JP5647159B2 - Prior distribution calculation device, speech recognition device, prior distribution calculation method, speech recognition method, program - Google Patents

Description

  The present invention relates to a prior distribution calculation device, a prior distribution calculation method, and a program for generating a prior distribution commonly used in a feature space and an acoustic model space, and a speech recognition device, a speech recognition method, and a program using the prior distribution.

  Adaptation techniques have been developed to prevent the adverse effects of various variables (eg, speakers, noise, communication channels, microphones, etc.) that affect the speech recognition input signal. In particular, model-based adaptation techniques are often used as adaptation techniques because all parameters of an acoustic model can be adapted by linear transformation using a transformation matrix.

  As the adaptation technique of the linear transformation format based on the model, Unconstrained_Maximum_Likelihood_Linear_Regression (UMLRR, unconstrained maximum likelihood linear regression; hereinafter referred to as MLLR) (Non-Patent Document 1) and Constrained_Maximum_LikeliLedReliable_Least )It has been known. The former is an adaptation method in a model space, and the latter is a feature space. CMLLR is also called feature_space_MLLR (fMLLR, feature space maximum likelihood linear regression) because it can express a conversion formula in the feature space. In particular, this method is effective for Speaker_Adaptive_Training (SAT, speaker adaptive learning) (Non-patent Document 3), and has an advantage that the amount of memory used and the amount of calculation can be reduced.

  However, the transformation matrix estimation method that does not use the prior distribution such as MLLR (maximum likelihood linear regression) described above cannot perform reliable estimation when the amount of adaptive data is small, and thus cannot reduce the recognition rate and cannot recognize itself. Occurs. Therefore, a method using prior distribution has been proposed to solve this problem.

  Typical methods using the prior distribution include Maximum_A_Posterori_Linear_Regulation (MAPLR) (Non-Patent Document 4), Structural_MAPLR (SMAPLR) (Non-Patent Document 5), and feature_space_MAPLR (fMAPLR) (Non-Patent Document 6). MAPLR and SMAPLR are adaptive methods in the acoustic model space, and fMAPLR is an adaptive method in the feature space. For the prior distribution, the distribution of the conversion matrix of the speakers included in the learning data using each method is used.

  A conventional SAT using MLLR requires a large amount of memory and calculation. This is because the transformation matrix of the node selected by the adaptive data is estimated based on the tree structure often used in MLLR. In general, the amount of learning data is much larger than the amount of test data, and since there are many nodes selected from the tree structure, the number of transformation matrices for each speaker to be estimated increases. Furthermore, since MLLR has different conversion matrices for average and variance, it requires twice the amount of calculation and memory compared to CMLLR.

Leggetter, C. and Woodland, P.C.Maximum likelihood linear regression for speaker adaptation of continuous density hidden Markov models.Computer Speech and Language, 9 (2): 171--185, 1995. Gales, M.J.F.Maximum likelihood linear transformations for HMM-based speech recognition.Computer Speech and Language, 12: 75--98, 1998. Anastasakos, T. and McDonough, J. and Makhoul, J. Speaker adaptive training: A maximum likelihood approach to speaker normalization.Proc. Of ICASSP, pages 1043--1046, 1997. Siohan, O. and Chesta, C. and Lee, C.H.Joint maximum a posteriori adaptation of transformation and HMM parameters.IEEE Trans. On Speech and Audio Processing, 9 (4): 417--428, 2001. Siohan, O. and Myrvoll, T.A. and Lee, C.H.Structural maximum a posteriori linear regression for fast HMM adaptation.Computer Speech & Language, 16 (1): 5--24, 2002. Lei, X. and Hamaker, J. and He, X. Robust feature space adaptation for telephony speech recognition.Proc. Of INTERSPEECH, pages 773--776, 2006.

  When speaker adaptation is performed by combining both the adaptation method in the model space and the adaptation method in the feature space, the recognition performance improves when the amount of adaptation data is large compared to the adaptation method in the model or feature space alone. Can be achieved. However, if the prior distribution is not used even by the above combination method, it is impossible to perform highly reliable estimation when the amount of adaptive data is small. On the other hand, if the prior distribution is used for both the adaptation in the model space and the adaptation in the feature space, the prior distribution is calculated separately for each, and the calculation amount increases.

  Therefore, the present invention generates a prior distribution that can be commonly used for both adaptation in the model space and adaptation in the feature space, and reduces the amount of calculation of the prior distribution, speech recognition An object is to provide a device, a prior distribution calculation method, a speech recognition method, and a program.

  The prior distribution calculation device of the present invention includes a feature vector extraction unit, a first transformation matrix estimation unit, a feature vector conversion unit, an MLE acoustic model learning unit, a second transformation matrix estimation unit, and a prior distribution calculation unit. Prepare.

  The feature vector extraction unit extracts feature vectors for each speaker from the input speech of a plurality of speakers. The first transformation matrix estimation unit estimates a first transformation matrix for each speaker by feature space maximum likelihood linear regression using a feature vector and an initial acoustic model previously learned from data of all speakers. The feature vector conversion unit converts the feature vector of the corresponding speaker using the first conversion matrix for each speaker. The MLE acoustic model learning unit learns the acoustic model by the maximum likelihood method using the feature vector converted by the feature vector conversion unit. The second transformation matrix estimation unit performs second transformation for each speaker by feature space maximum likelihood linear regression using the feature vector extracted by the feature vector extraction unit and the acoustic model learned by the MLE acoustic model learning unit. Estimate the matrix. The prior distribution calculation unit calculates a multivariate normal distribution of the matrix using the second transformation matrix, and outputs the hyperparameter of the prior distribution using the multivariate normal distribution as the prior distribution.

  According to the prior distribution calculation device of the present invention, it is possible to generate a prior distribution that can be commonly used for both adaptation in the model space and adaptation in the feature space, and reduce the amount of calculation of the prior distribution.

1 is a block diagram illustrating a configuration of a prior distribution calculation apparatus according to Embodiment 1. FIG. 3 is a flowchart illustrating the operation of the prior distribution calculation apparatus according to the first embodiment. FIG. 3 is a block diagram illustrating a configuration of a speech recognition apparatus according to a second embodiment. 9 is a flowchart illustrating the operation of the speech recognition apparatus according to the second embodiment. The block diagram which shows the structure of the speech recognition apparatus of the modification 1. FIG. 9 is a flowchart showing the operation of the speech recognition apparatus according to the first modification.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the structure part which has the same function, and duplication description is abbreviate | omitted.

  Hereinafter, the prior distribution calculation apparatus according to the first embodiment will be described in detail with reference to FIGS. 1 and 2. FIG. 1 is a block diagram showing the configuration of the prior distribution calculation apparatus 1 of the present embodiment. FIG. 2 is a flowchart showing the operation of the prior distribution calculation apparatus 1 of the present embodiment. The prior distribution calculation apparatus 1 of the present embodiment includes a feature vector extraction unit 10, a first transformation matrix estimation unit 20, a feature vector conversion unit 30, an MLE acoustic model learning unit 40, and a normalized acoustic model storage unit 50. A second transformation matrix estimation unit 60, a prior distribution calculation unit 70, and an initial acoustic model storage unit 80. The first transformation matrix estimation unit 20 includes a statistic G calculation unit 21, a statistic k calculation unit 22, a transformation matrix estimation unit 23, and an iterative learning unit 24. The MLE acoustic model learning unit 40 includes an average update unit 41 and a distributed update unit 42. The prior distribution calculation unit 70 includes parameter C calculation means 71 and parameter V calculation means 72. In the initial acoustic model storage unit 80, an acoustic model learned using data of all speakers is stored in advance as an initial acoustic model.

  Hereinafter, the outline of the process will be described in three sections first, and specific processes of each component in each section will be described at the end of each section.

<1. Estimation of transformation matrix (processing of first transformation matrix estimation unit 20)>
The prior distribution calculation apparatus 1 according to the present embodiment uses an initial acoustic model learned in advance using data of all speakers and a feature vector obtained by converting each speaker's input speech for each speaker. The conversion matrix (first conversion matrix) of each speaker is estimated by maximum likelihood linear regression. First, a feature vector extracted from input speech is defined as o (t). At this time, o (t) represents the N-dimensional feature vector of the t-th frame. The prior distribution calculation apparatus 1 of the present embodiment converts this feature vector o (t) into a feature vector o (t) hat using a conversion matrix. The transformed feature vector o (t) hat is as follows.

The Q function for transform matrix estimation is defined as follows.

Here, the optimization problem of W (estimation of α) is described in detail in Non-Patent Document 2. The i-th row of the transformation matrix W is obtained as follows.

The i-th statistics G ⁽ⁱ⁾ and k ⁽ⁱ⁾ for transform matrix estimation are the extended feature vector ξ (t) of the input speech and the i-th average μ _i ^{(u )} And variance σ _i ^(u) .

After the estimation of the Row-by-row transformation matrix, iterative learning is performed by the maximum likelihood method using the following equation.

  In the present embodiment, the first transformation matrix estimation unit 20 executes the processing of the first section described above. Therefore, first, the feature vector extraction unit 10 extracts an N-dimensional feature vector o (t) from the input speech of S speakers (speakers 1,..., Speakers S, S is an integer of 2 or more) (S10). ). As described above, in the initial acoustic model storage unit 80, an acoustic model learned using data of all speakers is stored in advance as an initial acoustic model. The statistic G calculating means 21 calculates the statistic G by the equation (4) (SS21). The statistic k calculation means 22 calculates the statistic k using equation (5) (SS22). Next, the transformation matrix estimation means 23 estimates the transformation matrix using equation (3) (SS23). The processing of these sub-steps SS21 to SS23 is repeatedly executed for all possible values of the dimension (row) i, and the transformation matrix W is obtained. Next, the iterative learning means 24 performs iterative learning of the transformation matrix by the maximum likelihood method using Equation (6) (SS24). Through the above steps S10 and S20, a conversion matrix of speakers 1 to S is obtained.

<2. Acoustic Model Learning (Processing of Feature Vector Conversion Unit 30 and MLE Acoustic Model Learning Unit 40)>
The prior distribution calculation apparatus 1 of the present embodiment uses the conversion matrix of speakers 1 to S to convert each speaker's feature vector, and uses the converted feature vector (learning data) for maximum likelihood. The acoustic model is learned by the method. The Q function for acoustic model learning (SAT) is defined as follows.

The transformed feature vector o ^(s) (t) is calculated by the following equation.

The processing of Expression (7) is processing in which only the feature vector is different compared to the conventional maximum likelihood method (MLE, Maximum_Likelihood_Estimation). That is, learning similar to conventional MLE is performed using o ^(s) (t) hat instead of the original feature vector o ^(s) (t).

The update formula for mean and variance is:

  In the present embodiment, the feature vector conversion unit 30 executes the processing according to the equation (8) among the processing of the second section described above, and the MLE acoustic model learning unit 40 executes the remaining processing. Therefore, the feature vector conversion unit 30 converts the feature vector generated by the feature vector extraction unit 10 using the equation (8) based on the conversion matrix estimated by the first conversion matrix estimation unit 20 (S30). Next, the learning of the acoustic model by MLE (maximum likelihood) is repeated until the likelihood converges while updating the average and variance for each learning count. The average updating unit 41 obtains an average from the feature vector after conversion using Equation (9) (SS41). The variance updating unit 42 obtains variance from the converted feature vector using Equation (10) (SS42). The acoustic model in which the speaker variation is normalized, which is learned using the characteristics transformed using the transformation matrix of each speaker, is stored in the normalized acoustic model storage unit 50 for the next processing ( S50).

<3. Calculation of Prior Distribution (Processing of Second Transformation Matrix Estimation Unit 60 and Prior Distribution Calculation Unit 70)>
The prior distribution calculation apparatus 1 according to the present embodiment uses the acoustic model learned by the MLE acoustic model learning unit 40 to obtain a transformation matrix (second transformation matrix) for each speaker using fMLLR, as in the first section. Ask. The prior distribution calculation apparatus 1 according to the present embodiment obtains a prior distribution (average and variance for each dimension) of the obtained transformation matrix (second transformation matrix) of each speaker.

First, for the prior distribution calculation, the same processing (Equations (3) to (6)) as in the first transformation matrix estimation unit 20 is performed again by fMLLR using the acoustic model learned by the MLE acoustic model learning unit 40. Thus, the conversion matrix (second conversion matrix) of each speaker is obtained. The calculation of the prior distribution uses a multivariate normal distribution of a matrix described in Non-Patent Document 4. This multivariate normal distribution is defined as:

Here, the distributed hyperparameter is assumed to be a unit matrix as in Non-Patent Document 4. Let S be the total number of speakers, and hyperparameters C and V are obtained from the S transformation matrices using the following equations.

  In the present embodiment, among the processes of the third section described above, the second conversion matrix estimation unit 60 executes the conversion matrix estimation process, and the prior distribution calculation unit 70 executes the process according to Expression (12). Therefore, the second transformation matrix estimation unit 60 estimates the transformation matrix (second transformation matrix) based on the equations (3) to (6) using the acoustic model learned by the MLE acoustic model learning unit 40. . This process is the same as the process of the first transformation matrix estimation unit 20. Next, under the assumption that the obtained second transformation matrix follows the multivariate normalized distribution of the matrix, the parameter C calculation means 71 calculates the hyperparameter C using the second transformation matrix according to the equation (12). (SS71). The parameter V calculation means 72 calculates the hyperparameter V using the second transformation matrix according to the equation (12) (SS72).

  Thus, according to the prior distribution calculation apparatus 1 of the present embodiment, a prior distribution that can be commonly used for both adaptation in the model space and adaptation in the feature space is generated, and the amount of calculation of the prior distribution is reduced. can do.

  Hereinafter, the speech recognition apparatus according to the second embodiment will be described in detail with reference to FIGS. 3 and 4. FIG. 3 is a block diagram showing the configuration of the speech recognition apparatus 100 of this embodiment. FIG. 4 is a flowchart showing the operation of the speech recognition apparatus 100 of this embodiment. The speech recognition apparatus 100 according to the present embodiment is characterized in that the feature space and the model space are simultaneously adapted by commonly using the prior distribution obtained in advance by the prior distribution calculation apparatus 1 according to the first embodiment. The speech recognition apparatus 100 according to the present embodiment includes a feature vector extraction unit 110, a feature vector storage unit 115, a feature vector conversion unit 120, an initial transformation matrix storage unit 125, a speech recognition unit 130, and a recognition data storage unit. 140, a recognition result storage unit 145, a feature space statistics calculation unit 150, a feature space transformation matrix estimation unit 155, a tree structure determination unit 160, a model space statistics calculation unit 170, and a model space transformation matrix estimation unit 175, an acoustic model update unit 180, and a prior distribution storage unit 190. The recognition data storage unit 140 includes an acoustic model 141, a language model 142, and a word dictionary 143. The model space statistic calculator 170 includes a statistic G tilde calculator 171, a statistic k tilde calculator 172, and a smoothed statistic calculator 173. The prior distribution storage unit 190 stores in advance the hyperparameters C and V of the prior distribution generated by the method described in the first embodiment.

  Hereinafter, an outline of processing of the speech recognition apparatus 100 of the present embodiment will be described, and then specific processing contents of each component will be described.

<4. Simultaneous adaptation of feature space and acoustic model space by sharing prior distribution>
The speech recognition apparatus 100 according to the present embodiment recognizes input speech and performs statistical calculation based on the speech recognition result (unsupervised adaptation). In calculating statistics, the prior distribution obtained in advance by the method of the first embodiment is reflected. The speech recognition apparatus 100 of the present embodiment estimates the transformation matrix in the feature space and the model space from the calculated statistics. The speech recognition apparatus 100 according to the present embodiment performs re-recognition by updating the N-dimensional feature vector and the acoustic model, respectively, using the estimated feature space transformation matrix and model space transformation matrix.

The ML criterion Q function without prior distribution is defined as follows.

  Transformation in model space only considers the mean. In other words, the distribution is adapted in the feature space. Since it is difficult to directly optimize the expression (13) in different spaces, here, a method of performing optimization in order in the feature space and the model space is used for simultaneous optimization.

First, assuming that the model space transformation matrix W _r ^M is a unit transformation matrix [0 _n ^T I _{n × n} ] and there is no prior distribution in the model space, the Q function in the feature space using the prior distribution is It becomes like this.

The transformation matrix of the i-th row in the feature space can be estimated by the following equation.

統計量Ｇ^（ｉ）ハットと統計量ｋ^（ｉ）ハットは、それぞれ、Ｇ^（ｉ）、ｋ^（ｉ）の平滑化された統計量を意味する。Ｇ^（ｉ）、ｋ^（ｉ）は式（４）と式（５）を用いて計算したものである。 Here, the statistic is calculated using the prior distribution as follows.

Statistics ^{G (i)} hat and statistic ^{k (i)} hat, ^{respectively,} it means ^{G (i), k} smoothed statistic ^(i). G ⁽ⁱ⁾ and k ⁽ⁱ⁾ are calculated using the equations (4) and (5).

Next, the Q function in the model space using the prior distribution is as follows.

The hyperparameter V tilde and C tilde of the prior distribution in the model space are defined as follows.

The transformation matrix W _r ^{M in} the i-th row of the r-th recursive class is defined by the following equation.

The smoothed statistics G ⁽ⁱ⁾ bar and k ⁽ⁱ⁾ bar are calculated using the following equations.

Statistics G tilde and k tilde in the model space are calculated using the following equations.

Using the obtained transformation matrix, the average of the acoustic model is updated as in the following equation.

  At the time of recognition (test), the feature vector of the input speech is transformed using the transformation matrix obtained in the feature space, and recognition is performed based on the acoustic model updated by the model space transformation matrix.

  In the present embodiment, each component of the speech recognition apparatus 100 executes the above-described processing in section 4. First, the feature vector extraction unit 110 extracts an N-dimensional feature vector from the input audio signal (S110). Next, the feature vector storage unit 115 stores an N-dimensional feature vector (S115). The stored N-dimensional feature vector is used for feature vector conversion in both steps S120-1 and S120-2 described later. The feature vector conversion unit 120 converts the feature vector using a conversion matrix (S120-1). Here, the initial value of the transformation matrix is assumed to be stored in the initial transformation matrix storage unit 125, and the initial transformation matrix is used when the feature vector transformation unit 120 is operated for the first time (in step S120-1). To do. Since the initial transformation matrix is a unit transformation matrix (the bias is all 0 and the rotation matrix is a unit matrix), the feature vector o (t) before the transformation and the feature vector o (t) after the transformation are the same. . Next, the speech recognition unit 130 performs speech recognition using the acoustic model 141, the language model 142, and the word dictionary 143 stored in the recognition data storage unit 140, and generates a speech recognition result from the converted feature vector. (S130-1). The recognition result storage unit 145 stores the generated speech recognition result (S145-1). Next, the feature space statistic calculation unit 150 calculates the statistic G hat and k hat using the equation (16) (S150). The feature space transformation matrix estimation unit 155 estimates the transformation matrix of the feature space using Equation (15) (S155). Next, the feature vector conversion unit 120 converts the feature vector stored in the feature vector storage unit 115 using the feature space conversion matrix estimated by the feature space conversion matrix estimation unit 155 (S120-2). Similar to step S130-1, the speech recognition unit 130 generates a speech recognition result from the feature vector converted using the feature space conversion matrix in step S120-2 (S130-2). The recognition result storage unit 145 stores the generated speech recognition result (S145-2). Next, the tree structure determination unit 160 classifies the speech recognition result into a tree structure using Expression (18), and determines hyperparameter C tilde and V tilde (S160). Next, the statistic G tilde calculation means 171 calculates the statistic G tilde according to the equation (21) (SS171). Next, the statistic k tilde calculation means calculates the statistic k tilde using equation (22) (SS172). The smoothed statistic calculation means 173 calculates the statistic G tilde, the statistic k tilde, the hyperparameter C tilde, and the V tilde using the equation (20) to calculate the smoothed statistic G bar and k bar. (SS173). The model space transformation matrix estimation unit 175 estimates the transformation matrix of the model space according to the equation (19) using the statistics G bar and k bar (S175). The acoustic model update unit 180 calculates the average of the acoustic model according to the equation (23) using the estimated transformation matrix of the model space, and updates the acoustic model 141 (S180).

  As described above, according to the speech recognition apparatus 100 of the present embodiment, in order to simultaneously adapt the feature space and the model space using a predetermined common prior distribution, the calculation amount of the prior distribution is reduced. In addition to the common effects, the recognition rate is improved by using prior distribution when the amount of adaptive data is small, and speaker adaptation is performed by combining the feature space and model space when the amount of adaptive data is large. As a result, the recognition rate is improved, so that the recognition rate is improved regardless of the amount of adaptive data.

[Modification 1]
Hereinafter, with reference to FIG. 5 and FIG. 6, a speech recognition apparatus according to Modification 1 which is a modification of the speech recognition apparatus 100 according to Embodiment 2 will be described. FIG. 5 is a block diagram showing a configuration of a speech recognition apparatus 100 ′ according to this modification. FIG. 6 is a flowchart showing the operation of the speech recognition apparatus 100 ′ of this modification. Similar to the second embodiment, the speech recognition apparatus 100 ′ of the present modification is characterized in that the feature space and the model space are simultaneously adapted using the prior distribution previously obtained by the method of the first embodiment in common. The speech recognition apparatus 100 ′ of the present modification includes a feature vector extraction unit 110, a feature vector storage unit 115, a feature vector conversion unit 120, an initial transformation matrix storage unit 125, a speech recognition unit 130, and a recognition data storage. Unit 140, recognition result storage unit 145, feature space statistic calculation unit 150, feature space conversion matrix estimation unit 155, tree structure determination unit 160 ′, model space statistic calculation unit 170, model space conversion matrix An estimation unit 175, an acoustic model update unit 180, and a prior distribution storage unit 190 are provided. Since each component other than the tree structure determining unit 160 ′ performs the same operation as each component having the same number in the speech recognition apparatus 100 of the second embodiment, the description thereof is omitted.

Accordingly, steps S110 to S145-1 are executed in the same manner as in the second embodiment. Next, the tree structure determination unit 160 ′ classifies the speech recognition result obtained first using the equation (18) ′ into a tree structure, and determines hyperparameter C tilde and V tilde (S160′-1). . Equation (18) ′ is shown below.

  Thereafter, step S150 to step S145-2 are executed in the same manner as in the second embodiment. Next, the tree structure determination unit 160 'classifies the speech recognition result obtained a second time using the equation (18) into a tree structure, and determines hyperparameter C tilde and V tilde (S160'-2). Thereafter, step S170, step S175, and step S180 are executed as in the second embodiment.

  As described above, according to the speech recognition apparatus 100 ′ of the present modified example, the speaker adaptation is performed using the tree structure even in the feature space, so that when the amount of adaptive data is large, the recognition rate is further improved compared to the second embodiment. To do.

  In addition, the various processes described above are not only executed in time series according to the description, but may be executed in parallel or individually according to the processing capability of the apparatus that executes the processes or as necessary. Needless to say, other modifications are possible without departing from the spirit of the present invention.

  Further, when the above-described configuration is realized by a computer, processing contents of functions that each device should have are described by a program. The processing functions are realized on the computer by executing the program on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the present apparatus is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

Claims (5)

A feature vector extraction unit that extracts feature vectors for each speaker from input speech of a plurality of speakers;
A first transformation matrix estimation unit for estimating a first transformation matrix for each speaker by feature space maximum likelihood linear regression using the feature vector and an initial acoustic model previously learned from data of all speakers;
A feature vector conversion unit that converts a feature vector of a corresponding speaker using the first conversion matrix for each speaker;
An MLE acoustic model learning unit that performs acoustic model learning by a maximum likelihood method using the feature vector converted by the feature vector conversion unit;
A second transformation matrix is estimated for each speaker by feature space maximum likelihood linear regression using the feature vector transformed by the feature vector extraction unit and the acoustic model learned by the MLE acoustic model learning unit. A transformation matrix estimation unit;
Calculating a multivariate normal distribution of a matrix using the second transformation matrix, setting the multivariate normal distribution as a prior distribution, and outputting a hyperparameter of the prior distribution;
A prior distribution calculation device comprising: A feature vector extraction unit that extracts a feature vector from input speech;
A feature vector conversion unit that converts the feature vector using an initial conversion matrix comprising a unit conversion matrix or a feature space conversion matrix;
A data storage unit for recognition that stores an acoustic model;
A speech recognition unit that performs speech recognition using the acoustic model and the feature vector converted by the feature vector conversion unit;
A feature space statistic calculation unit that calculates a statistic used for estimating a transformation matrix of the feature space using a hyperparameter of a prior distribution;
A feature space transformation matrix estimation unit that estimates a feature space transformation matrix using the statistics calculated by the feature space statistics calculation unit;
A model space statistic calculator for calculating a statistic used for estimating a transformation matrix of a model space using the hyperparameter of the prior distribution;
A model space transformation matrix estimation unit for estimating a model space transformation matrix using the statistics calculated by the model space statistics calculation unit;
A speech recognition apparatus comprising: an acoustic model update unit that updates the acoustic model using the estimated model space transformation matrix;
Prior distribution used in common by the feature space statistic calculator and the model space statistic calculator is
Each speaker's input speech is converted into a feature vector for each speaker, and each speaker is subjected to feature space maximum likelihood linear regression using the feature vector and an initial acoustic model previously learned from the data of all speakers. The first transformation matrix is estimated, the feature vector of the corresponding speaker is transformed using the first transformation matrix for each speaker, and the feature vector transformed using the first transformation matrix is And the acoustic model is learned by the maximum likelihood method, and the feature vector before being transformed by the first transformation matrix and the learned acoustic model are used for each speaker by the feature space maximum likelihood linear regression. A speech recognition apparatus, wherein the second transformation matrix is estimated to be a multivariate normal distribution of a matrix calculated using the second transformation matrix. A feature vector extraction step of extracting feature vectors for each speaker from the input speech of a plurality of speakers;
A first transformation matrix estimation step for estimating a first transformation matrix for each speaker by feature space maximum likelihood linear regression using the feature vector and an initial acoustic model previously learned from data of all speakers;
A feature vector conversion step of converting a feature vector of a corresponding speaker using the first conversion matrix for each speaker;
An MLE acoustic model learning step of learning an acoustic model by a maximum likelihood method using the feature vector converted by the feature vector conversion step;
A second transformation matrix is estimated for each speaker by feature space maximum likelihood linear regression using the feature vector transformed in the feature vector extraction step and the acoustic model learned in the MLE acoustic model learning step. A transformation matrix estimation step;
Calculating a multivariate normal distribution of a matrix using the second transformation matrix, setting the multivariate normal distribution as a prior distribution, and outputting a hyperparameter of the prior distribution;
A prior distribution calculation method characterized by comprising: A feature vector extraction step for extracting a feature vector from the input speech;
A feature vector conversion step of converting the feature vector using an initial conversion matrix comprising a unit conversion matrix or a feature space conversion matrix;
A speech recognition step for performing speech recognition using an acoustic model and the feature vector converted by the feature vector conversion step;
A feature space statistic calculation step for calculating a statistic used for estimating a transformation matrix of the feature space using a hyperparameter of a prior distribution;
A feature space transformation matrix estimation step for estimating a feature space transformation matrix using the statistics calculated by the feature space statistics calculation step;
A model space statistic calculation step for calculating a statistic used for estimating a transformation matrix of the model space using the hyperparameter of the prior distribution;
A model space transformation matrix estimation step for estimating a model space transformation matrix using the statistics calculated by the model space statistics calculation step;
An acoustic model update step of updating the acoustic model using a transformation matrix of the estimated model space,
Prior distribution commonly used in the feature space statistic calculation step and the model space statistic calculation step is:
A feature vector for each speaker is extracted from input speech of a plurality of speakers, and each speaker is subjected to feature space maximum likelihood linear regression using the feature vector and an initial acoustic model previously learned from data of all speakers. The first transformation matrix is estimated, the feature vector of the corresponding speaker is transformed using the first transformation matrix for each speaker, and the feature vector transformed using the first transformation matrix is And the acoustic model is learned by the maximum likelihood method, and the feature vector before being transformed by the first transformation matrix and the learned acoustic model are used for each speaker by the feature space maximum likelihood linear regression. A speech recognition method, wherein the second transformation matrix is estimated to be a multivariate normal distribution of a matrix calculated using the second transformation matrix.   A program for causing a computer to function as the apparatus according to claim 1.

Images