JP2007086703A - Speech recognition device and speech recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognition device that can output possibilities of recognition units such as words representing speech contents in the form of probabilities and surely improve speech recognition performance as compared with the conventional method using a sound model such as a hidden Markov model; and also to provide a speech recognition method. <P>SOLUTION: As for a received speech signal, a likelihood of a conventional sound model is taken in the regression function of a penalized logistic regression machine (PLRM); the coefficient parameter W and the hyperparameter Λ of the PLRM are learnt and optimized; and those parameters are used to compute predicted probability of a recognition unit representing utterance contents of a speech signal. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a voice recognition device and a voice recognition method. Specifically, after letting the learning machine learn using the training voice data. The present invention relates to a speech recognition apparatus and a speech recognition method for calculating predictions based on probabilities of possible words and the like for newly given speech data and inferring words and the like from speech.

  In the conventional speech recognition, the utterance content is determined using the maximum posterior likelihood determination rule. The maximum a posteriori likelihood is calculated as a product of the likelihood of an acoustic model represented by a hidden Markov model HMM (Hidden Markov Model) and the probability of a language model represented by N-gram or the like. In these models, the HMM parameters are learned using an error minimization learning algorithm or an EM (Expectation Maximization) algorithm. (See Non-Patent Documents 1 and 2)

  On the other hand, a fine logistic regression machine (Penarized Logistic Regression Machine; hereinafter referred to as PLRM in this specification), which is a learning machine with excellent probabilistic predictive discriminating power, was proposed by Tanabe, and its performance was shown in the field of speech. Yes. (See Non-Patent Documents 3-8)

Seiichi Nakagawa, “Voice Recognition by Probabilistic Model”, The Institute of Electronics, Information and Communication Engineers, 1988. B, -H, Jung, W .; Chou, and C.C. -H. Lee, “Minimum classification error rate methods for speech re-regulation”, IEEE Trans, Speech and Audio Processing, vol. 5, pp. 257-265, May 1997. K. Tanabe, “Penalized logistic regression machines: New methods for statistical prediction 1”, ISM Cooperative Research Report 143, pp. 197 163-194, March 2001. K. Tanabé, “Penalized logic regression machines: New methods for statistical prediction 2”, 4th Information Theory Learning Theory Workshop (IBIS2001), pp. 199 71-76, July. 2001. K. Tanab, “Penalized logistic regression machines and Related Linear Numerical Algebra”, Kyoto University Institute of Mathematical Analysis, Proc. 1320, pp. 239-249, 2003. T. T. et al. Matsui and K.M. Tanab, “Speaker Identification with Dual Penalized Logistic Regression Machine”, Processeds of Odyssey, pp. 363-366, 2004. T. T. et al. Matsui and K.M. Tanabbe, “Probabilistic Spiker Identification with Dual Penalized Logic Regression Machine”, Proceedings of ICSLP, pp. III-1797-1800, 2004. T. T. et al. Matsui and K.M. Tanab, “Speaker Recognition Without Feature Extraction Process”, Processeds of Workshops on Strategic Modeling for Prospect Recognition. 79-84, 2004.

  With respect to speech recognition, when the HMM parameters are obtained using an error minimization learning algorithm or EM algorithm, it may converge to a solution that is not necessarily optimal, and the method using the HMM further improves the recognition performance. I can expect. In general, speech recognition is affected by ambient noise and the like, so it is difficult to recognize with 100% probability. However, since the likelihood of a conventional acoustic model such as HMM does not give an index of probability, speech recognition performance is not high. Therefore, there has been a demand for a speech recognition apparatus and a speech recognition method that can output recognition results based on probability and can reliably improve speech recognition performance as compared with conventional methods.

  The present invention provides a speech recognition apparatus and a speech that can output a recognition result based on a probability for a recognition unit such as a word indicating speech content and can reliably improve speech recognition performance as compared with a conventional method using an acoustic model such as an HMM. An object is to provide a recognition method.

  In order to solve the above problem, in the speech recognition apparatus according to claim 1, for example, as shown in FIG. 1, a feature extraction unit 4 that extracts a feature vector time series such as a mel cepstrum coefficient time series from a speech signal; Acoustic model likelihood calculation for calculating a likelihood function group related to a recognition unit such as a word based on an acoustic model such as a hidden Markov model based on a feature vector time series related to training speech data extracted by the feature extraction unit 4 The first PLRM acoustic learning is performed by taking the likelihood function group obtained by the unit 8 and the acoustic model likelihood calculating unit 8 into the regression function φ and maximizing the PLRM coefficient likelihood with the PLRM coefficient parameter W. PLRM punishment is performed by incorporating the likelihood function group obtained by the unit 6 and the acoustic model likelihood calculating unit 8 into the regression function φ and by using the optimization method such as the steepest descent method. The second PLRM acoustic learning unit 5 that learns by maximizing the associated likelihood, the coefficient parameter W optimized by the first PLRM acoustic learning unit 6 and the second PLRM acoustic learning unit 5 are optimized. A speech recognition unit 9 that calculates the probability of the recognition unit expressed by probability using the hyperparameter Λ, and the acoustic model likelihood calculation unit 8 incorporates the hyperparameter Λ into parameters related to the acoustic model. Compute the likelihood function group.

Here, the audio signal includes audio data recorded in the storage means in addition to the audio signal received by the receiving means. The acoustic model is a conventional probability model capable of calculating the likelihood, and is typically a hidden Markov model (HMM), but is not limited thereto. Learning the parameter W means determining the parameter W so that the likelihood with fines calculated for all training speech data is maximized. Learning the parameter Λ has the highest likelihood with fines. The parameter Λ is determined so as to increase. The acoustic model likelihood calculation unit 8 only needs to be configured so that the hyperparameter Λ can always be calculated without taking the hyperparameter Λ into the parameter related to the acoustic model and calculating the likelihood function group.
If comprised in this way, the recognition result by probability can be output about recognition units, such as a word which shows the audio | voice content, and it can use also as the parameter | index of the reliability. In addition, it is possible to provide a speech recognition apparatus in which the PLRM and the acoustic model are combined and the speech recognition performance can be reliably improved as compared with a conventional method using an acoustic model such as HMM.

  The speech recognition method according to claim 2 is extracted by a feature extraction step (S011) for extracting a feature vector time series such as a mel cepstrum coefficient time series from a speech signal and a feature extraction step as shown in FIG. An acoustic model likelihood calculation step (S002, S005) for calculating a likelihood function group related to a recognition unit such as a word based on an acoustic model such as a hidden Markov model based on a feature vector time series related to the training speech data; A first PLRM acoustic learning step (S003) in which the likelihood function group obtained in the acoustic model likelihood calculation step is taken into the regression function φ to learn the coefficient parameter W of the PLRM by maximizing the likelihood with PLRM fineness (S003). , S006), and the likelihood function group obtained in the acoustic model likelihood calculation step is taken into the regression function φ and the hyperparameter Λ of the PLRM is calculated by the steepest descent method The second PLRM acoustic learning step (S004) for learning by maximizing the PLRM fine likelihood with which optimization method, the coefficient parameter W and the second PLRM optimized in the first PLRM acoustic learning step A speech recognition step (S008) that calculates a probability representation for the recognition unit using the hyperparameter Λ optimized in the acoustic learning step, and the acoustic model likelihood calculation step includes the hyperparameter Λ A likelihood function group is calculated by taking in parameters relating to the acoustic model.

Here, the acoustic model likelihood calculation step, the first PLRM acoustic learning step, and the second PLRM acoustic learning step may be repeated any number of times. In general, the greater the number of repetitions, the better the voice recognition performance. The acoustic model likelihood calculation step may be configured to calculate at any step without always taking the hyperparameter Λ into the parameters related to the acoustic model and calculating the likelihood function group.
If comprised in this way, the recognition result by a probability can be output about recognition units, such as a word which shows an audio | voice content, and it can also be used as an index of the reliability. In addition, by combining the PLRM and the acoustic model, it is possible to provide a speech recognition method that can surely improve the speech recognition performance as compared with a conventional method using an acoustic model such as an HMM.

  A program according to a third aspect is a program for causing a computer to execute the speech recognition method according to the second aspect.

  According to the present invention, a recognition result based on probability can be output for a recognition unit such as a word indicating speech content, and can also be used as an index of reliability thereof. Also, a PLRM and an acoustic model are combined to form an HMM or the like. It is possible to provide a speech recognition apparatus and a speech recognition method that can improve the speech recognition performance reliably as compared with the conventional method using the acoustic model.

  Embodiments of the present invention will be described below with reference to the drawings. Speech learning by the learning machine PLRM is performed by using the likelihood to the M-th candidate obtained from a speech recognition apparatus using a typical hidden Markov model (HMM) as a conventional acoustic model as an input regression function of the PLRM. A word candidate representing the utterance content is determined probabilistically. Also, the parameters of the acoustic model such as HMM, which is a hyperparameter in PLRM, are adjusted by an optimization method such as steepest descent based on the evaluation criterion called fine likelihood defining PLRM, and PLRM and acoustic The model combines to enhance the learning effect.

  FIG. 1 shows a block diagram of a configuration example of a speech recognition apparatus 100 according to an embodiment of the present invention. In FIG. 1, 1 is a voice, 2 is voice data for training, and 100 is a voice recognition device. The speech recognition apparatus 100 includes a reception unit 3 that receives a speech signal, a feature extraction unit 4 that extracts a feature vector time series such as a mel cepstrum coefficient time series from the received speech signal 1 or training speech data 2, An acoustic model database 7 that stores feature vector time series information related to the utterance content of a recognition unit such as a word, parameters related to HMM, and the like, and a feature vector time series acquired by the feature extraction unit 4 (particularly in training speech data 2 The feature vector time series) with reference to the feature vector time series information stored in the acoustic model database 7, the likelihood of the HMM related to the recognition unit candidate is calculated based on the HMM, or the parameter given from the PLRM learning unit 30 An acoustic model likelihood calculation unit 8 that calculates a likelihood function group for the values of W and Λ, and an acoustic model likelihood calculation unit 8 A first PLRM acoustic learning unit 6 that learns by incorporating a likelihood function group related to recognition unit candidates into the regression function φ and maximizing the PLRM coefficient-likelihood parameter coefficient W, and the likelihood A second PLRM acoustic learning unit 5 that learns the function group by incorporating the function group into the regression function φ and maximizing the PLRM hyper-parameter Λ by the optimization method such as the steepest descent method and the like. The probability parameter for the recognition unit is calculated using the coefficient parameter W optimized in the 6th part of PLRM acoustic learning and the hyperparameter Λ optimized in the second PLRM acoustic learning 5, and the utterance content is calculated. A speech recognition unit 9 that performs probabilistic determination of the above, a display unit 10 that displays a calculation result of the speech recognition unit 9, the second PLRM acoustic learning unit 5, the first PLRM acoustic learning unit 6, and the speech unit It comprises a storage unit 11 that appropriately stores calculation results, intermediate data, and the like in the recognition unit 9, and a control unit 11 that controls each unit constituting the speech recognition device 100 to perform the function as the speech recognition device 100. .

  Here, learning the parameter W means that the parameter W is determined so as to maximize the fine likelihood calculated for all training speech data, and learning the parameter Λ means that the fine likelihood is The parameter Λ is determined so as to be the largest. The likelihood function group is obtained by providing a plurality of values of parameters of the acoustic model from outside the likelihood calculation unit or by learning and determining the likelihood of the plurality of acoustic models for those values. It means to calculate.

  Among these, the first PLRM acoustic learning unit 6 and the second PLRM acoustic learning unit 5 constitute a PLRM learning unit 30, and the first PLRM acoustic learning unit 6 and the second PLRM acoustic learning unit 5 use the parameters. Learning is repeated while exchanging information W and Λ to improve recognition performance. The acoustic model database 7 and the acoustic model likelihood calculation unit 8 constitute a part of the acoustic model unit 40, and the acoustic model unit 40 has a speech recognition function and a likelihood calculation function based on the HMM. The PLRM learning unit 30 receives a likelihood function group to be incorporated into the regression function φ from the acoustic model unit 40 and provides the parameter Λ to the acoustic model unit 40. The PLRM learning unit 30 and the speech recognition unit 9 constitute a part of the PLRM unit 20 that executes processing based on the PLRM, and the PLRM unit 20 has various functions of the PLRM.

  FIG. 2 shows a processing flow example of the speech recognition method in the present embodiment. The frame shows an example of a processing flow in the PLRM learning unit 30.

  The receiving unit 3 receives the audio signal 1 (S010). A feature vector time series X such as a mel cepstrum coefficient time series is extracted from the audio signal 1 or training audio data 2 received by the receiving unit 3 (S011).

  Training voice data 2 is used during learning. The acoustic model likelihood calculation unit 8 inputs the feature vector time series X, learns based on the HMM using an error minimization learning algorithm or an EM algorithm, and sets an initial value of the second parameter Λ of the acoustic model. Obtain (S001). The acoustic model likelihood calculation unit 8 refers to the feature vector time series information related to the utterance content of the recognition unit (such as a word) stored in the acoustic model database 7 for the feature vector time series X of the training speech data 2, The likelihood of the HMM is calculated for a plurality of (M) recognition unit candidates, and an optimized likelihood function group is obtained. In addition to the word, the recognition unit includes a predicate, a phrase, a single sound, and the like.

で表され、ＰＬＲＭの罰金付き尤度は、

で表される（非特許文献３、４、５参照）。In the first PLRM acoustic learning unit 6, the likelihood of the calculated acoustic model of each candidate is used as the PLRM regression function φ (x; Λ) = (φ ₁ (x; Λ), φ ₂ (x; Λ), ..., Φ _M (x; Λ)) are input to the PLRM (S002). In PLRM, the acoustic model parameter Λ is treated as a PLRM hyperparameter. In PLRM, the probability of a recognition unit such as a word is

The PLRM fine likelihood is

(See Non-Patent Documents 3, 4, and 5).

とのバランスを取るために用いられるパラメータである。Γ、Σ、δはデータに応じて、適宜、事前に選択しておく。なお、Σの選択については後述する。Where K is the total number of recognition units such as words (1 ≦ M ≦ K), W = (w ₁ , w ₂ ,..., W _K ) is a coefficient parameter (first parameter) of PLRM, and its elements w _k is an M-dimensional weight vector. θ = (W, Λ). Also, D = {(x ⁿ , y ⁿ )} _n = ₁ ,. . . , _N represents represents and N training speech signal x ^n, a set of labels y ⁿ of recognition units, such as the word. Γ is a matrix that corrects the bias due to the class of learning data, and is, for example, a diagonal matrix corresponding to the number of training samples in which the kth diagonal component has the kth label. Σ is a positive definite matrix, δ is

It is a parameter used to balance Γ, Σ, and δ are appropriately selected in advance according to the data. The selection of Σ will be described later.

  The first PLRM acoustic learning unit 6 learns and optimizes the PLRM coefficient parameter W (S003). That is, the parameter W is determined so that the likelihood with fine calculated for all the training speech data is maximized. In the second PLRM acoustic learning unit 5, the coefficient parameter W of the PLRM is fixed, and the acoustic model parameter Λ is learned by maximizing the likelihood with PLRM fineness using an optimization method such as the steepest descent method. (S004).

The acoustic model unit 40 receives the parameter Λ obtained by the PLRM learning unit 30, and the acoustic model likelihood calculation unit 8 uses this to calculate the likelihood of the acoustic model. The PLRM learning unit 30 receives the likelihood of each recognition unit candidate calculated by the acoustic model unit 40 and inputs the likelihood to the first PLRM acoustic learning unit 6 as a regression function of the PLRM (S005). The parameter W is learned (S006). Here, the speech recognition unit 9 can also obtain an expression based on the probability p _k (θ) of a recognition unit such as a word by using (Expression 1) using the parameters W and Λ. However, it is preferable to learn the parameters W and Λ by repeating S004 to S006 to improve the validity of the probability p _k (θ). That is, the second PLRM acoustic learning unit 5 fixes the newly obtained PLRM coefficient parameter W and optimizes the acoustic model parameter Λ using an optimization method such as the steepest descent method (S004). . Then, the steps from S004 to S006 are sufficiently repeated, and when it is determined that the coefficient parameter W of the PLRM and the parameter Λ of the acoustic model have converged, the acoustic learning by the PLRM is terminated (S007).

  During the test, the audio signal 1 is received by the receiving unit 3 (S010), and the feature extraction unit 4 extracts a feature vector time series such as a mel cepstrum coefficient time series (S011). The feature vector time series is input to the learned PLRM unit 20, and is estimated based on the probability of recognition unit candidates such as words corresponding to the input speech (S008). A candidate indicating the maximum probability is displayed as a recognition result on the display unit 10 (S009). The whole element of the calculated probability can be used as an index that gives the certainty of the recognition result.

Next, selection of Σ will be described. The simplest option is Σ = I. Here, I represents a unit matrix. Another option is the excessive learning was considered Σ = (1 / N) ΦΦ T to training speech data. (See Non-Patent Documents 3, 4, and 5)

Finally, a recognition experiment using the voice recognition apparatus will be described.
Experiments were performed on the E set of the TI46 database (Table 1). This set includes 1433 training sounds and 2291 test sounds. A 39-dimensional feature vector consisting of a 13-dimensional mel cepstrum coefficient time series and its primary and secondary regression coefficients was extracted from each speech signal. The feature vector extraction used a 25 ms Hamming window and a 10 ms window shift. Each word was modeled by an HMM with a 5 Gaussian mixture distribution. For [delta], [delta] = 0.01 If the preliminary experiment sigma ₁ = I is, in the case of _{Σ 2 = (1 / N)} ΦΦ T is set to [delta] = 5000.

Table 2 shows the result of comparing the word recognition rates of the method of the present invention and the conventional method.

  Further, the present invention can be realized as a program for causing a computer to execute the speech recognition method according to the embodiment, and also as a computer-readable recording medium on which the program is recorded. The program may be used by being recorded in a ROM built in the computer, or may be recorded on a recording medium such as an FD, CD-ROM, built-in or external magnetic disk, read by the computer, and used via the Internet. May be downloaded to a computer and used.

  Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and various modifications can be made to the embodiments without departing from the spirit of the present invention. It is obvious.

  For example, in the above embodiment, for example, the second PLRM acoustic learning step of learning the second parameter Λ of the acoustic model by maximizing the likelihood with a PLRM fine by an optimization method such as the steepest descent method The example in which the first PLRM acoustic learning step (S006) for learning by optimizing the parameter W of the PLRM has been described from (S004), but the number of repetitions is arbitrary. Moreover, although the conventional acoustic model demonstrated the example of the hidden Markov model (HMM), you may use the likelihood of other acoustic models, such as an acoustic model based on a Bayesian network. The training voice data is not limited to the E set used in the experiment.

  The present invention can be used for speech recognition.

It is a block diagram of the structural example of the speech recognition apparatus in embodiment of this invention. It is a figure which shows the example of a processing flow of the speech recognition method in embodiment of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Voice signal 2 Training voice data 3 Reception part 4 Feature extraction part 5 2nd PLRM acoustic learning part 6 1st PLRM acoustic learning part 7 Acoustic model database 8 Acoustic model likelihood calculation part 9 Speech recognition part 10 Display part 11 Storage unit 12 Control unit 20 PLRM unit 30 PLRM learning unit 40 Acoustic model unit 100 Speech recognition device W Coefficient parameter Λ Hyper parameter

Claims (3)

A feature extraction unit that extracts a feature vector time series such as a mel cepstrum coefficient time series from a speech signal;
Acoustic model likelihood calculation for calculating a likelihood function group related to a recognition unit such as a word based on an acoustic model such as a hidden Markov model based on a feature vector time series related to training speech data extracted by the feature extraction unit Part;
The likelihood function group obtained by the acoustic model likelihood calculation unit is taken into a regression function, and a coefficient parameter of a penalty logistic regression machine (hereinafter referred to as PLRM in this claim) is used as a coefficient parameter, and a likelihood with a PLRM penalty is used. A first PLRM acoustic learning unit that learns by maximizing;
The likelihood function group obtained by the acoustic model likelihood calculation unit is taken into the regression function, and the hyperparameter of the PLRM is learned by maximizing the likelihood with PLRM fineness by an optimization method such as a steepest descent method. A second PLRM acoustic learning unit;
Using the coefficient parameter optimized by the first PLRM acoustic learning unit and the hyperparameter optimized by the second PLRM acoustic learning unit, a probability representation of the possibility for the recognition unit is calculated. A voice recognition unit for
The acoustic model likelihood calculating unit calculates the likelihood function group by taking the hyper parameter into a parameter related to the acoustic model;
Voice recognition device. A feature extraction step of extracting a feature vector time series such as a mel cepstrum coefficient time series from a speech signal;
Acoustic model likelihood calculation for calculating a likelihood function group related to a recognition unit such as a word based on an acoustic model such as a hidden Markov model based on a feature vector time series related to training speech data extracted in the feature extraction step The process;
The likelihood function group obtained in the acoustic model likelihood calculation step is taken into the regression function, and the coefficient parameter of the Penalized Logistic Regression Machine (hereinafter referred to as PLRM in this claim) is set as the coefficient parameter of the PLRM fine. A first PLRM acoustic learning step that learns by maximizing;
The likelihood function group obtained in the acoustic model likelihood calculation step is taken into the regression function, and the hyperparameter of the PLRM is learned by maximizing the likelihood with PLRM fine by an optimization method such as a steepest descent method. A second PLRM acoustic learning step;
Using the coefficient parameter optimized in the first PLRM acoustic learning step and the hyper parameter optimized in the second PLRM acoustic learning step, a probability representation of the probability for the recognition unit is calculated. A voice recognition step to perform;
The acoustic model likelihood calculating step calculates the likelihood function group by taking the hyperparameter into a parameter related to the acoustic model;
Speech recognition method.   A program for causing a computer to execute the speech recognition method according to claim 2.

Images