JPH0786758B2 - Voice recognizer - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device using a standard pattern created by a user's utterance.

[0002]

2. Description of the Related Art At present, in the field of voice recognition, research and development of an unspecified speaker recognition system aiming at recognizing anyone's voice is being actively conducted. In these recognition systems, the hidden Markov model (Hidden Marko) is used.
v Model, hereinafter abbreviated as HMM), a neural network (Neural Network, hereinafter abbreviated as NN), and other recognition methods are widely used. For details of the HMM, for example, “Speech recognition by probabilistic model”
It is explained in detail in the Institute of Electronics, Information and Communication Engineers (hereinafter referred to as Reference 1) by Seiichi Nakagawa, 1988. Further, regarding the speech recognition by the NN, for example, in "Speech / Hearing and Neural Network Model" edited by Shunichi Amari, 1990, Ohmsha (hereinafter referred to as Reference 2) is detailed. In these methods, a standard pattern is learned using learning data composed of words, sentences, and the like uttered by a large number of speakers in advance, and recognition is performed using the learned standard pattern.

The unspecified speaker system, unlike the specified speaker system in which the user is specified, has an advantage that the user does not need to register the utterance in advance. However, in recent years
The following problems were pointed out. First, most speakers have poorer recognition performance than the specific speaker system. Furthermore, there is a speaker (unique speaker) whose recognition performance is significantly poor. In order to solve such a problem, a study has recently been started to apply a speaker adaptation technique, which has been conventionally used in a specific speaker system, to an unspecified speaker system.

Speaker adaptation refers to a method of adapting the recognition system to a new user (unknown speaker) by using a smaller amount of learning data than that used for learning. For details of the speaker adaptation method, “Speaker Adaptation Technology in Speech Recognition”, Sadahiro Furui, Journal of Television Society, Vol. Four
3, No. 9, 1989, pp. 929-934 (hereinafter
Reference 3 )). Speaker adaptation can be roughly divided into two methods. One is unsupervised speaker adaptation, and the other is supervised speaker adaptation. The former is
Since it is not necessary to specify the vocabulary spoken by an unknown speaker in advance, it has the advantage of being easy to use, but the recognition performance after adaptation does not reach the latter. Therefore, currently, supervised adaptation in which the vocabulary to be spoken is designated in advance is the mainstream.

[0005]

In the voice recognition system, if a standard pattern is prepared for each word, an unknown word cannot be recognized. In the case of the specific speaker system, the user may utter the unknown word each time it appears, but in the case of the unspecified speaker system, it is true that many speakers utter the unknown word. It's impossible. Therefore, in many unspecified speaker systems, units such as phonemes and syllables smaller than words (hereinafter, subwords) are standard pattern units (hereinafter, recognition units). When utterances of words and sentences are recognized, they are connected to create a standard pattern of words and sentences. Standard patterns can be prepared for unknown words.

In the case of performing supervised speaker adaptation in such an unspecified speaker voice recognition system using subwords as a recognition unit, there are the following problems. The acoustic feature quantity corresponding to the recognition unit depends on the context, that is,
It depends on what recognition unit follows before and after the recognition unit. If the training data contains many vocabulary utterances, the data contains utterances in various contexts, and the standard patterns learned using the data are almost independent of the context. ing.

However, in the case of adaptation, the amount of learning data needs to be extremely small compared to that used for ordinary learning in order to reduce the burden of vocalization on the user. If there is a small amount of training data, naturally, the types of contexts contained in it are limited, and the standard patterns learned from that data depend on the context included in the vocabulary of the training data (hereinafter, learning vocabulary). become. Such a standard pattern has poor recognition performance for utterances in a context that did not appear in the learning vocabulary.

According to the present invention, a mapping between a pattern that depends on the context of the learning vocabulary and a pattern that does not depend on the context of the learning vocabulary is created in advance using the utterance of the reference speaker, and a new By converting the pattern that depends on the context of the learning vocabulary created from the utterance of the learning vocabulary of the user into one that does not depend on the learning vocabulary using the mapping, and using the converted pattern as the standard pattern, The purpose is to improve recognition performance.

[0009]

A speech recognition apparatus according to the present invention is a speech recognition apparatus that uses a standard pattern created from a user's utterance, and inputs and references the utterances of a large number of vocabularies of a reference speaker. Reference speaker learning vocabulary-independent pattern output part for reference speaker learning vocabulary-independent pattern creation part and reference speaker learning vocabulary-dependent pattern input for reference speaker learning vocabulary-dependent pattern input A vocabulary-dependent pattern creating unit, which inputs the reference speaker learning vocabulary independent pattern and the reference speaker learning vocabulary dependent pattern to convert the reference speaker learning vocabulary dependent pattern to the reference speaker learning vocabulary independent pattern A conversion map creating unit for outputting a conversion map; a new user learning vocabulary dependent pattern creating unit for inputting a utterance of a new user's learning vocabulary and outputting a new user learning vocabulary dependent pattern; The user learning vocabulary-dependent pattern is converted by the conversion map and outputs a standard pattern, and a recognition unit that inputs the utterance of a new user and performs recognition using the standard pattern. It is characterized by

[0010]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of a standard pattern creating apparatus according to the present invention.

Reference speaker learning vocabulary independent pattern creating unit 1
02 inputs the vocalization data of many vocabulary of the reference speaker,
The reference speaker learning vocabulary independent pattern RI is output.

Reference speaker learning vocabulary-dependent pattern generator 1
01 inputs the vocalization data of the learning vocabulary of the reference speaker,
The reference speaker learning vocabulary dependent pattern RD is output.

The transformation mapping creating unit 103 inputs the reference speaker learning term vocabulary independent pattern RI and the reference speaker learning term vocabulary dependent pattern RD, and transforms the learning term vocabulary dependent pattern into a learning term vocabulary independent pattern M. Is output.

New User Learning Vocabulary Dependent Pattern Creation Unit 1
04 inputs the vocalization data of the new user vocabulary, and outputs the new user learning vocabulary dependence pattern PD of the unknown speaker.

The pattern conversion unit 105 inputs the learning vocabulary dependent pattern PD of the new user, converts it by the conversion map M, and outputs the converted pattern PI as a standard pattern.

The recognition unit 106 inputs the utterance at the time of recognition of the new user, performs recognition using the standard pattern PI, and outputs the recognition result.

The operation of the speech recognition apparatus according to the present invention will be described in detail below.

The utterance of the speaker input to the voice recognition device is
Through a process such as AD conversion and voice analysis, it is converted into a time series of feature vectors for each unit called a frame having a certain time length. Frame length is typically 10ms to 10
It is about 0 ms. The feature vector is a feature amount extracted from the voice spectrum at that time, and is usually 10 to 100 dimensions. The time series of this feature vector is
Here, it is called vocal data.

An HMM will be described below as an example of the recognition method. HMM is one of the models of voice information source. The HMM is prepared for each recognition unit. here,
A phoneme is taken as an example of the recognition unit. When recognizing a word or sentence, the HMMs of each phoneme are concatenated and the H
Create an MM. The HMM for each phoneme is usually 1
To 10 states and state transitions between them.
Normally, the start state and the end state are defined, and a symbol is output from each state and a state transition is performed every unit time. The speech of each phoneme is represented as a time series of symbols output from the HMM during the state transition from the start state to the end state. The appearance rate of symbols is defined for each state, and the transition probability is defined for each transition between states. A symbol is generated according to the appearance probability for each state, and transitions between states according to the transition probability. By setting the probability of the initial state to a certain value and multiplying the appearance probability and the transition probability for each state transition, the probability that utterance occurs from the model can be obtained. On the contrary, when a utterance is observed, it is possible to calculate the occurrence probability assuming that it is generated from a certain HMM. In speech recognition by HMM, H is applied to each recognition candidate.
When an MM is prepared and a utterance is input, the occurrence probability is calculated for each HMM, the maximum HMM is determined as the generation source, and the recognition candidate corresponding to that HMM is used as the recognition result.

A significant feature of the HMM is that there is an algorithm called Baum-Welch algorithm that learns parameters such as transition probability and appearance probability by giving a voice corresponding to the model. The Baum-Welch algorithm is detailed in Reference 1. now,
Assuming that the symbols have a continuous distribution, the appearance probability is represented by a continuous probability density distribution function. When a Gaussian mixture distribution is used as the continuous density distribution function, its parameters are the mean vector and variance vector of each distribution, and the weighting coefficient that determines the weight between each distribution. The distribution number of the mixture distribution in each state is usually about 1 to 10. The mean vector and variance vector of each distribution have the same dimensions as the vocal data. The weighting factor of each distribution is a scalar. The parameters that can be learned are the mean and variance vectors of these continuous mixed distributions, the weighting factors, and the transition probabilities.
In the following example, the case of learning the average vector of each distribution is taken as an example. In this case, the standard pattern is an average vector of each distribution of each state for each phoneme.

Now, take one of the phonemes and consider the mean vector μ of one distribution of that one state.

First, the average vector μ _RD depending on the context of the learning vocabulary is learned by using the vocalization data of the learning vocabulary of the reference speaker prepared in advance. Reference speaker is 1
There may be one or more people. In the case of multiple speakers, the average vector may be learned for each speaker, or 1 for all reference speakers.
One may learn one mean vector. Here, the latter case will be described. As the learning method, the above-mentioned Baumu-Welch algorithm can be used. It is also possible to make the utterances correspond to each distribution by the Viterbi algorithm and average all utterance data corresponding to each distribution as an average vector. The Viterbi algorithm is detailed in Reference 1. The above corresponds to the reference speaker learning vocabulary dependence pattern creation unit 101.

Next, the average vector μ _RI independent of the context of the vocabulary for learning is learned using a large number of utterances of a large number of reference speakers prepared in advance. The learning method is the same as above. This stage corresponds to normal learning in an independent speaker system. The above corresponds to the reference speaker learning vocabulary independent pattern creating unit 102.

Next, using the average vectors of the two types of reference speakers created above, a mapping from the average vector dependent on the learning vocabulary to the average vector independent of the learning vocabulary is created. For example, the following mapping is used.

[0025]

[Equation 1]

Where μ is an average vector depending on the input learning vocabulary,

[0027]

[Equation 2]

Is a mean vector independent of the output learning vocabulary. Various other maps are possible. The mapping created here is called a conversion mapping. The above is
It corresponds to the conversion mapping creation unit 103.

The processing up to this point can be performed before a new user uses it. The following is the process when a new user uses it.

First, by using the vocabulary of the learning vocabulary uttered by the new user, an average vector μ _PD which is adapted to the utterance of the new user and which depends on the context of the vocabulary for learning is calculated. learn. The learning method is similar to that of the reference speaker. The above corresponds to the new user learning vocabulary dependence pattern creating unit 104.

Next, the average vector depending on the learning vocabulary of the new user is transformed by using a transformation map, and the average vector independent of the context of the learning vocabulary is estimated.
The above corresponds to the pattern conversion unit 105.

The above procedure is performed for each distribution of each state of each phoneme.

After creating the standard pattern in this way,
When the voice recognition device is used, the input new user's utterance is recognized using the HMM having the estimated average vector. The HMM recognition method is described in detail in Reference 1. The above corresponds to the recognition unit 106.

Here, there are a plurality of reference speakers, and all the speakers are 1
Although an example of learning one average vector μ _RD , μ _RI has been described, the average vector may be learned for each speaker. In that case, the second term of the equation (2) may be obtained by summing all reference speakers by appropriately weighting the differences of μ _RI and μ _RD among the plurality of speakers.

Further, here, the mapping is created by using only one distribution in the mixture distribution, and the average vector is converted, but it is also possible to create the mapping by simultaneously using other distributions. . As other distributions, other distributions of the same state,
Alternatively, the distribution of other states, the distribution of other recognition units, and the like are possible. In such a case, a plurality of average vectors corresponding to the distribution are converted using the created mapping.

This time, a non-linear mapping is taken as an example, but a linear one or a non-linear one can be used.

The number of reference speakers is variable. It is possible for one person.

For simplification, the reference speaker learning vocabulary dependence pattern, the reference speaker learning vocabulary independent pattern, and the new user learning vocabulary dependence pattern are all set to the same learning method. Even if they are different, the present invention can be applied.

Although an example of learning only the average vector is shown here, other variances, weights, transition probabilities, etc. can be easily learned by the same method. In addition, it is possible to learn a plurality of parameters at the same time. Here, the phoneme is taken as an example of the recognition unit, but the present formula can be easily applied to other recognition units such as a syllable and a semisyllabic other than the phoneme.

Although the HMM has been described as an example of the recognition method, the present method should be applied to other recognition methods such as NN and DP matching when learning parameters. Is easily possible.

[0041]

EFFECTS OF THE INVENTION A standard pattern created using a utterance of a limited number of learning terms depends on the context of the learning terms. Correcting the effects of context using maps created from the voicing of multiple vocabularies of other speakers makes it possible to estimate standard patterns independent of context. A standard pattern with high recognition performance can be created with a smaller amount of training utterances than before.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of a voice recognition device according to the present invention.

[Explanation of symbols]

 101 Reference Speaker Learning Vocabulary Dependent Pattern Creating Unit 102 Reference Speaker Learning Vocabulary Independent Pattern Creating Unit 103 Transformation Mapping Creating Unit 104 New User Learning Vocabulary Dependent Pattern Creating Unit 105 Pattern Converting Unit 106 Recognition Unit

Claims (1)

[Claims] 1. A speech recognition apparatus using a standard pattern created from a user's utterance, wherein a reference speaker learning vocabulary independent pattern is output by inputting utterances of many vocabularies of a reference speaker. A learning vocabulary independent pattern creating unit, a reference speaker learning vocabulary dependent pattern creating unit that inputs the utterance of the learning vocabulary of the reference speaker and outputs a reference speaker learning vocabulary dependent pattern, and the reference speaker learning pattern A conversion mapping generation unit that inputs a vocabulary independent pattern and the reference speaker learning vocabulary dependent pattern, and outputs a conversion map from the reference speaker learning vocabulary dependent pattern to the reference speaker learning vocabulary independent pattern; A new user learning vocabulary dependence pattern creating unit for inputting the utterance of the learning vocabulary of the user and outputting a new user learning vocabulary dependence pattern; and a conversion copy of the new user learning vocabulary dependence pattern. Was converted by the pattern conversion section for outputting a reference pattern, enter the utterance of the new user, the speech recognition device comprising a recognition unit for recognizing, in that they are composed of using the standard pattern.