JPH06266384A - Acousting model adaption system - Google Patents

Abstract

PURPOSE:To improve performance such as the speech recognition rate of a tutor-less speaker adaption system. CONSTITUTION:A probability model (all-phoneme Ergodic HMM) is generated by the ergodic combination of all phonemes learnt from the input speech of a standard speaker by a phoneme bigram probability value. Then, the mean value vector mu of an output probability distribution as a parameter of the probability model is learnt by a maximum likelihood estimating method of Baum- Welch by using an input speech whose vocalization contents are unknown and the mean value vector mu is corrected by a moving vector field smoothing system. Then, the learning and the correction of the mean value vector is repeated until the value of output likelihood to input speed data is converged.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an acoustic model adaptation system, and more particularly to an unsupervised adaptation system for adapting an acoustic model to a feature space of an input speech according to a speaker, a speech style, a speech environment, or the like.

[0002]

2. Description of the Related Art Conventionally, a method based on mapping of a vector quantization codebook or a principle of continuous distribution type Hidden Markov Model (HMM; hidden Markov mode) has been used as an acoustic model adaptation method without teacher data relating to utterance contents.
The technique applied to the speaker based on the distribution of the voice pattern, such as the method applied to L), is the technical report of IEICE Technical Report SP88-21, SP88-122, SP90-6.
7, etc.

[0003]

However, in the unsupervised speaker adaptation method, it is difficult to perform accurate mapping because the linguistic information relating to the utterance content is not used. Therefore, there is a problem that the unsupervised speaker adaptation method is generally inferior in performance and efficiency as compared with the supervised speaker adaptation method using voice data whose utterance content is known.

The present invention solves these problems, and provides an unsupervised speaker adaptation method that adapts to an acoustic model and a feature space of the input speech by using an input speech whose utterance content is unknown. The purpose is to improve performance and efficiency.

[0005]

The gist of the acoustic model adaptation method according to the present invention is a plurality of acoustic models that express the characteristics of speech for use in speech recognition. In the acoustic model adaptation method in which what is learned by the voice of the standard speaker is adapted to the feature space of the input voice whose utterance content is unknown, all of the acoustic models are connected to each other by a desired transition probability, and the acoustic model The object is to create a probabilistic model which itself is self-connected with a desired transition probability, and re-learn all or part of the parameters of the probabilistic model with the input speech.

Further, in the acoustic model adaptation method, a discrete distribution type, a continuous distribution type or a semi-continuous distribution type HMM of phonemes is used as the acoustic model.

In the acoustic model adaptation method,
A phoneme bigram probability value obtained from desired text data is used as the initial value of the transition probability.

On the other hand, in the above acoustic model adaptation method,
A context-dependent phoneme model is used as the acoustic model, and a context-dependent phoneme bigram probability value obtained from desired text data is used as an initial value of the transition probability.

Further, in the above acoustic model adaptation method,
This is to use a motion vector field smoothing method when retraining all or some of the parameters of the above probabilistic model with the above input speech.

[0010]

According to the acoustic model adaptation method according to the present invention, a probabilistic model in which an acoustic model is connected by a certain transition probability is created, and various parameters of the probabilistic model use the input speech whose utterance content is unknown. The acoustic model is adapted to the feature space of the input speech by being retrained. Therefore, the performance such as the voice recognition rate is improved.

[0011]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, an embodiment of an acoustic model adaptation system according to the present invention will be described with reference to the drawings.

FIG. 2 is a conceptual diagram showing a stochastic model in the unsupervised speaker adaptation system which is an embodiment of the present invention.

As shown in FIG. 2, first, a mixed continuous distribution type phoneme HMM1 learned by using a standard speaker's input voice.
Are connected to the ergodic with a certain transition probability a _ij to create one large-scale probabilistic model. here,
49 phonemes HMM1 including silence are used. However, in FIG. 1, only four phonemes HMM1 are shown for the sake of simplicity. Hereinafter, this probabilistic model is referred to as "whole phoneme ergodic HMM".

That is, this whole phoneme ergodic HM
M is a transition probability a with all 49 phoneme HMM1s.
_The phonemes HMM1 themselves are connected to each other by _ij , and are also self-connected by a certain transition probability a _ij . The phoneme HMM1 is a kind of acoustic model that expresses the characteristics of a voice used for voice recognition.

Since the transition probabilities a _ij between the phonemes HMM1 and in each phoneme HMM1 in this all-phoneme ergodic HMM correspond to the phoneme bigram probability values, the initial values of the transition probabilities a _ij are: A phoneme bigram probability value obtained from some text data is used.

In FIG. 2, the transition probability a _ij is shown only at one location, but the same applies to other phonemes HMM1 and within the phoneme HMM1. Further, a _ij represents the transition probability from the i-th phoneme HMM1 to the j-th phoneme HMM1. Therefore, when i = j, it represents a self-transition within the same phoneme HMM1.

Therefore, this whole phoneme ergodic H
The MM is a probabilistic model in which a language model and an acoustic model are fused, and expresses all language speech. Here, since the voice uttered by the speaker is a “language voice” including acoustic information and linguistic information, the whole phoneme ergodic HMM is obtained by using the input voice of the input speaker whose utterance content is unknown. It is possible to learn each of the parameters by the maximum likelihood estimation method.

As described above, since the linguistic information about the utterance content can be stochastically used by performing the learning, it is possible to improve the performance of speaker adaptation as compared with the conventional method which does not use the linguistic information at all. Is possible.

By the way, in this unsupervised speaker adaptation method, when a large amount of learning data exists, all parameters in the all-phoneme ergodic HMM are re-learned by an input voice whose utterance content is unknown, It is possible to adapt acoustic and language models simultaneously.
The parameters here include transition probability a _ij between phoneme HMMs, and transition probability a _ij within phoneme HMM1.
_ij , mean value vector of output probability distribution, covariance matrix, and weighting coefficient of mixture distribution.

Therefore, a two-stage learning method is conceivable in which after learning the transition probability between the phoneme HMM1s which are the parameters of the acoustic model, the learning probability between the phoneme HMM1s which is the parameter of the language model is learned. In this case, a certain amount of learning data is needed. For this reason, in the following examples, the transition probability a between the phonemes HMM1 is fixed and the adaptation effect in the phoneme HMM1 which is considered to have the highest adaptation effect among the parameters of the phoneme HMM1 is premised on adaptation by a smaller amount of learning data. A case of re-learning only the average value vector μ of the output probability distribution will be described.

FIG. 1 is a flow chart showing a learning algorithm in such a case. As shown in FIG. 1, first, in step S1, phoneme HMMs are created for all phonemes by the voice of a standard speaker, and in step S2, phoneme bigram probability values are calculated using text data.

Next, in step S3, these phoneme HMMs are connected to an ergodic by each phoneme bigram probability value, and the whole phoneme ergodic HM shown in FIG. 2 is connected.
Create M.

Next, in step S4, the Baum-Welch (Baum-Welch) is used by using the input voice whose utterance content is unknown.
ch) Learn the average value vector μ of the output probability distribution by the maximum likelihood estimation method.

Next, in step S5, the average value vector μ of the output probability distribution is corrected by the moving vector field smoothing method. Since the moving vector field smoothing method is disclosed in detail in IEICE Technical Research Report SP92-16, it will be briefly described here.

First, a difference vector between the average value vector of the output probability distribution of the all-phoneme ergodic HMM retrained by the input speech by the maximum likelihood estimation method and the average value vector before adaptation is input from the standard speaker space. We consider it as a movement vector to the speaker space, and let the set be a movement vector field. In the case of unsupervised learning, there is a possibility that the mean value vector of the output probability distribution is re-learned due to erroneous phoneme data, so it is considered that this includes an estimation error. This estimation error also occurs when the learning sample is small. Therefore, it is considered that the direction of the movement vector obtained in this way is discontinuous. Furthermore, when the learning sample is small, there is also an average value vector of the output probability distribution that is not retrained.

Therefore, by introducing a "constraint condition of continuity" into the movement vector field, the movement vector is smoothed, and thereby the average value vector of the output probability distribution is corrected. Further, the movement vector with respect to the unlearned average value vector is interpolated by interpolation or extrapolation of another movement vector. Here, the smoothing strength of the movement vector is controlled by the value of fuzziness, and the larger this value, the stronger the smoothing. Therefore, when the fuzziness value is infinite, all phoneme models move in parallel.

Then, in step S6, it is determined whether or not the value of the output likelihood of the input voice data of the all-phoneme ergodic HMM has converged. If it has not converged, the process returns to step S4. That is, until the value of the output likelihood for the input voice data converges, the above step S
Repeat 4 and S5.

Therefore, if the value of the output likelihood has converged, the process proceeds to step S7, and steps S4 to S6 are performed.
The phoneme HMM1 in the all-phoneme ergodic HMM retrained in (3) is disconnected to be decomposed into each phoneme HMM.

By the above method, the phoneme HMM of the standard speaker
Is adapted to the feature space of the input voice using the input voice whose utterance content is unknown.

Next, the experimental results when one standard speaker model is adapted to another input speaker model by this unsupervised speaker adaptation method are shown below.

As a phoneme HMM for a standard speaker, a mixed continuous distribution HMM having four states, three loops and three mixtures was used. The number of phoneme HMMs was 49. For the phoneme HMM learning, 5240 words were used as the important words of the standard speaker and 216 words were used as the balance words. The phoneme bigram probability value was obtained from the text data. The speech utterance of the input speaker was used for speaker adaptation, and the evaluation was performed by a phoneme recognition experiment in 2560 words different from the adaptive learning of the input speaker.

As a result, the phoneme recognition rate, which was 70.2% in the standard speaker model before speaker adaptation, was 80.4% when learning with 25 words by this speaker adaptation method, 1
It was proved that the speaker adaptation method of the present invention was effective, which was 83.3% when learned by using 00 words and 87.6% when learned by using 200 words.

The embodiment of the present invention has been described above.
The present invention is not limited to the above-mentioned embodiments, but can be implemented in other modes.

For example, although the mixed continuous distribution type HMM is used as the acoustic model in the above-mentioned embodiment, a single continuous distribution type HMM may be used, or a discrete distribution type HMM may be used. Alternatively, a context-dependent phoneme model may be used as the acoustic model, and a context-dependent phoneme bigram probability value obtained from some text data may be used as the initial value of the transition probability. Regarding the context-dependent phoneme model, IEICE Technical Report SP91
"Stable Robust Phoneme Recognition by Smoothing Based on Phoneme Environment Tree Structure of Single Gaussian HMM", and S91-8.
Since it is disclosed in detail in "Automatic Generation of Hidden Markov Network by Sequential State Division with respect to Phoneme Text and Time" in Section 8, this is incorporated herein.

In addition, in the above embodiment, the phoneme H before adaptation is applied.
Although the model learned by one standard speaker is used as MM, an unspecified speaker model in which voice data of several speakers is learned may be used. Further, it can be applied not only to speaker adaptation, but also to speech style adaptation, speaker environment adaptation, and the like.

[0036]

As described above, according to the acoustic model adaptation method according to the present invention, it is possible to adapt the existing acoustic model to the feature space of the input speech by using the input speech whose utterance content is unknown. It will be possible. For this reason, the recognition performance is improved such that a speech recognition rate comparable to that of the supervised speaker adaptation method can be obtained. Furthermore, not only speaker adaptation, but also when applied to utterance style adaptation, utterance environment adaptation, etc., the recognition performance is similarly improved.

[Brief description of drawings]

FIG. 1 is a flowchart showing an algorithm of an embodiment of an acoustic model adaptation system according to the present invention.

FIG. 2 is a conceptual diagram showing a stochastic model in the acoustic model adaptation method shown in FIG.

[Explanation of symbols]

1 phoneme HMM a _ij phoneme bigram probability value

Claims (5)

[Claims] 1. A plurality of acoustic models for expressing speech features for use in speech recognition, which are learned by one or two or more standard speaker's voices and are used as input voices whose utterance contents are unknown. In an acoustic model adaptation method that adapts to a feature space, all of the acoustic models are connected to each other by a desired transition probability, and the acoustic model itself also creates a probability model in which it is self-connected by a desired transition probability. An acoustic model adaptation method characterized in that all or part of parameters of a model are retrained by the input speech. 2. The acoustic model adaptation method according to claim 1, wherein a discrete distribution type, a continuous distribution type or a semi-continuous distribution type hidden Markov model of phonemes is used as the acoustic model. 3. The acoustic model adaptation method according to claim 1, wherein a phoneme bigram probability value obtained from desired text data is used as an initial value of the transition probability. 4. The context-dependent phoneme model is used as the acoustic model, and the context-dependent phoneme bigram probability value obtained from desired text data is used as an initial value of the transition probability. The acoustic model adaptation method described. 5. The moving vector field smoothing method is used when retraining all or some of the parameters of the probabilistic model with the input speech. The acoustic model adaptation method described.