JPH08241092A - Speaker adaptation method for acoustic model and device therefor - Google Patents

Abstract

PURPOSE: To realize a recognition system of high performance which minimizes the error rate of an acoustic HMM. CONSTITUTION: This device has a characteristic parameter extraction section 1 which forms and holds the acoustic HMM(hidden Markov model) 6 for nonspecific speakers formed by learning from the speeches of many speakers and extracts a characteristic parameter from speech data 5 of the person to be recognized and a likelihood maximizing adaptation section 2 which optimizes the parameter of the acoustic HMM for the nonspecific persons so as to maximize the likelihood for the speech of a person to be recognized. Further, the device has an identification error minimizing adaptation section 3 which obtaines the acoustic HMM of the min. in the identification error by defining the differentiable loss function from the parameter of the acoustic HMM having the parameter maximized in the likelihood and the time series speech data of the characteristic parameter of the person to be recognized and selecting the parameter so as to minimize the function and an adaptation acoustic HMM accumulation section 4.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method and apparatus, and more particularly, to a speech HMM for an unspecified speaker corresponding to a recognition category such as a phoneme or a word, which is obtained by modeling a speech feature of speech by an HMM. The present invention relates to a speaker adaptation method and apparatus for an acoustic HMM adapted to a specific recognition target speaker.

[0002]

2. Description of the Related Art Conventionally, in unspecified speaker voice recognition, attempts have been made to improve recognition performance by adapting a recognition system to a recognition target speaker. In order to maximize the likelihood of the acoustic HMM for the unspecified speaker with respect to the voice of the recognition target speaker, for example, the document “Seiji Nakagawa:“ Speech recognition by probabilistic model ”IEICE, 1988” (hereinafter referred to as Document 1). Baum-Wel (Baum-Wel)
ch) The parameters of the acoustic HMM are estimated according to the algorithm to adapt the sound HMM to the recognition target speaker.

[0003]

In the above-described conventional speaker adaptation method of the acoustic model, it is the actual situation that good performance is not obtained with respect to the recognition error rate.

It is an object of the present invention to provide a speaker adaptation method for an acoustic model and a device therefor which can obtain a better performance in terms of recognition error rate.

[0005]

A speaker adaptation method for an acoustic model according to the present invention extracts acoustic characteristics of a speech, statistically models the characteristic quantity, and classifies the acoustic characteristics into phonemes, words and other recognition categories. In order to construct a corresponding acoustic model, an acoustic model for an unspecified speaker, which has been learned by using the voices of a large number of speakers, is represented by a hidden Markov model, which is abbreviated as HMM. In the speaker adaptation method of the acoustic model, the parameters of the acoustic HMM for the unspecified speaker are optimized using the speech of the speaker so that the likelihood for the speech of the recognition target speaker is maximized. There is a step of adapting the parameters of the acoustic HMM for the unspecified speaker optimized so that the likelihood for the speaker's voice is maximized so that the recognition error for the voice of the recognition-target speaker is minimized.

Further, the step of adapting the parameters of the acoustic HMM for the unspecified speaker so as to minimize the recognition error with respect to the voice of the recognition target person defines a differentiable loss function so that this value decreases. The present invention also includes the step of sequentially updating the parameters of the acoustic HMM to obtain an optimum value.

The speaker adaptation device of the acoustic model of the present invention comprises:
In order to construct an acoustic model corresponding to phonemes, words, and other recognition categories by extracting the acoustic features of speech and statistically modeling the features, we learned the learning using speech from many speakers. The acoustic model for a specific speaker is represented by a hidden Markov model, which is abbreviated as HMM, and the voice of the speaker to be recognized is used to generate the acoustic HM for the unspecified speaker.
In a speaker adaptation device of an acoustic model for optimizing the parameters of M so that the likelihood for the speech of the recognition target speaker is maximized, the optimal is such that the likelihood for the speech of the recognition target speaker is maximized. It has an adapting means for adapting the parameters of the converted acoustic HMM for the unspecified speaker so that the recognition error with respect to the voice of the recognition target speaker is minimized.

Further, the adaptation means for adapting the parameters of the acoustic HMM for the unspecified speaker so as to minimize the recognition error with respect to the voice of the recognition target person defines a differentiable loss function, and this value decreases. As described above, a device including means for sequentially updating the parameters of the acoustic HMM to obtain an optimum value is also included in the speaker adaptation device of the acoustic model of the present invention.

[0009]

After the parameters of the acoustic HMM for an unspecified speaker are optimized using the sounds of a large number of speakers so as to maximize the likelihood with respect to the sounds of a specific speaker, Since adaptation is performed so that the recognition error with respect to the voice of the recognition target speaker is minimized, adaptation of the acoustic model speaker with a low error rate becomes possible.

[0010]

Embodiments of the present invention will now be described with reference to the drawings.

FIG. 1A is a flow chart of an embodiment of the acoustic model speaker adaptation method of the present invention, and FIG. 1B is FIG.
It is a flowchart of the embodiment which used the loss function for the identification error rate minimization adaptation of step 13 of (A).

In this speaker adaptation method of an acoustic model, an acoustic model for an unspecified speaker learned by using the voices of a large number of speakers is an acoustic HMM (Hidden Mar) which is a hidden Markov model.
kovModel) (step 11). Next, using the voice of the specific speaker to be recognized, the parameters of the acoustic HMM are optimized so that the likelihood for the voice of the speaker to be recognized is maximized (step 12). Further, the parameters of the acoustic HMM output in step 12 are adapted so that the identification error with respect to the voice of the recognition target speaker is minimized (step 13). The acoustic HMM output from step 13 is stored (step 14).

Also, the identification error minimization adaptation step 13 is performed.
Is the loss function

[0014]

[Equation 1] To reduce

[0015]

[Equation 2] Here, ε _t is a coefficient for adjusting the update amount, and is set experimentally.

The embodiment including the step of sequentially updating Λ _t in Equation (6) to obtain the optimum value is also included in the present invention.

FIG. 2 is a block diagram of an apparatus to which the acoustic model speaker adaptation method of the present invention is applied.

This acoustic model speaker adaptation apparatus includes a hidden Markov model (hereinafter referred to as an acoustic model) for an acoustic model for an unspecified speaker in which voices of a large number of speakers are recorded and statistically modeled for the feature amount of the voice. HMM (HIDDEN MARKOV MODEL)
6, a feature parameter extraction unit 1 for converting the input voice into an expression format using a feature parameter such as a cepstrum when the voice data 5 of the recognition target speaker is input, an output of the feature parameter extraction unit 1, and the HMM 6 Is input, a highly likely acoustic HMM adapted to the recognition target speaker by the voice data converted into the recognition target speaker in time series, and the speech converted into the recognition target speaker in time series. A likelihood maximization adaptation unit 2 that outputs data, and an identification error minimization adaptation unit 3 that receives the output of the likelihood maximization adaptation unit 2 and modifies the parameters of the acoustic HMM so as to minimize the identification error. , An adaptive acoustic HM for accumulating the output of the identification error minimization adaptation unit 3
It has an M storage unit 4.

In the likelihood maximization adaptation unit 2, the HMM parameter Λ = of the input acoustic HMM 6 is

[0020]

[Outside 1] So that the likelihood function L _K (·) shown in Equation 1 is maximized,
For example, the Baum / Welch algorithm (see Reference 1) is used to adapt the speaker to be recognized.

Where α is a state, m is a mixture distribution,

[0022]

[Outside 2] Is the weighting coefficient of the mixture distribution m of the state α,

[0023]

[Outside 3] Is the average value,

[0024]

[Outside 4] Is the variance value,

[0025]

[Outside 5] Represents the transition probability from the state α to the state β.

[0026]

(Equation 3) Here, X = {x ₁ , x ₂ , x ₃ ... x _T ) is voice data converted into time series of feature parameters, t is time, π is initial state probability, Q = {q ₀ , q ₁ , ... Q _T ) represent a state transition sequence.

In addition,

[0028]

[Equation 4] Is the probability of occurrence.

In the identification error minimization adaptation section 3, the acoustic HMM is modified so that the identification error function d _k (.) Is minimized. Here, the discrimination function g _K (·) and the discrimination error function d
_k (·) is defined by equations (3) and (4), respectively.

[0030]

(Equation 5) L _k is a likelihood function based on the Viterbi algorithm (according to Reference 1).

In addition, in the identification error minimization adaptation section 3, for example, in the parameter correction of the acoustic HMM, the identification error function d
Instead of _k (•), differentiable loss function l shown in Equation 5
(D _k ) is defined and equation (6) is used so that this function decreases.
The acoustic parameters of are sequentially updated. Where equation (5),
(6) is a reference [BH Juang and Katagiri: "Discriminat
ivetrain-ing "J. Acoust. Soc. Jpn., 13, 6, pp. 333-3
39, 1992.].

[0032]

(Equation 6) Here, ε _t is a coefficient for adjusting the update amount, and is set experimentally.

Next, the operation results of this embodiment will be described with reference to experimental examples.

First, an acoustic HMM for an unspecified speaker was created. The acoustic HMM of this example is a semi-continuous HMM with a mixture distribution number of 256, and there are 43 types of acoustic HMM independent of the phonological environment. To create an acoustic HMM for unspecified speakers, a male 35
Using a total of 7,016 sentences as the name, the HMM parameters were estimated by the Baumwelch algorithm.

In addition, for speaker adaptation, the speaker-specific sound H is used.
It was evaluated by the phoneme recognition rate in continuous speech when data of 10 and 50 sentences were used for two men different from the speaker used for creating the MM. In the phonological recognition experiment, the voice content is written down, the Viterbi alignment is taken for the speech section, and it is assumed that it is the correct phonological section, and the one showing the maximum likelihood among all the phonological HMMs in the phonological section. The recognition result. The phonological discrimination experiment was performed using 100 sentences different from the same sentence set used for speaker adaptation. As a characteristic parameter, sample frequency 12 KHz, frame length 32 ms, frame period 8 m
s, LPC (Linear Prediction Coefficient) The cepstrum was extracted in the analysis order 16.

Further, in speaker adaptation by likelihood maximization,
The learning was repeated 5 times, and only the average value of each mixture distribution was estimated. Further, in the speaker adaptation by minimizing the discrimination error, the number of iterations was set to 10 and the average value and weighting coefficient of each mixture distribution were estimated. In both cases, in each iteration, the HMM parameters were updated all at once after applying the pre-learning data.

The results of the above experiment are shown in Table 1, and it is clear that this method is more effective than the speaker adaptation method of the conventional acoustic model.

[0038]

[Table 1]

[0039]

As described above, according to the present invention, the acoustic HMM for the unspecified speaker is subjected to the likelihood maximization adaptation and then the HMM parameter identification error minimization adaptation of the acoustic HMM is performed. Speaker adaptation is enhanced, and a high-performance acoustic recognition system can be realized.

[Brief description of drawings]

1A is a flowchart of an embodiment of an acoustic model speaker adaptation method of the present invention, and FIG. 1B is step 1 of FIG.
3 is a flowchart of an embodiment using a loss function for the identification error rate minimization adaptation of FIG.

FIG. 2 is a block diagram of an embodiment of a speaker adaptation device for an acoustic model of the present invention.

[Explanation of symbols]

 1 Feature Parameter Extraction Unit 2 Likelihood Maximization Adaptation Unit 3 Identification Error Minimization Adaptation Unit 4 Adapted Acoustic HMM Storage Unit 5 Speech Data of Speaker to be Recognized 6 Acoustic HMM for Unspecified Speaker

Claims (4)

[Claims] 1. To extract acoustic features of a voice and statistically model the feature amount to construct an acoustic model corresponding to a phoneme, a word and other recognition categories, voices of a large number of speakers are extracted. The acoustic model for the unspecified speaker learned by using it is represented by a hidden Markov model, which is abbreviated as HMM, and the speech of the speaker to be recognized is used to set the parameters of the acoustic HMM for the unspecified speaker. In the speaker adaptation method of the acoustic model that optimizes the likelihood for the speech of the recognition target speaker to be maximum, and is optimized to maximize the likelihood for the speech of the recognition target speaker. A speaker adaptation method for an acoustic model, comprising a step of adapting a parameter of an acoustic HMM for an unspecified speaker so as to minimize a recognition error with respect to a voice of a recognition target speaker. 2. The step of adapting the parameters of the acoustic HMM for an unspecified speaker so as to minimize the recognition error with respect to the speech of the recognition target speaker defines a differentiable loss function, and this value is reduced. 2. The speaker adaptation method for an acoustic model according to claim 1, further comprising the step of sequentially updating the parameters of the acoustic HMM to obtain an optimum value. 3. A plurality of speakers' voices are extracted in order to extract an acoustic feature of the voice and statistically model the feature amount to construct an acoustic model corresponding to a phoneme, word or other recognition category. The acoustic model for the unspecified speaker learned by using it is represented by a hidden Markov model, which is abbreviated as HMM, and the speech of the speaker to be recognized is used to set the parameters of the acoustic HMM for the unspecified speaker. In a speaker adaptation device for an acoustic model that optimizes the likelihood for the speech of the recognition-target speaker to be maximum, and is optimized so that the likelihood for the speech of the recognition-target speaker is maximized. A speaker adaptation apparatus for an acoustic model, comprising an adaptation means for adapting parameters of an acoustic HMM for an unspecified speaker so that a recognition error with respect to a voice of a recognition target speaker is minimized. 4. An adapting means for adapting the parameters of the acoustic HMM for an unspecified speaker so as to minimize a recognition error with respect to a voice of a recognition target person defines a differentiable loss function, and this value is A means for obtaining an optimum value by sequentially updating the parameters of the acoustic HMM so as to decrease.
Speaker adaptation device of the described acoustic model.