possible to synthesize a voice so that the minimum and maximum values of parameter fluctuation within one utterance unit are close to that of the learning voice and the learning voice is reproduced more accurately.

  (5) Further, the speech synthesizer of the present invention is characterized in that the basic frequency time-series data correction unit corrects the first basic frequency time-series data for each time-series section in which voice is continuous. . Thereby, the fundamental frequency time series data that strongly reflects the output result of the second fundamental frequency distribution parameter generation unit 133 can be generated in shorter time units.

  (6) The speech synthesis method of the present invention is a speech synthesis method for generating a synthesized speech waveform from speech synthesis information configured as a set of phonemes, and is a first synthesis method based on given speech synthesis information. Predicting and generating first basic frequency time-series data using feature vector distribution information, and using second feature vector distribution information based on the given speech synthesis information, Predicting and generating two fundamental frequency time-series data, and modifying the first fundamental frequency time-series data using the second fundamental frequency time-series data, the modified A synthesized speech waveform based on the first basic frequency time-series data is generated. Thereby, even when the calculation accuracy of the predicted fundamental frequency parameter is insufficient, the deviation of the fundamental frequency distribution that occurs with the learning speech 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 first feature vector based on the given speech synthesis information, a process for predicting and generating the first basic frequency time-series data, and based on the given speech synthesis information The second basic frequency time-series data is predicted and generated using the distribution information of the second feature vector, and the second basic frequency time-series data is used to generate the first basic frequency time-series data. Including a process of correcting the sequence data, and generating a synthesized speech waveform based on the corrected first basic frequency time-series data. Thereby, even when the calculation accuracy of the predicted fundamental frequency parameter is insufficient, the deviation of the fundamental frequency distribution that occurs with the learning speech can be suppressed.

  According to the present invention, even when the calculation accuracy of the predicted fundamental frequency parameter is insufficient, the deviation of the fundamental frequency distribution that occurs with the learning speech can be reduced.

1 is a block diagram of a speech synthesizer according to the present invention. It is a flowchart which shows an example of operation | movement of the speech synthesizer which concerns on this invention.

  An embodiment of the present invention will be described. 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. In the following, all fundamental frequencies are handled on the logarithmic fundamental frequency axis.

[First Embodiment]
(Functional configuration of speech synthesizer)
An example of a functional configuration of the speech synthesizer 100 will be described with reference to the drawings. FIG. 1 is a block diagram of the speech synthesizer 100. The speech synthesizer 100 generates a synthesized speech waveform from speech synthesis information such as a speech synthesis symbol. As shown in FIG. 1, the speech synthesizer 100 includes a state duration sequence generator 110, a spectrum feature distribution parameter generator 120, a fundamental frequency time series data generator module 130, a spectrum feature time series data generator 140, and a speech waveform. A generation unit 150 is provided.

  The state duration sequence generator 110 predicts the duration (number of frames) of each state constituting the speech unit for each speech unit corresponding to the speech synthesis symbol string, collects one utterance unit, Output as a continuous length sequence. On the other hand, the spectral feature distribution parameter generation unit 120 generates a spectral feature distribution parameter for each state in the state sequence corresponding to the speech synthesis symbol string.

  The voiced / unvoiced parameter generation unit 131 generates a voiced / unvoiced probability parameter in the state sequence corresponding to the speech synthesis symbol string. The first fundamental frequency distribution parameter generation unit 132 generates a fundamental frequency feature distribution parameter for each state in the state sequence corresponding to the speech synthesis symbol string. The first fundamental frequency distribution parameter generation unit 132 is a predictor that predicts a fundamental frequency distribution parameter using learning speech.

  On the other hand, the second fundamental frequency distribution parameter generation unit 133 similarly generates a fundamental frequency feature distribution parameter of each state in the state sequence corresponding to the speech synthesis symbol string. The spectral feature and the fundamental frequency feature include not only static features but also dynamic features such as delta features and delta-delta features. The second fundamental frequency distribution parameter generation unit 133 is a predictor that predicts a fundamental frequency distribution parameter using learning speech.

  Note that all the above features assume a normal distribution, and the distribution parameters are each composed of an average and a variance. 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.

  Next, the spectrum feature time series data generation unit 140 generates time series data of a spectrum feature vector from the state duration sequence and the spectrum feature distribution parameter for each state. In this generation process, it is assumed that the distribution of each characteristic parameter in each state continues for the number of frames stored in the state duration length sequence, and the parameter time series that maximizes the likelihood seen in one utterance unit. It is done by asking for. This can be obtained by calculating equation (12).

  Similarly, the first fundamental frequency time-series data generation unit 134 calculates the fundamental frequency parameter from the state duration length series and the fundamental frequency feature distribution parameter for each state output by the first fundamental frequency distribution parameter generation unit 132. Time series data (basic frequency time series data) is generated. However, with respect to the fundamental frequency, the parameter time-series data generation processing is performed only for the voiced state with reference to voiced / unvoiced information for each state. In this way, the first basic frequency time-series data is predicted and generated using the distribution information of the first feature vector based on the given speech synthesis information. As the first feature vector, a vector including a dynamic feature representing a temporal change as an element is used.

  In this process, since the delta feature or the delta delta feature cannot be defined for a voiced frame adjacent to an unvoiced frame, the distribution value of the distribution of the delta feature and the delta delta feature is set to infinity in such a frame. . As a result, when the mathematical expression (12) is expanded and actually calculated, the delta features of those frames and the terms related to the delta-delta features become 0, and the dynamic features of such frames may be ignored in the calculation. it can. In the following description, it is assumed that voiced / unvoiced information is also embedded in the basic frequency time-series data by setting an appropriate special value for the basic frequency data even in the case of unvoiced and treating it as a voiceless code. . As the “special value”, 0, for example, 65535, which is the maximum value when the fundamental frequency is represented by a 16-bit unsigned integer value, can be set.

  Similarly, the second fundamental frequency time-series data generation unit 135 uses the above-mentioned state duration sequence and the fundamental frequency feature distribution parameter for each state output by the second fundamental frequency distribution parameter generation unit 133 to determine the time of the fundamental frequency parameter. Generate series data. However, unlike the first basic frequency time-series data generation unit 134, dynamic characteristics are not considered here, so that the generated basic frequency time-series data takes a constant value in the state and every time the state is switched. Stepwise time-series data with discontinuous values. In this way, the basic frequency time-series data is predicted and generated using the distribution information of the second feature vector based on the given information for speech synthesis. As the second feature vector serving as a long-term feature reference, a vector having a static feature that does not represent a temporal change in the fundamental frequency as an element is used.

  On the other hand, unlike the first basic frequency time-series data generation unit 134, since there is no influence of distribution dispersion and digits are not easily lost, a relatively accurate output result can be obtained in terms of calculation accuracy. It is assumed that voiced / unvoiced information is also embedded in the output result in second fundamental frequency time-series data generation unit 135 as well as first fundamental frequency time-series data generation unit 134.

  The fundamental frequency time series data correction unit 136 converts the fundamental frequency time series data (first fundamental frequency time series data) generated by the first fundamental frequency time series data generation unit 134 into the second fundamental frequency time series data. Correction is performed using the basic frequency time-series data (second basic frequency time-series data) generated by the generation unit 135.

  Thereby, even when the calculation accuracy of the predicted fundamental frequency parameter is insufficient, the deviation of the fundamental frequency distribution that occurs with the learning speech can be reduced. As a result, it is possible to realize speech synthesis close to learning speech even in a processor that is not capable of floating-point arithmetic, which is mounted on a portable terminal or white goods. In particular, fundamental frequency time-series data generated from feature vectors containing dynamic features representing temporal changes as elements, and fundamental frequency time-series data generated from vectors containing static features that do not represent temporal changes as elements. When using and correcting, it is possible to reduce the deviation of the fundamental frequency distribution that easily occurs between the learning speech and the synthesized speech.

  In the correction process, first, an average value of only a predetermined time series section is obtained for the basic frequency time series data generated by the second basic frequency time series data generation unit. Next, the average value of the basic frequency time-series data generated by the first basic frequency time-series data generating unit 134 is equal to the average value of the output result of the second basic frequency distribution parameter generating unit 133 obtained previously. Thus, a certain constant is added to all elements of the basic frequency time-series data, and the result of addition is output.

  Thereby, the shift in the frequency direction of the time series data of the basic frequency parameter can be corrected using the average value. As a result, it is possible to generate fundamental frequency time-series data that strongly reflects the output result of the second fundamental frequency distribution parameter generation unit 133 on a long-term average. The predetermined time series section is, for example, a voiced section.

  The voiced / unvoiced parameter generation unit 131, the first fundamental frequency distribution parameter generation unit 132, the second fundamental frequency distribution parameter generation unit 133, the first fundamental frequency time series data generation unit 134, and the second fundamental frequency time The sequence data generation unit 135 and the basic frequency time series data correction unit 136 constitute a basic frequency time series data generation module 130.

  The audio waveform generation unit 150 includes a sound source 151 and an articulation filter 152. The speech waveform generator 150 synthesizes a speech waveform by driving the articulation filter 152 controlled based on the spectral feature time-series data with the sound source waveform output from the sound source 151 based on the modified basic frequency time-series data. In this way, a synthesized speech waveform based on the time series data of the modified first feature vector is generated.

(Operation of speech synthesizer)
Next, an example of the operation of the speech synthesizer 100 will be described. FIG. 2 is a flowchart showing an example of the operation of the speech synthesizer 100. Hereinafter, a flow of processing for generating a synthesized speech waveform from speech synthesis symbols will be described in order.

  As shown in FIG. 2, first, the speech synthesizer 100 generates a state duration sequence based on the input speech synthesis symbol (step S1). Next, a spectrum feature distribution parameter is generated based on the speech synthesis symbol (step S2). Then, using the generated state duration sequence and the spectrum feature distribution parameter, time series data of the spectrum feature parameter is generated (step S3). Also, voiced / unvoiced parameters are generated based on the speech synthesis symbols (step S4).

  Then, based on the speech synthesis symbol, a first fundamental frequency distribution parameter is generated (step S5). Using the generated first fundamental frequency distribution parameter, state duration sequence, and voiced / unvoiced parameter, time series data of the first fundamental frequency parameter is generated based on the speech synthesis symbol (step S6).

  On the other hand, a second fundamental frequency distribution parameter is generated based on the speech synthesis symbol (step S7). Then, using the generated second fundamental frequency distribution parameter, state duration sequence, and voiced / unvoiced parameter, time-series data of the second fundamental frequency parameter is generated based on the speech synthesis symbol (step S8). Then, the time series data of the first fundamental frequency parameter is corrected using the time series data of the second fundamental frequency parameter (step S9). Finally, a speech waveform is generated using the time-series data of the spectrum feature parameters and the time-series data of the modified first basic frequency parameter (step S10). In this way, a synthesized speech waveform can be output from the input of a speech synthesis symbol.

[Second Embodiment]
In the above embodiment, the basic frequency parameter time-series data correction unit 136 performs only translation on the frequency axis, but correction other than translation may be performed. For example, for the output of the second fundamental frequency distribution parameter generation unit 133, the variance is calculated in addition to the average value, and the variance of the fundamental frequency time series data output by the fundamental frequency parameter time series data correction unit 136 is The characteristic parameters may be expanded or contracted in the frequency axis direction so as to be equal to the variance of the output of the second fundamental frequency distribution parameter generation unit 133. Thereby, the minimum and maximum values of parameter fluctuation within one utterance unit can be brought close to that of the learning voice, and the voice can be synthesized so as to reproduce the learning voice more accurately.

[Third Embodiment]
In the above embodiment, the first fundamental frequency distribution parameter generation unit 132 outputs the static feature distribution and the dynamic feature distribution parameters of the fundamental frequency at the same time. However, the first fundamental frequency distribution parameter generation unit 132 outputs only the dynamic feature amount. It is good. In this case, the first fundamental frequency distribution parameter generation unit 132 needs some restriction on the static feature quantity distribution or the initial value restriction of the fundamental frequency. However, it may be some suitable constant distribution or value. The fundamental frequency time series data generated by the first fundamental frequency time series data generation unit 134 is finally based on the output of the second fundamental frequency distribution parameter generation unit 132 in the fundamental frequency parameter time series data correction unit 136. The frequency range is corrected to an appropriate frequency range. In this case, the predictor learning of the first fundamental frequency distribution parameter generation unit 132 may refer to only dynamic features, or may simultaneously refer to static features and dynamic features as in the conventional case.

[Fourth Embodiment]
In the above embodiment, only the static feature distribution is output by the second fundamental frequency distribution parameter generation unit 133 and is used by the second fundamental frequency time-series data generation unit. If the accuracy of the long-time feature of the output of the second fundamental frequency time-series data generation unit 135 is higher than the accuracy of the long-time feature of the output of the fundamental frequency time-series data generation unit 134, the first The distribution of other features may be used for generating the basic frequency time-series data of 2. It may also take into account two or more dimension feature vectors. In that case, the process of the second fundamental frequency time series data generation unit is the same as the process in the first fundamental frequency time series data generation unit 134.

[Fifth Embodiment]
In the above-described embodiment, the control unit is one utterance unit. However, for example, the correction amount may be independently obtained and corrected in a shorter time unit with a voiced continuous state as a unit. Thereby, the fundamental frequency time series data that strongly reflects the output result of the second fundamental frequency distribution parameter generation unit 133 can be generated in shorter time units. However, it is necessary to increase the accuracy of the output of the second fundamental frequency distribution parameter generation unit 133 as compared with the case where correction is made in units of one utterance as the interval for obtaining the average becomes shorter.

[Sixth Embodiment]
For static features, an average value between several frames before and after may be used instead in order to remove short-term noise. The feature vector including the dynamic feature is not limited to a three-dimensional feature vector as long as the feature vector can be determined by the basic frequency time-series data generation module 130. For example, a two-dimensional vector composed of static features and delta features, or a high-dimensional feature vector including higher-order features may be used.

[Seventh Embodiment]
In the above embodiment, the discussion on the logarithmic frequency axis is performed, but the fundamental frequency parameter is not limited to this. For example, when only a sufficiently narrow range of fundamental frequencies is handled, such as limiting speakers, the same operation may be performed on the linear frequency axis.

[Eighth Embodiment]
In the above embodiment, an example of speech synthesis based on a sound source / filter model is shown. However, only a part related to generation of basic frequency time-series data may be used in combination with another speech synthesis method. For example, in connection synthesis using fundamental frequency time series data as an explicit speech synthesis target, the present invention can be used as a predictor of fundamental frequency time series data.

  In the above embodiment, the speech synthesizer 100 has two fundamental frequency distribution parameter generation units as predictors, but may have three or more. However, since the data size to be processed increases in proportion to the number of predictors, it is preferable that there are two predictors.

DESCRIPTION OF SYMBOLS 100 Speech synthesizer 110 State continuation length series generation part 120 Spectral feature distribution parameter generation part 130 Fundamental frequency time series data generation module 131 Voiced / unvoiced parameter generation part 132 1st fundamental frequency distribution parameter generation part 133 2nd fundamental frequency distribution Parameter generating unit 134 First basic frequency time-series data generating unit 135 Second basic frequency time-series data generating unit 136 Basic frequency time-series data correcting unit 140 Spectrum feature time-series data generating unit 150 Speech waveform generating unit 151 Sound source 152 Articulation filter

Claims (7)

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,
-Out based on the given speech synthesis information using the distribution information of the first feature vector including dynamic characteristic representing a temporal change as elements, generated by predicting a first fundamental frequency time-series data A first fundamental frequency time-series data generator;
-Out based on the given speech synthesis information using the distribution information of the second feature vector comprising a static feature which does not represent a direct temporal change as an element, a second fundamental frequency time-series data A second fundamental frequency time-series data generation unit that predicts and generates
Using the second fundamental frequency time-series data, a fundamental frequency time-series data correction unit that corrects a frequency shift of the frequency of the first fundamental frequency time-series data,
A speech synthesizer for generating a synthesized speech waveform based on the modified first basic frequency time-series data. The first basic frequency time-series data generation unit uses distribution information of a first feature vector including dynamic features representing temporal changes as elements,
The second basic frequency time-series data generation unit uses distribution information of a second feature vector including a static feature that does not directly represent a temporal change as an element. Speech synthesizer.   The fundamental frequency time-series data correction unit matches an average value of the first fundamental frequency time-series data with an average value of the second fundamental frequency time-series data for each predetermined time-series section. The speech synthesizer according to claim 1 or 2.   The fundamental frequency time-series data correction unit matches the variance of the first fundamental frequency time-series data with the variance of the second fundamental frequency time-series data for each predetermined time-series section. Item 4. The speech synthesizer according to Item 3.   The said basic frequency time series data correction part corrects said 1st basic frequency time series data for every time series area where voiced continues, The one in any one of Claims 1-4 characterized by the above-mentioned. Speech synthesizer. A speech synthesis method for generating a synthesized speech waveform from speech synthesis information configured as a set of phonemes,
-Out based on the given speech synthesis information using the distribution information of the first feature vector including dynamic characteristic representing a temporal change as elements, generated by predicting a first fundamental frequency time-series data Steps,
-Out based on the given speech synthesis information using the distribution information of the second feature vector comprising a static feature which does not represent a direct temporal change as an element, a second fundamental frequency time-series data Predicting and generating
Using the second fundamental frequency time-series data to correct a shift in the frequency direction of the first fundamental frequency time-series data,
A speech synthesis method, comprising: generating a synthesized speech waveform based on the modified first basic frequency 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,
-Out based on the given speech synthesis information using the distribution information of the first feature vector including dynamic characteristic representing a temporal change as elements, generated by predicting a first fundamental frequency time-series data Processing,
-Out based on the given speech synthesis information using the distribution information of the second feature vector comprising a static feature which does not represent a direct temporal change as an element, a second fundamental frequency time-series data Processing to predict and generate
Using the second basic frequency time-series data to correct a frequency shift in the frequency direction of the first basic frequency time-series data,
A speech synthesis program for generating a synthesized speech waveform based on the modified first basic frequency time-series data.

Images