JP5763414B2 - Feature parameter generation device, feature parameter generation method, and feature parameter generation program - Google Patents

Description

  The present invention relates to a feature parameter generation apparatus, a feature parameter generation method, and a feature parameter generation program that reproduce a time-series distribution of feature parameters corresponding to symbol series information.

  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 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 parameter time series (time series data of feature parameters) into 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, when the statistical distribution of the value xi of the speech synthesis parameter x in frame i is independent of other types of speech synthesis parameters and follows a normal distribution, the average of the distribution is μi, and the variance is σi2, the length of the speech Is a total of 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 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 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 explanation, the feature vector in frame i is assumed to be oi. 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, 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 log likelihood p (X) satisfies the following relationship.

Solving Equation (8) and Equation (9) with respect to X yields the following equation:

  That is, by calculating Equation (10), a parameter time series in consideration of dynamic features based on the maximum likelihood criterion can be obtained. The same applies when the speech synthesis parameter x is extended to a multidimensional vector. The configuration of such a processing apparatus is expressed as shown in FIG. 3, for example.

  However, the parameter time series obtained by such a method often gathers closer to the average value side of the entire series than the data from which the distribution information is based, and tends to be a parameter time series with a small dynamic range. One reason for this is that the feature parameter time series generation processing does not include restrictions on the distribution of values when viewed as a whole series. For this reason, it is conceivable that a large error occurs in the distribution when viewed as a whole series between the data used for the distribution estimation and the obtained data.

Therefore, a method is known in which the variance of the entire sequence is constrained with respect to the variance of the values of the entire sequence to obtain the maximum likelihood parameter time series (see Non-Patent Document 5). ). In this method, instead of Equation (8), the log likelihood given by the following equation is maximized. Note that the distribution of dispersion refers to the distribution of dispersion values with respect to conditions.

  However, p (X) is the same as the equation given by Equation (8), and pv (X) in Equation (11) is the likelihood of X of the variance distribution of the elements of X. The likelihood of. Further, ω is a weighting coefficient that adjusts the ratio of the effect of the likelihood calculated from the distribution of the variance of the elements of X and the likelihood calculated from the conventional distribution of X at time i. Thereby, more preferable parameter time-series data in which not only the feature distribution at each time but also the dynamic range of the entire series is reproduced is obtained.

  The speech synthesis technique has been described above as an example, but the parameter time series generation technique can be applied to other uses. For example, instead of acoustic features, feature parameters related to the mouth shape are modeled in the same way, and parameter time series is generated by this technology, so that a smooth mouth can be obtained from a small amount of data compressed in the time axis direction as well as speech. Can be generated (see Non-Patent Document 6). Alternatively, the input information is not phonemes, but discrete symbol information about movement (for example, “extends the index finger from a closed hand”), and modeling is performed in combination with feature quantities related to movement. It can also be used to generate natural motion data such as animation characters and robots (see Non-Patent Document 7).

"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 Toda Tomoki, Tokuda Keiichi, “Speech parameter generation algorithm considering intra-sequence variation for HMM speech synthesis”, IEICE Technical Report, SP2005-52, Aug. 2005, pp. 1-6. Masashi, Masanori Tamura, Kiminari Miyagawa, Takashi Masuko, Takao Kobayashi, Keiichi Tokuda “Simultaneous Generation of Speech and Lip Images from Text Based on HMM”, Proceedings of the 1998 Acoustical Research Conference of the Acoustical Society of Japan, 2- P-6, vol. 1, Mar. 1998., pp. 305-306 Kenji Mori, Yoshihiko Nankaku, Chiyomi Miyajima, Keiichi Tokuda, Tadashi Kitamura “Finger Text Video Generation Based on Hidden Markov Models” FIT2005 (4th Forum on Information Science and Technology), K-092, Aug. 2005, pp. 569-570 .

  Unlike the case where such a restriction is not provided, the parameter time-series data having the maximum likelihood cannot be directly calculated when the restriction on the element value distribution of the characteristic parameter time series data to be generated is provided. Therefore, it is necessary to obtain approximately by an iterative method such as Newton-Raphson method. In such an iterative method, the likelihood value of the parameter time-series data changes when the generated data in the parameter time-series changes even at one point. Thereby, for example, in each process repeated by the Newton-Raphson method, each off-diagonal element of the Hessian matrix of the likelihood function related to variance (a matrix having the second partial differential of the objective function as an element) does not become zero. That is, the Hessian matrix is not a diagonal matrix. However, in each iteration of the above iterative process, it is necessary to obtain the inverse matrix of the Hessian matrix, and the cost increases due to the inverse matrix calculation of the non-diagonal matrix.

  On the other hand, a quasi-Newton method based on the BFGS formula or the like is known as a method that can avoid the inverse matrix operation. However, in the quasi-Newton method, the size of the row and column of the matrix to be calculated is proportional to the number of frames of the speech to be synthesized, and requires a memory proportional to the square thereof. It cannot be processed at high speed in a limited environment.

  On the other hand, as another method for facilitating the inverse matrix calculation, there is a method in which all the off-diagonal elements of the Hessian matrix in the Newton-Raphson method are set to 0 as attempted in Non-Patent Document 5. In this case, the memory usage can be reduced and the inverse matrix calculation becomes easy. However, setting all the off-diagonal elements to 0 is equivalent to assuming that the data at each time is independent. To do. However, since dynamic features are determined by data at a plurality of times and are not feature information independent at each time, an appropriate solution cannot be corrected by such a method. As a result, solution convergence becomes very slow, and the processing speed cannot be increased.

  In these methods, processing that is completely different from processing that does not consider the distribution of the time-series data to be obtained is performed, so that a new circuit or software is required. In fact, even when the variance of the time series data to be obtained is taken into account, the result is somewhat similar to the result when it is not taken into account. By doing so, solution convergence can be accelerated. However, in that case, as shown in FIG. 4, the circuit and software of the conventional method and the iterative method are required, and cannot be adopted in a device having insufficient processing ability or capacity such as a portable terminal or an embedded device.

  The present invention has been made in view of such circumstances, a feature parameter generation device, a feature parameter generation method, and a feature parameter generation program that take into account the variance of feature parameter time series while suppressing an increase in circuit / software scale. The purpose is to provide.

  (1) In order to achieve the above object, a feature parameter generation device of the present invention is a feature parameter generation device that reproduces a time series distribution of feature parameters corresponding to symbol series information, and the feature parameter distribution is In order to obtain maximum likelihood, the initial value series data is modified to generate feature parameter time series data, and the likelihood of the distribution of the feature parameter distribution and the feature parameter time series data is further increased. Of the configurations for repeatedly correcting the time-series data of the generated characteristic parameters, the initial value correction and the repeated correction are performed by the same configuration.

  As described above, the feature parameter generation device of the present invention is excellent in the reproducibility of feature parameters because it takes into account the dispersion of time series data of feature parameters. Then, since the parameter time-series data generation result with the maximum likelihood of the feature parameter distribution is repeatedly corrected, the convergence can be further accelerated. In addition, since the initial value correction and the repeated correction are performed with the same configuration, it is not necessary to provide redundant circuits and software, and the circuit scale and software scale can be reduced. Therefore, it is particularly advantageous for portable terminals and embedded devices.

  (2) Further, the feature parameter generation apparatus of the present invention includes a first intermediate data generation unit that generates an initial intermediate data sequence from a symbol information sequence so that the distribution of the feature parameter is maximum likelihood, and the initial use A correction data generation unit that generates correction series data from the intermediate data series, a data correction unit that corrects the time series data of feature parameters corresponding to the symbol information series using the correction series data, and feature parameters A second intermediate data generation unit that generates an intermediate data sequence for updating from the time series data of the modified feature parameter so that the likelihood of the distribution of the distribution and the time series data of the feature parameter becomes larger; The correction data generation unit generates correction sequence data from the update intermediate data sequence, and the data correction unit generates the correction sequence data. It is characterized by repeating a series of operations to correct the time-series data of said modified feature parameter using the data.

  As a result, the correction data generation unit operates in both cases of initial correction and repeated correction of the time series data of the feature parameter, so that it is not necessary to provide redundant circuits and software, and the circuit scale・ The software scale can be reduced.

  (3) In addition, the feature parameter generation device of the present invention is further characterized by further including a processing control unit that repeatedly corrects the time series data of the corrected feature parameters until a predetermined condition is satisfied. Thereby, the time-series data of the characteristic parameters can be converged to a certain reproduction level.

  (4) In the feature parameter generation device of the present invention, the time series distribution of the feature parameters includes a time series dynamic feature distribution of the feature parameters, and the first intermediate data generation unit includes the dynamic features. An intermediate data sequence is generated from the symbol information sequence so that the distribution of the feature parameter including the maximum likelihood, and the second intermediate data generation unit includes the distribution of the feature parameter including the dynamic feature and the time series data of the feature parameter. An intermediate data sequence is generated from the symbol information sequence so that the likelihood of dispersion is greater. Thereby, the reproducibility of the time series data of the feature parameter can be improved in consideration of the dynamic feature.

  (5) Further, in the feature parameter generation device of the present invention, the second intermediate data generation unit includes a weighted sum of the log likelihood of the distribution of the feature parameter and the log likelihood of the variance of the time series data of the feature parameter. Is used to generate an intermediate data series for update. In this way, it is possible to adjust the reproducibility when reproducing the variance of the time series data of the characteristic parameters by the weight.

  (6) In the feature parameter generation device of the present invention, the second intermediate data generation unit has a condition on the assumption that the likelihood of variance of the time series data of the feature parameter does not change before and after the correction. In the following, the update intermediate data series is generated. Thereby, the processing load can be reduced by simplifying the calculation of the correction of the time series data of the characteristic parameters.

  (7) In the feature parameter generation device of the present invention, the correction data generation unit generates correction sequence data by an operation equivalent to calculating a product of an inverse matrix for a symmetric band matrix and a vector. It is characterized by. Thereby, the processing load can be reduced by simplifying the calculation of the correction of the time series data of the characteristic parameters.

  (8) Also, the feature parameter generation device of the present invention inputs speech synthesis information describing the type of unit speech included in a series of unit speech sequences as the symbol information sequence, and the modified feature parameter A feature is that a synthesized speech waveform is generated as sequence data. A voice waveform can be generated with an appropriate dynamic range by inputting a symbol describing the type of phoneme and prosodic features.

  (9) The feature parameter generation method of the present invention is a feature parameter generation method that reproduces a time-series distribution of feature parameters corresponding to a symbol information sequence by a computer, and the feature parameter distribution is the maximum likelihood. A step of generating an initial intermediate data sequence from the symbol information sequence, a step of generating correction sequence data from the initial intermediate data sequence, and a time series of feature parameters corresponding to the symbol information sequence Correcting the data using the correction series data; and from the time series data of the modified feature parameters so that the likelihood of the distribution of the feature parameters and the variance of the time series data of the feature parameters is increased. Generating an intermediate data series for update, wherein the intermediate data series for update Generates a corrective series data, is characterized by repeated sequence of steps to modify the time series data of said modified feature parameter. In this way, it is possible to reproduce the feature parameter in consideration of the variance of the feature parameter time series while suppressing an increase in circuit / software scale.

  (10) A feature parameter generation program of the present invention is a feature parameter generation program that causes a computer to execute and reproduce a time-series distribution of feature parameters corresponding to a symbol information series, wherein the feature parameter distribution is the maximum likelihood. Processing for generating an initial intermediate data sequence from the symbol information sequence, processing for generating correction sequence data from the initial intermediate data sequence, and a time series of feature parameters corresponding to the symbol information sequence From the time series data of the modified feature parameter so that the likelihood of the process of modifying the data using the correction series data and the distribution of the feature parameter distribution and the time series data of the feature parameter is increased. Generating intermediate data series for update, and correcting from the intermediate data series for update To generate a sequence data is characterized by repeating a series of processing for correcting the time series data of said modified feature parameter. In this way, it is possible to reproduce the feature parameter in consideration of the variance of the feature parameter time series while suppressing an increase in circuit / software scale.

  The time series data of the characteristic parameters can be reproduced in consideration of dispersion while suppressing the entire circuit and program scale. In addition, by using a device or program that has been speeded up for processing in an iterative method, the calculation processing can be easily speeded up.

It is a block diagram which shows the characteristic parameter production | generation apparatus which concerns on this invention. It is a flowchart which shows operation | movement of the characteristic parameter generation apparatus which concerns on this invention. It is a block diagram which shows the characteristic parameter production | generation apparatus of the conventional method. It is a block diagram which shows the characteristic parameter production | generation apparatus which combined the conventional method and the iterative method.

(Features of feature parameter generator)
The feature parameter generation device of the present invention reproduces a time series distribution of feature parameters corresponding to symbol series information. For example, speech synthesis information describing the type of unit speech included in a series of unit speech sequences is input as a symbol information sequence, and a synthesized speech waveform is generated as time-series data of the modified feature parameters. Then, it is possible to input a symbol describing the type of phoneme and prosodic features and generate a speech waveform with an appropriate dynamic range.

  In addition, instead of the acoustic feature quantity, a feature parameter related to the mouth shape can be similarly modeled, a parameter time series can be generated, and a moving image of mouth movement can be generated. It is also possible to model feature quantities and generate natural motion data such as animation characters and robots.

  Such a feature parameter generation device corrects initial value series data to generate feature parameter time series data so that the distribution of the feature parameters becomes maximum likelihood. Then, in the configuration in which the time series data of the generated feature parameters is repeatedly corrected so that the likelihood of the distribution of the feature parameters and the distribution of the time series data of the feature parameters is increased, the correction of the initial value and the correction of the repetition are performed. Are performed with the same configuration. The maximum likelihood means the maximization of the likelihood function of the distribution, and the likelihood function is a function that is defined as the distribution is more likely as the value is larger. Examples of such functions are Gaussian functions and mixed Gaussian functions defined by linear weighted sums of Gaussian functions.

  Thereby, since the dispersion of the time series data of the characteristic parameters is taken into consideration, the reproducibility of the characteristic parameters is excellent. Then, since iterative correction is performed using the parameter time-series data generation result with which the distribution is likely, the convergence can be further accelerated. In addition, since the initial value correction and the repeated correction are performed with the same configuration, it is not necessary to provide redundant circuits and software, and the circuit scale and software scale can be reduced. Therefore, it is particularly advantageous for portable terminals and embedded devices.

  As an example, the feature parameter generation device generates feature parameter time series data that maximizes the weighted sum of the log likelihoods of the feature parameter distribution at each time and the distribution of the variance of the time series data of the feature parameters. Output. Thereby, feature parameter time-series data in which both standards are balanced is generated. Details of the present invention will be described below with reference to the drawings.

(Configuration of feature parameter generator)
FIG. 1 is a block diagram of the feature parameter generation device 100. The feature parameter generation apparatus 100 includes a first intermediate data generation unit 110, a correction data generation unit 120, a data correction unit 130, a second intermediate data generation unit 140, and a processing control unit 150.

  Hereinafter, for the sake of simplicity, the case where the data at each time in the feature parameter time series data is a one-dimensional value will be described. The length of the time series data is n.

The first intermediate data generation unit 110 generates an initial intermediate data sequence from the symbol information sequence so that the distribution of the characteristic parameters becomes maximum likelihood. That is, the first intermediate data generation unit 110 applies the n × n symmetric matrix shown in the following equations (12) and (13) of the final output to the feature parameter distribution information at each time and n The calculation for obtaining the dimension vector is performed, and the result is output as intermediate data for correcting the feature parameter distribution for the first time.

  This process is the same as when the time series variance of feature parameters is not considered. The feature parameter distribution information is obtained and input based on the symbol series information. “Generating an intermediate data sequence for initial use so as to have maximum likelihood” includes, for example, calculating an intermediate data sequence for initial use using a mathematical formula created on the basis of maximum likelihood as described above. The generation of the initial intermediate data series using some maximum likelihood criterion.

The correction data generation unit 120 generates time series data for correcting the feature parameter. Specifically, when an n × n symmetric matrix (here, S) and an n-dimensional vector (x) are input, an n-vector dimensional vector shown in the following equation (14) is calculated. Output.

  Therefore, correction sequence data is generated from the initial intermediate data sequence at the first correction, and correction sequence data is generated from the update intermediate data sequence at the second and subsequent corrections. In both the first and second corrections, the correction data generation unit 120 operates, so that it is not necessary to provide redundant circuits and software, and the circuit scale and software scale can be reduced.

  The data correction unit 130 receives time series data and correction data of a certain feature parameter as input, and outputs time series data obtained by adding the parameter values at each time as a correction result of the time series data of the feature parameter. . At this time, the correction data may be multiplied by α (0 <α) as a multiplier corresponding to the step parameter in the Newton-Raphson method. If α is large, the possibility of a worsening of the result due to the correction is increased instead of faster correction. If α is small, the possibility of the deterioration of the result due to the correction is reduced, but the convergence is delayed. Therefore, it is preferable to set α having an appropriate size.

  Specifically, the correction data generation unit 130 generates correction sequence data by an operation equivalent to calculating a product of an inverse matrix for a symmetric band matrix and a vector. Note that “corresponding” is not limited to the above-described calculation method, but means that a calculation method that does not actually obtain an inverse matrix explicitly is included when calculating the product of an inverse matrix of a symmetric matrix and a vector. ing. That is, for example, a method called a modified Cholesky decomposition is used to decompose a symmetric band matrix into a product of a lower triangular matrix, a diagonal matrix, and an upper triangular matrix, and calculate the product of each inverse matrix and vector in order.

  With the function as described above, the data correction unit 130 corrects the time series data of the characteristic parameters corresponding to the symbol information series at the time of the first correction using the correction series data. At the time of update in repeated correction, the time series data of the corrected feature parameter is further corrected using the correction series data.

The second intermediate data generation unit 140 generates an update intermediate data sequence from the modified feature parameter time-series data so that the likelihood of the distribution of the feature parameters and the distribution of the feature parameter time-series data is increased. Generate. Specifically, the second intermediate data generation unit 140 uses the Hessian matrix (second-order partial differentiation of the reference value based on the characteristic parameter at each time as an element when the first predetermined reference is maximization of a reference value. N × n symmetric matrix) and an n-dimensional vector whose elements are first-order partial differentials of reference values according to feature parameters at each time. In response to the input of the n-dimensional vector, the correction data generation unit 120 outputs a vector for correcting the time series data of the feature parameter in the Newton-Raphson method. This vector is expressed by the following equation (15), where H is a Hessian matrix and n-dimensional vector whose element is a first-order partial differential of a reference value according to a feature parameter at each time is ∂p '(X) / ∂X. Given. When −H = S and ∂p ′ (X) / ∂X = X, Equation (15) matches Equation (14), so that the correction data generation unit 120 can calculate this vector. it can.

  However, the matrix of Expression (12) is usually a symmetric band matrix. If the modified data generation unit 120 uses an apparatus, a program, or the like that does not consider the distribution of time series data of feature parameters, the input of the n-dimensional vector to the apparatus, the program, etc. is the symmetric band matrix of Equation (12). If the processing is premised on the shape, the correction data generation unit 120 cannot perform the calculation processing of Expression (15) as it is.

  On the other hand, if it is assumed that the likelihood value for the variance of the time series data of the feature parameter does not change before and after the processing in the modified data generation unit 120, the first-order partial coefficient of the logarithmic likelihood function of the variance is 0. It becomes. That is, the Hessian matrix for the log likelihood function of the variance of the time series data of the feature parameters is a diagonal matrix.

  In particular, as the iterative process proceeds, this assumption becomes more reasonable because the modification of the feature parameter becomes smaller and the change in the variance value becomes smaller. Therefore, the Hessian matrix relating to the entire log likelihood function, which is one of the outputs of the corrected data generation unit 120, is the weighted sum of the symmetric band matrix and the diagonal matrix having the same form as the formula (12), that is, the formula (12 ) In the same form as () (a matrix in which the row and column elements that are always 0 in Equation 12 are always 0).

  As described above, the second intermediate data generation unit 140 generates the intermediate data series for update under the condition that the likelihood of the variance of the time series data of the feature parameters does not change before and after the correction. Thereby, the processing load can be reduced by simplifying the calculation of the correction of the time series data of the characteristic parameters.

  The second intermediate data generation unit 140 can also perform processing without making assumptions regarding the distribution of time-series data of feature parameters. In that case, for the Hessian matrix for the log-likelihood function of variance, a matrix obtained by replacing some elements with 0 so that the symmetric band matrix has the same form as Equation (12) is output. it can.

  As described above, the second intermediate data generation unit 140 uses the weighted sum of the log likelihood of the feature parameter distribution and the log likelihood of the variance of the time series data of the feature parameters to update the intermediate data sequence for update. Is generated. In this way, it is possible to adjust the reproducibility when reproducing the variance of the time series data of the characteristic parameters by the weight. And the correction data generation part 120 can perform the process of Numerical formula (13).

  The process control unit 150 repeatedly corrects the time-series data of the corrected feature parameters until a predetermined condition is satisfied. That is, the process control unit 150 controls each unit to repeat the process in the second intermediate data generation unit 140, the process in the correction data generation unit 120, and the process in the data correction unit 130 until a predetermined criterion is satisfied. Examples of the predetermined criteria include (1) a feature parameter time-series correction data that is output from the correction data generation unit 120, and a mean square sum of the elements is equal to or less than a certain threshold value. There is to repeat. In this way, the time series data of feature parameters can be converged to a certain reproduction level.

(Operation of feature parameter generator)
Next, the operation of the feature parameter generation device 100 configured as described above will be described. First, the feature parameter generation device 100 uses the decision tree or the like to determine the time series information of the feature parameter distribution and the distribution information of the variance of the time series data of the feature parameter from the symbol series information such as the input symbol string in the previous stage of the above configuration. Generate (step S1).

  Then, time series data with all elements set to 0 is set as an initial series of time series data of feature parameters (step S2). Then, the first intermediate data generation unit 110 generates an initial intermediate data series from the above information so that the distribution of the characteristic parameters becomes maximum likelihood (step S3).

  Next, the correction data generation unit 120 generates correction series data from the initial intermediate data series (step S4). Then, the data correction unit 130 corrects the time series data of the characteristic parameters corresponding to the symbol information series using the correction series data (step S5).

  Next, it is determined whether or not a repetitive process stop condition is satisfied (step S6). That is, it is determined whether or not the predetermined criterion is satisfied. When the stop condition is not satisfied, the second intermediate data generation unit 140 corrects the feature parameter time series so that the likelihood of the distribution of the feature parameters and the time series data of the feature parameters becomes larger. An intermediate data series for update is generated from the data (step S7), and the process returns to step S4. On the other hand, when the stop condition is satisfied, the time series data of the generated feature parameter is output (step S8), and the operation is terminated.

  In this way, in the feature parameter generation device 100, the correction data generation unit 120 generates correction sequence data from the update intermediate data sequence, and the data correction unit 130 uses the correction sequence data to correct the feature parameters. A series of operations for correcting time series data is repeated. As a result, the correction data generation unit 120 operates both in the initial correction and the repeated correction of the time series data of the characteristic parameters, so that it is not necessary to provide redundant circuits and software. Scale and software scale can be reduced. The time series data of the characteristic parameters generated in this way is output as, for example, voice.

  Note that the time-series distribution of feature parameters may include a time-series dynamic feature distribution of feature parameters. The first intermediate data generation unit 110 generates an intermediate data sequence from the symbol information sequence so that the distribution of feature parameters including dynamic features is maximum likelihood, and the second intermediate data generation unit 140 generates dynamic features. An intermediate data sequence is generated from the symbol information sequence so that the distribution of the feature parameter and the likelihood of dispersion of the time series data of the feature parameter are increased. Thereby, the reproducibility of the time series data of the feature parameter can be improved in consideration of the dynamic feature.

  The dynamic features are not limited to those defined by Equation (2) and Equation (3). It suffices if the inverse matrix of Expression (12) can be calculated, and the number of dimensions of oi is arbitrary. As such an example, there is a method in which the result of discrete cosine transform for an appropriate subsequence of {xi} is used as an element of oi.

  Also, the variance of the time series data of the feature parameters may not be the variance obtained from the entire sequence. For example, for each time, there is a method of calculating the variance of only a certain number of frames before and after the time and using the average value thereof. In this case, the time series data of the characteristic parameters can be reproduced more accurately by determining the distribution from the calculated values in the same manner.

  In the above embodiment, one-dimensional value time-series data is generated. However, the same applies to the case of expansion to the generation of multi-dimensional value (vector) data. Moreover, in the above embodiment, the process of each part is realizable by making a computer run a program.

100 feature parameter generation device 110 intermediate data generation unit 120 correction data generation unit 130 data correction unit 140 intermediate data generation unit 150 processing control unit

Claims (8)

A feature parameter generation device for reproducing time series data of feature parameters corresponding to symbol series information,
A first intermediate data generation unit that generates an initial intermediate data sequence composed of a first symmetric band matrix and a first vector from a symbol information sequence so that the distribution of feature parameters becomes maximum likelihood;
Correction data for generating first correction series data by an operation equivalent to calculating a product of an inverse matrix for the first symmetric band matrix and the first vector from the initial intermediate data series A generator,
A data correction unit for correcting time series data of feature parameters corresponding to the symbol information series using the first correction series data;
For updating the characteristic parameter distribution and the characteristic parameter time-series data variance, the second characteristic symmetric band matrix and the second vector are updated from the corrected characteristic parameter time-series data . A second intermediate data generation unit for generating an intermediate data series,
A second correction sequence is generated by an operation corresponding to the correction data generation unit calculating a product of the inverse matrix for the second symmetric band matrix and the second vector from the update intermediate data sequence. A feature parameter generation device characterized in that data is generated and the data correction unit repeats a series of operations for correcting the time series data of the corrected feature parameter using the second correction series data. The feature parameter generation apparatus according to claim 1 , further comprising a processing control unit that repeatedly corrects the time series data of the corrected feature parameters until a predetermined condition is satisfied. The feature parameter time-series distribution includes a feature parameter time-series dynamic feature distribution, and the first intermediate data generation unit performs symbol information so that the feature parameter distribution including the dynamic feature is a maximum likelihood. An intermediate data sequence is generated from the sequence, and the second intermediate data generation unit generates the distribution of the feature parameter including the dynamic feature and the likelihood of the variance of the time series data of the feature parameter from the symbol information sequence. 3. The characteristic parameter generation apparatus according to claim 1, wherein the intermediate data series is generated. The second intermediate data generation unit generates an intermediate data sequence for update using a weighted sum of the log likelihood of the distribution of the feature parameter and the log likelihood of the variance of the time series data of the feature parameter. The feature parameter generation device according to claim 1 , wherein the feature parameter generation device is a feature parameter generation device. The second intermediate data generation unit generates the intermediate data series for update under a condition that the likelihood of variance of the time series data of the feature parameters does not change before and after the correction. The feature parameter generation device according to claim 4 , wherein the feature parameter generation device is a feature. Voice synthesis information describing the type of unit voice included in a series of unit voice strings is input as the symbol information series, and a synthesized voice waveform is generated as time-series data of the modified feature parameters. The feature parameter generation device according to any one of claims 1 to 5 . A feature parameter generation method for reproducing time series data of a feature parameter corresponding to a symbol information series performed by a computer,
Generating an initial intermediate data sequence consisting of a first symmetric band matrix and a first vector from a symbol information sequence so that the distribution of feature parameters is maximum likelihood;
Generating first correction series data from the initial intermediate data series by an operation equivalent to calculating a product of an inverse matrix for the first symmetric band matrix and the first vector ; ,
Correcting time series data of characteristic parameters corresponding to the symbol information series using the first correction series data;
For updating the characteristic parameter distribution and the characteristic parameter time-series data variance, the second characteristic symmetric band matrix and the second vector are updated from the corrected characteristic parameter time-series data . Generating an intermediate data series,
Wherein the intermediate data sequence for updating, and generates the inverse matrix for the second symmetric band matrix, the second correcting series data by operation corresponding to calculating the product of the second vector, the A feature parameter generation method characterized by repeating a series of steps for correcting the time series data of the corrected feature parameters using the second correction series data . A feature parameter generation program for causing a computer to reproduce time series data of feature parameters corresponding to symbol information series,
A process of generating an initial intermediate data sequence composed of a first symmetric band matrix and a first vector from a symbol information sequence so that the distribution of feature parameters becomes maximum likelihood;
Processing for generating first correction series data by an operation corresponding to calculating a product of an inverse matrix for the first symmetric band matrix and the first vector from the initial intermediate data series ; ,
Processing to correct time series data of feature parameters corresponding to the symbol information series using the first correction series data;
For updating the characteristic parameter distribution and the characteristic parameter time-series data variance, the second characteristic symmetric band matrix and the second vector are updated from the corrected characteristic parameter time-series data . Processing to generate an intermediate data series,
Wherein the intermediate data sequence for updating, and generates the inverse matrix for the second symmetric band matrix, the second correcting series data by operation corresponding to calculating the product of the second vector, the A feature parameter generation program that repeats a series of processes for correcting the time series data of the corrected feature parameter using the second correction series data .

Images