JP2976795B2 - Speaker adaptation method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speaker adaptation system for quickly adapting a recognition device to a user in speech recognition, and more particularly to a teacher in a speech recognition system using a mixed continuous distribution model type HMM. It is related to a speaker adaptation method without.

[0002]

2. Description of the Related Art In recent years, research on recognition of voice patterns by machines has been conducted, and various techniques have been proposed. Among these, typical recognition methods often used include a method called DP matching (dynamic programming matching) and a method called a hidden Markov model (HMM). As a speech recognition system using a speech recognition technique such as DP matching or HMM, an unspecified speaker recognition system for the purpose of recognizing anyone's voice has been actively researched and developed.

An unspecified speaker system has an advantage that, unlike a specific speaker system in which a user is specified, the user does not need to register an utterance in advance. However, in recent years,
The following problems were pointed out. First, the recognition performance of most speakers is inferior to that of the specific speaker system, and there are speakers (singular speakers) whose recognition performance is significantly poor. In order to solve such problems,
In recent years, studies have begun to apply speaker adaptation techniques, which have been used in specific speaker systems, to unspecified speaker systems.

[0004] Speaker adaptation refers to a method of adapting a recognition system to a new user (an unknown speaker) using a smaller amount of adaptation data than used for learning. For details of the speaker adaptation method, see "Speaker Adaptation Technology in Speech Recognition", Sadahiro Furui, Television Society of Japan, Vol.
o. 9, 1989, pp. 929-934.

[0005] Speaker adaptation can be broadly divided into two approaches. One is supervised speaker adaptation and the other is unsupervised speaker adaptation. Here, the teacher refers to a phoneme notation sequence representing the utterance content of the input utterance. What is supervised adaptation?
This is an adaptation method in a case where a phoneme notation sequence for an input utterance is known. In the adaptation, it is necessary to instruct an unknown speaker in advance on a vocabulary vocabulary.

On the other hand, unsupervised adaptation is an adaptation method in the case where the phoneme notation sequence for an input utterance is unknown, and does not limit the utterance content of the input utterance for an unknown speaker, that is, for an unknown speaker. There is no need to give instructions on what to say,
It is an easy-to-use method for the user because the adaptation can be performed by using the input speech that is actually using the speech recognition without being aware of the unknown speaker.

In general, unsupervised adaptation has lower recognition performance after adaptation than supervised adaptation, and thus, supervised adaptation is often used at present.

A conventional speech recognition apparatus using the supervised adaptation technique will be described below with reference to FIG.

The utterance of the speaker input to the speech recognition device 6-1 is input to an input pattern creating unit 6-2, where the utterance is subjected to AD conversion,
Through a process such as voice analysis, the data is converted into a time series of feature vectors for each unit called a frame having a certain time length. The time series of the feature vectors is referred to as an input pattern here. The frame length is usually about 10ms to 100ms. The feature vector is obtained by extracting the feature amount of the voice spectrum at that time, and is usually 10 to 100 dimensions.

On the other hand, HM is stored in the standard pattern storage section 6-6.
M (Hidden Markov Model) is stored. HMM
Is a model of a speech information source, and its parameters can be learned using speech of a speaker. The HMM will be described in detail in the description of the recognition unit 6-3.

An HMM is usually prepared for each recognition unit. Here, a phoneme is taken as an example of the recognition unit. As the HMM in the standard pattern storage unit 6-6, a different speaker HMM learned using utterances of different users, or an unspecified speaker HMM previously learned using utterances of many speakers is used. .

Now, when 1000 words are to be recognized, that is,
It is assumed that a correct answer of one word is obtained from recognition candidates of 1000 words. When recognizing a word, the HMM of each phoneme is connected to create an HMM of a recognition candidate word. In the case of 1000 word recognition, a word HMM for 1000 words is created. In this description, 1000 words are taken as an example, but it is not particularly necessary to be 1000 words, and any number of words may be used. In addition, continuous syllables may be used as the recognition target, for example. A continuous syllable is an HMM that uses a network to represent the concatenation of all syllables that appear in Japanese (or that language in other foreign language recognition), and recognizes virtually all utterances that appear in Japanese. It is possible to These processes are performed by the vocabulary pattern creating unit 6-5.

The recognition unit 6-3 includes a vocabulary pattern creation unit 6
The input pattern is recognized using the word HMM created in -5. HMM is a model of speech information source,
In order to cope with various fluctuations of the voice pattern, a statistical idea is introduced in the description of the standard pattern. For a detailed description of HMM, see "Speech Recognition by Stochastic Model", Seichi Nakagawa, edited by IEICE (Showa 63)
Pages 46, 55-60, 69-74. The HMM of each phoneme is usually composed of 1 to 10 states and state transitions between them. Usually, a start state and an end state are defined, and a symbol is output from each state for each unit time, and a state transition is performed. The speech of each phoneme is represented as a time series of symbols output from the HMM during a state transition from a start state to an end state. A symbol appearance probability is defined for each state, and a transition probability is defined for each transition between the states. The transition probability parameter is a parameter for expressing the temporal fluctuation of the voice pattern. The appearance probability parameter expresses the fluctuation of the timbre of the voice pattern. By setting the probability of the start state to a certain value and multiplying the appearance probability and the transition probability for each state transition, the probability that the utterance occurs from the model can be obtained.
Conversely, when an utterance is observed, assuming that it is generated from a certain HMM, its occurrence probability can be calculated.

In speech recognition by HMM, an HMM is prepared for each recognition candidate, and when an utterance is input, each HMM is input.
In the MM, the occurrence probability is obtained, the HMM having the maximum value is determined as the generation source, and a recognition candidate corresponding to the HMM is used as a recognition result.

The output probability parameters include a discrete probability distribution expression and a continuous probability distribution expression. Here, a continuous probability expression will be used as an example. In the continuous probability distribution expression, a mixed continuous distribution, that is, a distribution obtained by adding a plurality of Gaussian distributions with weight is used. Output probability parameter, transition probability parameter,
Parameters such as weights of a plurality of Gaussian distributions are learned in advance by applying a learning voice corresponding to the model and using an algorithm called a Baum-Welch algorithm.
The Baum-Welch algorithm is described in detail in Reference 1. In the following example, the output probability is a mixed continuous probability distribution.

The processing at the time of word recognition will be described below using mathematical expressions. Wherein when X = x ₁ input pattern X expressed as a sequence of _{vectors, x 2, x 3, ......} x t, and ...... x _T (1). Here, T is the total number of frames of the input pattern. Now, a recognition candidate word W _1, W _2, and the W _N.
N is the number of recognition candidate words. Word HM for each word W _N
Matching between M and input pattern X is performed as follows. In the following, the suffix n is omitted unless necessary.

First, in the word HMM, the transition probability from the state j to the state i is a _ji , and the mixing weight of the output probability distribution is λ
_im , an average vector of each element Gaussian distribution (called a frame distribution) is μ _im , and a covariance matrix is Σ _im . Here, t represents the input time, i, j represents the state of the HMM, and m represents the mixed element number. The following recurrence formula calculation for the forward probability α (i, t) is performed.

[0018]

Where

[0020]

[0021]

The likelihood of an input pattern for a word W _n is

[0023]

[0024] Here, I is the final state. This processing is performed for each word model, and a recognition result word for the input pattern X is obtained.

[0025]

Is

[0027]

Given by The recognition result word is sent to the recognition result output unit.

The recognition result output unit 6-4 performs processing such as outputting the recognition result on a screen or transmitting a control command corresponding to the recognition result to another device.

The speech recognition apparatus has been described using the HMM as an example.

Next, a supervised speaker adaptation technique for the speech recognition apparatus will be described. In the supervised speaker adaptation, a word or sentence to be uttered is instructed to a user in advance, and the parameters of the HMM are updated using the word notation and the input speech. This is called supervised adaptation in the sense that the correct word for the utterance is known in advance.
As a supervised speaker adaptation method, Japanese Patent Application No. 2-203437 "Standard pattern adaptation method" or Japanese Patent Application No. 4-203669
Although there is a method described in "Speech Recognition Apparatus", here, a method based on Japanese Patent Application No. 2-203437 will be briefly described.

In supervised speaker adaptation, it is necessary to minimize the amount of input utterance necessary for adaptation in order to reduce the burden on the speaker. However, the HMM generally has a large number of parameters, and when trying to adapt all parameters with a small amount of adaptation utterance, the accuracy of parameter estimation is poor due to lack of data, and recognition performance may not be improved. Therefore, in the supervised speaker adaptation here, H
Only the average vector of the output probability distribution among the parameters of the MM is adapted. The average vector was selected because it is considered that this has the greatest effect on the recognition performance among the parameters of the HMM.

For the sake of simplicity, a case where the output probability distribution is a single Gaussian distribution will be described, and a case where the output probability distribution is a Gaussian mixture distribution will be described later.

The supervised adaptation is divided into the following two stages. First, the first stage will be described.

First, an adaptation initial HMM is prepared in the adaptation initial standard pattern storage unit 6-11 in advance. Initialization H
As the MM, for example, a phoneme HMM of an unspecified speaker prepared in advance by the utterance of many speakers is used, and may be the same as the phoneme HMM stored in the standard pattern storage unit 6-6, or may be different. Good. Also, each phoneme HM is added to the adaptation unit 6-9.
A buffer B1 (i) having a dimension of a feature vector for each state of M, and 1 for adding the number of feature vectors
A dimensional buffer B2 (i) is prepared. Then, the following processing is performed for each input utterance.

First, similarly to the recognition, an input pattern is created from an input voice in an input pattern creating section 6-8. As described above, in the case of supervised adaptation, since the correct word is known in advance, the adaptation dictionary creating unit 6-7 determines the adaptation dictionary from the input correct word notation and the created input pattern. Create Next, vocabulary pattern creating section 6-10
Is the phoneme sequence of the adaptation dictionary and the initialized HMM for each phoneme.
To create a word HMM corresponding to the input pattern. Then, the adaptation unit 6-9 performs likelihood calculation between the input pattern and the adaptation word HMM. Here, the expression
Instead of (2) and (5),

[0037]

[0038]

The recurrence formula is used. This is called the Viterbi algorithm. In parallel with equation (7),

[0040]

Is calculated, and the state at the previous time in each state at each time is stored in the array Ψ. After the calculation of the last frame T in Expression (9) is completed, the state corresponding to each frame is obtained in order from the last frame to the first frame using Ψ. That is, when the state corresponding to the frame t is represented as S (t),

[0042]

[0043]

Is as follows. This process is called backtracking. By this processing, a state corresponding to the feature vector at each time is obtained. Then for each feature vector x _t at each time,

[0045]

[0046]

Is performed, and the B1 and B2 buffers are added. This process is repeated for the number of utterance words for adaptation.

After the above association processing is completed for all input utterances, the feature vectors of the frames associated with each state i of each phoneme HMM are averaged over all input patterns, and the state adaptation is performed. The mean vector after

[0049]

Then,

[0051]

Is calculated.

In the second stage, HMMs corresponding to phonemes not included in the adaptation utterance are adapted using a technique called spectral interpolation. In the spectral interpolation, the average vector corresponding to the phoneme that did not appear during the adaptation utterance was calculated as
The estimation is performed using the difference between the average vector of the appearing phonemes before and after the adaptation.

Assume that a set of average vectors of each state of the HMM included in the adaptation utterance is set A, and a set of average vectors of each state of the HMM not included is set B. First, the adaptation vector delta ^A is calculated for all the states of the set A. The adaptation vector is the average vector τ after the adaptation.
And ^A, is defined as the difference between the average vector mu ^A prior adaptation. Next, in order to adapt the adaptation vector of the state of the set B, it is obtained by interpolating the adaptation vector of the state of the set A. The algorithm is as follows. 1. In the state j of the set A, the average vector τ _j ^A after the adaptation has already been obtained. The adaptation vector Δ _j ^A is given by the following equation.

[0055]

Here, A is a subscript indicating that the state j belongs to the set A. The adaptation vector Δ _j ^A is calculated for all states in set A. 2. For state i of set B, the adaptation vector Δ _i ^B is
It is obtained by interpolating the adaptation vector of the state j of the set A.

[0057]

[0058] weights w _i ^j to adaptation vector delta _j ^A is defined as a function of the distance d _i ^j with mu _i ^B and mu _j ^A. For example, w
_i ^j is defined as follows.

[0059]

[0060] where m is a constant representing the dependence on the distance d _i ^j weights w _i ^j. The adaptation vector Δ _i ^B is calculated for all the states belonging to the set B. 3. The average vector τ _i ^B of the new speaker state i is given by the following equation.

[0061]

[0062] Here, mu _I ^B is the mean vector adaptation initial HMM. 4. Repeat steps 2-3 for all states in set B. The above procedure can be applied to an HMM in which the output probability distribution is a Gaussian mixture distribution by separately treating a plurality of component distributions in the state. In the first stage of backtracking, among the component distributions in the state, those having the largest value obtained by multiplying the appearance probability of the corresponding feature vector by the weighting coefficient are selected and classified into the set A. Component distributions without corresponding adaptation data are classified into set B. The second-stage spectral interpolation is performed on the component distribution of the set B. That is, the adaptation vector of the component distribution of the set B is obtained by spectrum interpolation using the adaptation vector of the component distribution of the set A over all states.

Here, an example in which only the average vector is adapted has been described, but other variances, weights, transition probabilities, and the like can be easily adapted in the same manner. Also,
It is possible to adapt a plurality of these parameters at the same time.

The HMM after the adaptation is stored in place of the previous HMM in the standard pattern storage unit 6-6. In this case, the HMM may be stored for each speaker, but in that case, as a pre-processing of the recognition, the user selects the HMM or automatically selects the HMM using the utterance of the user. You need a way to do that.

Up to this point, the HMM using phonemes as recognition units
Has been described, but when words or sentences are used as a recognition unit, the adaptation can be easily performed by creating an HMM corresponding to each word or sentence. When the recognition unit and the input utterance unit are the same (for example, a word-based HMM and a word utterance), it is not necessary to connect the adaptation initial HMM, and the adaptation can be performed by performing likelihood calculation as it is.

The conventional supervised adaptation has been briefly described above.

[0067]

The above-described conventional supervised adaptation method has higher performance than the unsupervised adaptation method.
However, the user has to utter the words instructed to the device as training separately from the utterance at the time of use, and there is a disadvantage that the burden is large.

It is an object of the present invention to provide a method of unsupervised adaptation in a speaker adaptation of a speech recognition system that has a performance comparable to supervised adaptation without the user being aware of it. .

[0069]

According to a first aspect of the present invention, there is provided a speaker adaptation system for adapting a standard pattern used for speech recognition by using an input speech pattern. A standard pattern storage unit that holds a word standard pattern created based on a reference; an input pattern creation unit that creates an input pattern by performing voice analysis on an input voice; and stores the created input pattern in a word in the standard pattern storage unit. A recognition unit for recognizing a word by using a standard pattern, a recognition result output unit for outputting a recognition result in the recognition unit, and an initial word standard pattern adapted by a predetermined method with reference to the notation of the recognition word And an adaptation standard pattern selecting unit that prepares the likelihood calculation of the input pattern based on the adaptation initial word standard pattern. And supervised adaptation unit by adaptation after word reference pattern obtained by the time the calculation result to update the word reference pattern in the reference pattern memory,
An input pattern storage unit for storing the input pattern;
Input the input pattern in the input pattern storage unit
The recognition unit, the recognition result output unit, and the adaptation
Operation of the standard pattern selection unit and the supervised adaptation unit
A predetermined variable reaches a predetermined reference value
It has a repetition control unit to repeat until
I do.

According to a second aspect of the present invention, in a speaker adaptation method for adapting a standard pattern used for speech recognition using an input speech pattern, the speaker adaptation method is prepared based on a predetermined standard in units of sub-words such as syllables and phonemes. A standard pattern storage unit that holds the obtained subword standard pattern, a vocabulary pattern creating unit that creates a word standard pattern corresponding to the recognition candidate word using the subword standard pattern, An input pattern creation unit to be created, a recognition unit that recognizes the created input pattern using the word standard pattern created by the vocabulary pattern creation unit, and a recognition that outputs a recognition result in the recognition unit A result output unit and an adaptive initial word standard parameter by a method predetermined with reference to the notation of the recognized word. An adaptation standard pattern selecting unit for preparing a standard pattern, and a likelihood calculation of the input pattern based on the adaptation initial word standard pattern, and an adaptation subword standard pattern obtained from the likelihood calculation result. A supervised adaptation unit for updating the subword standard pattern in a pattern storage unit, and the input pattern
An input pattern storage unit for storing the input pattern;
The recognition is performed using the input pattern in the storage unit as an input.
Unit, the recognition result output unit, and the standard pattern selection for adaptation.
The operation of the selection unit and the supervised adaptation unit is determined in advance.
Until the variable reaches a predetermined reference value.
And a repetition control unit.

A third aspect of the present invention relates to a standard pattern used for speech recognition.
Speaker adaptation using the input speech pattern
Sub-words such as syllables, phonemes, etc.
Subword mark created according to predetermined criteria
A standard pattern storage unit for storing a reference pattern;
Words corresponding to recognition candidate words using word standard patterns
Vocabulary pattern creation unit for creating word standard patterns, and input
Performs voice analysis on voice and creates input patterns.
A force pattern creation unit and the created input pattern.
An input pattern storage unit for remembering the input pattern storage unit
In the input pattern and the vocabulary pattern creation unit in
Word recognition using the word standard pattern created above
And a recognition unit that outputs the recognition result.
A recognition result output unit that performs
Based on the word standard pattern in the vocabulary pattern creation unit
Performing the likelihood calculation of the input pattern and the likelihood calculation result
Of the subword standard pattern after adaptation
The sub-word standard in the standard pattern storage unit.
A supervised adaptation unit for updating the pattern and the recognition unit
The operation of the recognition result output unit and the supervised adaptation unit
Until a predetermined variable reaches a predetermined reference value.
Characterized by having a repetition control unit for repeating
You.

A fourth aspect of the present invention relates to a standard pattern used for speech recognition.
Speaker adaptation using the input speech pattern
Sub-words such as syllables, phonemes, etc.
Subword mark created according to predetermined criteria
A standard pattern storage unit for storing a reference pattern;
Words corresponding to recognition candidate words using word standard patterns
A vocabulary pattern creation unit that creates a word standard pattern,
The recognition candidate words created by the specified method
Basic standard pattern notation that stores the corresponding basic standard pattern
Memory unit and input pattern by performing voice analysis on input voice
An input pattern creation unit for creating the input, and the created input
An input pattern storage unit for storing a pattern;
The input pattern and the vocabulary pattern in the turn storage unit.
The word standard pattern created by the
And a recognition unit for recognizing words.
A recognition result output unit for outputting a recognition result;
The basic mark in the corresponding basic standard pattern storage unit;
Perform likelihood calculation of the input pattern based on the quasi-pattern
Adapted subword obtained from the likelihood calculation result
By the standard pattern in the standard pattern storage unit
Supervised adaptation unit that updates the subword standard pattern
With the recognition unit, the recognition result output unit, and the teacher
Variables for which the operation of the adaptation unit is predetermined are predetermined.
Equipped with a repetition control unit that repeats until the reference value is reached
It is characterized by having.

[0073]

[0074]

Next, the present invention will be described with reference to the drawings.

The description is based on the assumption that standard pattern storage units 101 to 501, input pattern creation units 102 to 502, recognition units 103 to 503, recognition result output units 104 to 504, and vocabulary pattern creation shown in FIGS. Part 207,
Reference numerals 309, 409, and 510 denote a standard pattern storage unit 6-6, an input pattern creation unit 6-2, a recognition unit 6-3, and a recognition result output unit 6-4 described in the description of the related art, respectively.
Since this is the same as the vocabulary pattern creating unit 6-5, in the present embodiment, only a brief description will be given .

Also, in the supervised speaker adaptation here,
As in the case described in the related art, it is assumed that only the average vector of the output probability distribution, which has the greatest effect on the recognition performance among the parameters of the HMM, is adapted.

FIG . 1 is a block diagram of an embodiment of the first speaker learning method . The input utterance and the recognition unit of the HMM are words. The standard pattern storage unit 101 holds a word HMM of each recognition candidate word. The word HMM is an HMM of an unspecified speaker or an HMM of another speaker. The input pattern creation unit 102 performs speech analysis on the input speech to create an input pattern X. The created input pattern X is stored in the recognition unit 103 in the standard pattern storage unit 10.
1 is recognized using the word HMM, and the recognition result output unit 104 outputs the recognition result.

The adaptation standard pattern selection unit 105 prepares an adaptation initial word HMM by referring to the notation of the recognition result word. The adapted initial word HMM is a word HMM of an unspecified speaker learned in advance by the utterances of many speakers or a word HMM of a different speaker learned by utterances of other speakers. The word HMM of 101 may be used, or another word HMM may be used.

The supervised adaptation unit 106 calculates likelihood using the input pattern X and the adapted initial word HMM,
After performing one or more input patterns,
The average vector after the adaptation is calculated, and the HMM after the adaptation is obtained. For the detailed operation of the supervised adaptation unit, refer to the adaptation unit 6-9 in the description of the related art. The post-adaptation HMM output from the supervised adaptation unit 106 is
It is output to the standard pattern storage unit 101 and the recognition H
It is stored instead of MM.

FIG . 2 is a block diagram of an embodiment of the second speaker learning method . It is assumed that the input utterance is a word. The standard pattern storage unit 201 holds the HMM of each phoneme. The vocabulary pattern creation unit 207 creates a word HMM corresponding to the recognition candidate word using the HMM of each phoneme. The input pattern creation unit 202 performs an audio analysis on the input speech to create an input pattern X. The created input pattern is recognized by the recognition unit 203 using the word HMM of the recognition candidate word, and the recognition result output unit 204 outputs the recognition result. The adaptation dictionary creating unit 205 creates an adaptation dictionary from the recognition result notation.

The supervised adaptation unit 206 first creates an adapted initial word HMM by connecting the adapted initial phonemes HMM using the adaptation dictionary. Adapted initial phoneme HMM
May be a phoneme HMM in the standard pattern storage unit 201 or another phoneme HMM. Next, the likelihood calculation is performed using the created adaptive initial word HMM and the input pattern,
After performing one or more input patterns,
The average vector after the adaptation is calculated, and the HMM after the adaptation is obtained. The adapted HMM is stored in the standard pattern storage unit 201.
And stored in place of the conventional recognition HMM.

FIG. 3 shows a third embodiment according to the second aspect of the present invention .
FIG. 2 is a block diagram of an embodiment of a speaker learning method according to the present invention. In FIG.
The difference from the second embodiment shown in FIG.
, The adaptation is repeatedly performed. The HMM after the adaptation created by the adaptation of the supervised adaptation unit 306 is compared with the recognition HMM used for the recognition.
It generally exhibits higher recognition performance for user utterances. Therefore, if the input pattern is recognized once more using the HMM after the adaptation, it is considered that a better recognition rate will be exhibited. Then, if the adaptation is performed using the adaptation dictionary created using the recognition result, there is a possibility that an adapted HMM having higher recognition performance may be created. At the time of repetition of recognition / adaptation, the input pattern is stored in the input pattern storage unit 307 at the time of the first adaptation in order to omit the calculation for creating the input pattern at each iteration, and the second and subsequent times are used. In the repetition, an input pattern output from the input pattern storage unit 307 is used. The number of repetitions is determined in advance, or a recognition result likelihood corresponding to the recognition result word in the recognition unit 303 is stored for each repetition, and the likelihood is saturated as compared with the likelihood in the previous repetition. Judgment is made, and if saturated, the repetition is stopped. This repetition means 308 can be similarly applied to the first embodiment shown in FIG. 1, and the same applies whether the phoneme is a recognition unit or the same recognition unit as an input utterance such as a word or a sentence. Applicable to

FIG. 4 shows a fourth embodiment according to the third aspect of the present invention .
FIG. 2 is a block diagram of an embodiment of a speaker learning method according to the present invention. In the fourth speaker learning method, the phoneme HMM stored in the standard pattern storage unit is used as the adapted initial phoneme HMM in the third speaker learning method. By performing the repetition, the phoneme HMM stored in the standard pattern storage unit has already been adapted to the user to some extent. Therefore, by using it as an initial model of adaptation, the adaptation is performed quickly. That is, there is an effect that the number of times of recognizing and adapting is reduced. This method can be easily applied even if the recognition unit is a word.

FIG. 5 shows a fifth embodiment according to the fourth aspect of the present invention .
FIG. 2 is a block diagram of an embodiment of a speaker learning method according to the present invention. In the fifth speaker learning method, in the third speaker learning method, a phoneme HMM stored in the basic standard pattern storage unit 509 is used as an adaptive initial HMM. The basic standard pattern is
Unspecified speaker HM learned in advance by uttering many speakers
M or another speaker HMM learned by uttering another user
Is used. This basic standard pattern is not updated by repetition. In the fourth speaker adaptation method, the HMM after the adaptation in the previous iteration is used as the adaptation initial HMM. However, although the adaptation is performed quickly, if there is a misrecognition during the recognition, the adaptation is made. However, there is a problem that the influence on the performance of the conversion becomes greater. However, in the fifth speaker adaptation method, the information obtained from the previous loop in the adaptation is only the adaptation dictionary serving as a teacher, and the number of repetitions is larger than that of the fourth speaker adaptation method. It is considered that the degree of misrecognition is low. This method can be easily applied even if the recognition unit is a word.

Each of the embodiments described above has a feature that it can be achieved by extending and changing the conventional device shown in FIG.

Here, a word has been described as an example of an object to be recognized, but a sentence or a conversation utterance can be adapted by the same means. Also, HMM is used as the recognition method.
Has been described as an example, but other recognition methods, such as NN,
In a recognition method such as DP matching, the recognition / adaptation unit can be configured using a similar method. Also, a method based on the method of Japanese Patent Application No. 2-203437 has been described as the adaptation means, but it is also possible to use another supervised adaptation method. Furthermore, in the recognition / adaptation means, a phoneme was taken as an example of a recognition unit.
This method can be easily applied to other recognition units such as syllables.

The results of an evaluation experiment of the above-described second speaker adaptation method will be described below. In the evaluation experiment, 5000 similar words were recognized by using a Gaussian mixture HMM using a semisyllable as a recognition unit. Here, the Gaussian mixture distribution number of the HMM is 2
As data of a large number of speakers, a single 250-word utterance taking into account the phoneme balance of 46 men and 39 women in total was used. In addition, three males and four females, which are not included in the above-mentioned 85 speakers, were used as the evaluation speakers, and the vocabulary was 250 words once each for the adaptation data and the evaluation data, which were different from those at the time of learning. Utterance was used. The vocabulary of the data for adaptation and evaluation is different from each other. The analysis conditions were as follows: sampling frequency 16 kHz, band 0.1-7.2 kHz,
Mel cepstrum analysis was used at a frame interval of 10 ms. The feature vectors are normalized power difference, mel-cepstral 10-dimensional, and mel-cepstral change 10-dimensional in total 2
It is one-dimensional. The initial HMM of the adaptation used an unspecified speaker model learned using utterance data of 85 speakers.

As a result of performing an evaluation experiment of unsupervised adaptation for the case where 5000 discrete words were used as the recognition target for adaptation, the performance was greatly improved, and the unspecified speaker recognition rate was 84.5 on average for 7 speakers. %, The recognition rate after unsupervised adaptation is 91.3% and the error is reduced by almost half when the number of adapted words is 250 words. Also, compared to the supervised adaptation, the number of words for each adaptation is only 1-2% lower.

[0089]

As described above, according to the present invention, it is possible to adapt the speech recognition device to the user without being conscious of the user and to obtain a high recognition performance, and at the same time to reduce the burden on the user. This has the effect of reducing the user interface, and further, if a supervised adaptation system already exists, it is possible to construct an unsupervised adaptation system with little effort by using it.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a first embodiment of the present invention.

FIG. 2 is a block diagram showing a second embodiment of the present invention.

FIG. 3 is a block diagram showing a third embodiment of the present invention.

FIG. 4 is a block diagram showing a fourth embodiment of the present invention.

FIG. 5 is a block diagram showing a fifth embodiment of the present invention.

FIG. 6 is a block diagram showing an embodiment of the prior art.

[Explanation of symbols]

 Reference Signs List 101 Standard pattern storage unit 102 Input pattern creation unit 103 Recognition unit 104 Recognition result output unit 105 Standard pattern selection unit for adaptation 106 Supervised adaptation unit 201 Standard pattern storage unit 202 Input pattern creation unit 203 Recognition unit 204 Recognition result output unit 205 adaptation dictionary creation unit 206 supervised adaptation unit 207 vocabulary pattern creation unit 301 standard pattern storage unit 302 input pattern creation unit 303 recognition unit 304 recognition result output unit 305 adaptation dictionary creation unit 306 supervised adaptation unit 307 Input pattern storage unit 308 Repetition control unit 309 Vocabulary pattern creation unit 401 Standard pattern storage unit 402 Input pattern creation unit 403 Recognition unit 404 Recognition result output unit 405 Adaptation dictionary creation unit 406 Supervised adaptation unit 407 Input pattern storage unit 408 Iteration control unit 409 Vocabulary pattern creation unit 501 Standard pattern storage unit 502 Input pattern creation unit 503 Recognition unit 504 Recognition result output unit 505 Adaptation dictionary creation unit 506 Supervised adaptation unit 507 Input pattern storage unit 508 Repetition control unit 509 Basic standard pattern storage unit 510 Vocabulary pattern creation unit 6-1 Speech recognition device 6-2 Input pattern creation unit 6-3 Recognition unit 6-4 Recognition result output unit 6-5 Vocabulary pattern creation unit 6-6 Standard pattern storage unit 6 -7 Adaptation dictionary creation unit 6-8 Input pattern creation unit 6-9 Adaptation unit 6-10 Vocabulary pattern creation unit 6-11 Adaptation initial standard pattern storage unit

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-5-88693 (JP, A) JP-A-2-198499 (JP, A) JP-A-4-280299 (JP, A) JP-A-63-1988 63098 (JP, A) JP-A-5-232985 (JP, A) JP-A-5-224691 (JP, A) Proceedings of the Acoustical Society of Japan October 1993 2-7-13 p. 95-96 Proceedings of the Acoustical Society of Japan March 1994 3-7-8 p. 103-104 Proceedings of the Acoustical Society of Japan September 1990 1-8-12 p. 23-24 Proceedings of the Acoustical Society of Japan March 1992 1-P-15 p. 133-134 (58) Field surveyed (Int. Cl. ⁶ , DB name) G10L 3/00 521 G10L 3/00 535 JICST file (JOIS)

Claims (4)

(57) [Claims] In a speaker adaptation method for adapting a standard pattern used for speech recognition using an input speech pattern, a word standard pattern created based on a predetermined standard for each recognition candidate word is held. A standard pattern storage unit that performs voice analysis on an input voice to generate an input pattern; and recognizes a word of the generated input pattern using a word standard pattern in the standard pattern storage unit. A recognition unit, a recognition result output unit that outputs a recognition result in the recognition unit, and an adaptation standard pattern selection unit that prepares an adaptation initial word standard pattern by a predetermined method with reference to the notation of the recognition word. And calculating the likelihood of the input pattern based on the adapted initial word standard pattern, and calculating the likelihood calculated by the likelihood calculation result. A supervised adaptation unit for updating the word standard pattern in the standard pattern storage unit with the optimized word standard pattern, and storing the input pattern
An input pattern storage unit;
The recognition unit and the recognition
A result output unit and the standard pattern selecting unit for adaptation and
Variables for which the operation of the adaptation unit with teacher is predetermined
Iterative system that repeats until the specified reference value is reached
A speaker adaptation method comprising a control unit. 2. A speaker adaptation method for adapting a standard pattern used for speech recognition using an input speech pattern, wherein a subword standard created based on a predetermined criterion in units of subwords such as syllables and phonemes. A standard pattern storage unit for holding a pattern, a vocabulary pattern creating unit for creating a word standard pattern corresponding to a recognition candidate word using the subword standard pattern, and an input pattern for creating an input pattern by performing speech analysis on input speech A creating unit, a recognition unit that recognizes a word using the word standard pattern created by the vocabulary pattern creating unit with the created input pattern, and a recognition result output unit that outputs a recognition result in the recognition unit. Preparing an adapted initial word standard pattern by a predetermined method with reference to the notation of the recognized word. An adaptation standard pattern selecting unit, and a likelihood calculation of the input pattern based on the adaptation initial word standard pattern, and a post-adaptation subword standard pattern obtained from the likelihood calculation result in the standard pattern storage unit. A supervised adaptation unit for updating the subword standard pattern, and storing the input pattern
An input pattern storage unit that performs
The recognition unit and the input pattern
A recognition result output unit and the standard pattern selection unit for adaptation;
And the variables in which the operation of the supervised
Repetition until it reaches a predetermined reference value
A speaker adaptation method characterized by comprising a control unit. 3. A standard pattern used for voice recognition is input sound.
Speaker adaptation method using voice patterns
And sub-words such as syllables, phonemes, etc.
Subword standard pattern created by the specified criteria
A standard pattern storage unit for storing the subword standard
Word standard patterns corresponding to recognition candidate words using patterns
Vocabulary pattern creation unit that creates
An input pattern that performs voice analysis and creates an input pattern
A creating unit and an input for storing the created input pattern
A pattern storage unit, and the input pattern storage unit
Created by the vocabulary pattern creating unit
To recognize words using the word standard pattern
And a recognition unit that outputs a recognition result in the recognition unit.
Result output unit and the vocabulary pattern corresponding to the recognized word
The input pattern based on the word standard pattern in the creating unit;
The likelihood calculation of the turn is performed and the likelihood calculation result is obtained.
The standardized subword standard pattern after adaptation
The subword standard pattern in the pattern storage unit is
Supervised adaptation unit to be updated, the recognition unit, and the recognition result
Predetermine the operation of the output section and the supervised adaptation section.
Until the variable reaches a predetermined reference value.
Speaker adaptation characterized by having an iterative control unit
System. 4. A standard pattern used for speech recognition is input sound.
Speaker adaptation method using voice patterns
And sub-words such as syllables, phonemes, etc.
Subword standard pattern created by the specified criteria
A standard pattern storage unit for storing the subword standard
Word standard patterns corresponding to recognition candidate words using patterns
A vocabulary pattern creation unit that creates a
Basics corresponding to the recognition candidate words created by the method
A basic standard pattern storage unit for storing standard patterns
Perform voice analysis on input speech to create input patterns To
An input pattern creation unit, and the created input pattern
An input pattern storage unit for storing the input pattern;
The input pattern and the vocabulary pattern creating unit in the unit
Of the word using the word standard pattern created by
A recognition unit for performing recognition, and output a recognition result of the recognition unit.
A recognition result output unit,
The basic standard pattern in the basic standard pattern storage unit
Calculating the likelihood of the input pattern based on the likelihood meter
Subword standard pattern after adaptation obtained from the calculation result
The sub word in the standard pattern storage unit.
Supervised adaptation unit for updating standard patterns and said recognition
Unit, the recognition result output unit, and the supervised adaptation unit.
A variable whose operation has been determined reaches a predetermined reference value.
It has a repetition control unit that repeats until
Speaker adaptation method to be used.