JPH09160586A - Learning method for hidden markov model - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method of learning environment-dependent phonemes HMM(hidden Markov model) which is smoothed and reduces the bias of learned data, while maintaining the advantages of the environment-dependent phonemes HMM. SOLUTION: A word HMM is learned in a step 6, separated into environment- dependent phonemes HMM in a step 7, and recombined in a step 9 to produce the word HMM. The environment-dependent phonemes HMM are learned by repeating such learning, separating, and connecting processes. Thereafter, whether or not the number of samples learned at the environment-dependent phonemes HMM has been sufficient is judged and, only when the number has been judged to be insufficient, a parameter is obtained in a step 12 as the weighted means of the environment-dependent phonemes HMM and the parameter of corresponding environment-independent phonemes HMM by use of a weighting factor corresponding to the number of the samples learned, and is substituted of the parameter of the environment-dependent phonemes HMM.

Description

Detailed Description of the Invention

[0001]

The present invention relates to a hidden Markov model (hereinafter referred to as "HM") used in a speech recognition method and the like.
M))).

[0002]

[0002] Conventionally, as a technology in such a field,
For example, the following documents have been described. Document 1; The Bell System Technical Journal, 62 "4"
(1983-4) AmericanTelephone and Telegrap
h Company, (US), S. E. Levinson (SELevi
nson), L. R. Rabiner (LR Rabiner), M. M. Songdi (MMSondhi) co-authored "An Introduct
ion to the Application of the Theory of Probabilist
ic Function of a Markov Process to Automatic Speech
"Recognition" p. 1035-1074 Article 2; Seiichi Nakagawa, "Speech recognition by probabilistic model" (Apr. 63-7), Institute of Electronics, Information and Communication Engineers, p. 55-61 Speech recognition techniques include classical pattern matching methods and statistical methods, and the latter has become mainstream in recent years. In the latter statistical method, Markov models with probabilistic finite states are proposed, and usually HMMs
It is called. In general, HMMs have multiple states (eg,
It consists of transitions between voice features and so on) and states. Furthermore, HM
M has a transition probability representing a transition between states, and an output probability of outputting a label accompanying transition (a typical characteristic parameter of speech, which usually has several tens to several thousands). . A speech recognition method using such an HMM is described in the document 1 and an example of the word speech recognition is shown in FIG.

FIG. 2 is a view showing an example of the structure of a word HMM used in the conventional speech recognition method. S ₁ in FIG.
S ₂ , S ₃ , and S ₄ indicate states of speech features and the like in the HMM, and a ₁₁ , a ₁₂ , a ₂₂ , a ₂₃ , a ₃₃ , a ₃₄ , a ₄₄ , a ₄₅
Is the state transition probability, b ₁ (k), b ₂ (k), b ₃ (
k) and b ₄ (k) represent label output probabilities for the label k, respectively. In HMM, state transition probability a
When state transition is performed at _ij (where i = 1,..., 4, j = 1,..., 5), a label is output with the label output probability b _j (k). In order to recognize a uttered word using an HMM, first, using learning data prepared for each word, the HMM is trained so as to output the label string of the word with the highest probability. Next, the label sequence of the uttered unknown word is input, and the word HM giving the highest output probability
Let M be the recognition result. In this type of speech recognition method, an HMM is given to the uttered word itself for learning, and the recognition result is judged by the likelihood (that is, the output probability of the label output). Such word HMMs guarantee excellent recognition accuracy, but an increase in the number of recognition vocabulary requires a large amount of learning data. In addition, there is a disadvantage that voices other than the learning target words can not be recognized at all. On the other hand, in phonetics, a word is usually represented by a series of phonetic elements called phonemes. Therefore, prepare an HMM for each phoneme, and
There is also a method of connecting a MM to generate a word HMM and performing word recognition. However, in the word speech actually uttered, each phoneme is influenced by neighboring phonemes, and feature parameters (for example, spectrum) are considerably deformed. Such deformation of spectrum due to articulatory combination may not be able to be represented by a phoneme HMM. Therefore, such a method of simply connecting phoneme HMMs to recognize words inevitably leads to a decrease in recognition rate.

In order to eliminate the influence of such articulatory coupling, a phoneme model which depends on the phonological environment before and after, ie,
A diphone and a triphone have been proposed. Here, the diphone refers to a phoneme (one-sided environment-dependent phoneme) for which either the preceding phoneme or the subsequent phoneme is known with respect to the target phoneme, and the triphone is a phoneme for which both the preceding phoneme and the subsequent phoneme are known. (Environment-dependent phoneme). When performing speech recognition, a diphone or triphone HMM is prepared, and a word HMM is constructed by linking these HMMs, and word recognition is performed. Environment-dependent phoneme HMMs are environment-independent phoneme HMMs
Compared to the above, it is possible to avoid the decrease in recognition rate due to the spectrum deformation due to articulatory coupling, but since there are a large number of models, it is necessary to prepare a large amount of learning data to learn the HMM. Furthermore, it is necessary to prepare information (ie, label information) indicating the section in which each diphone or triphone exists in the learning data. Moreover, when performing labeling work, for example, automatic work by a computer, satisfactory accuracy can not be obtained, and labeling is almost always performed manually. Then, the learning method which does not require label information conventionally is proposed. First of all,
Prepare environment-independent phoneme HMMs that are easy to learn. Then, the above environment independent phoneme HMMs are connected to the learning data of words (or clauses or sentences, the same shall apply hereinafter) of utterance contents known and not labeled to construct word HMMs, and these word HMMs are constructed. To learn Since the learning of word HMMs, the learning process can be realized if the word boundaries (ie, the start and end of the word) are known. Furthermore, these word HMMs are decomposed in the reverse procedure of concatenation to generate environment-dependent phoneme HMMs. In order to improve learning accuracy, environment-dependent phoneme HMMs are approximately generated by repeating the above-described connected learning and decomposition generation.

[0005]

However, the conventional environment-dependent phoneme HMM learning method has the following problems. For a particular environment-dependent phoneme HMM, the number of corresponding speech data is sometimes very limited, so the environment-dependent phoneme H obtained by learning as described above
MM is easily influenced by learning data. That is, there is a risk of being biased towards learning data. The environment-dependent phoneme HMM is learned by the above-described learning method, and in particular, with respect to the HMM learned from a small amount of learning samples, the parameters of the environment-dependent phoneme HMM and the parameters of the corresponding environment-independent phoneme HMM are weighted There is a method of taking an average and making the average value an environment-dependent phoneme HMM parameter. This is because the environment-independent phoneme HMM is easy to learn from a large amount of learning data, and there is little bias to the learning data. However, in the above-described averaging process of HMM parameters, since a parameter called a weighting factor is dependent on learning data and the value thereof is determined by experiment, it takes time to determine the weighting factor. In addition, depending on the setting of the weighting factor, emphasis is placed on environment independent phoneme HMM parameters, and there is also a possibility that the environment dependent phoneme HMM parameters obtained by learning may not be used. SUMMARY OF THE INVENTION The present invention provides a learning method of environment-dependent phoneme HMMs with less bias to learning data by setting weighting coefficients that is simple and effectively using learning results as problems that the above-described conventional techniques have. is there.

[0006]

[Means for Solving the Problems] In order to solve the above problems, in the first invention, when learning environment-dependent phoneme HMMs, environment-independent phoneme HMMs prepared in advance are connected to form words ( Or construct a clause or sentence) HMM.
Then, a learning process for learning the word (or phrase or sentence) HMM, a decomposition process for decomposing the learning result into an environment-dependent phoneme HMM after the learning process, and reconnection of the decomposed environment-dependent phoneme HMM In the HMM learning method for learning the environment-dependent phoneme HMM by repeating the learning process, the decomposition process and the connection process using the word (or phrase or sentence) HMM and the linking process, the following is performed. I am taking measures. That is, in the first invention, the number of learning samples used to learn the environment-dependent phoneme HMM is counted, and only when it is determined that the number of learning samples is insufficient, the decomposition is finally performed in the repetition. A weighted parameter is calculated by weighting the parameters of the environment-dependent phoneme HMM and the parameters of the environment-dependent phoneme HMM corresponding to the parameters with a weighting factor according to the number of learning samples, and Replacing the parameters of the environment-dependent phoneme HMM decomposed into the above-mentioned parameters,
I try to learn M.

The second invention is a learning method substantially the same as the first invention, but the method of calculating a parameter to be replaced is different when it is judged that the number of learning samples is insufficient. That is, in the second aspect of the invention, the learning sample includes the parameters of the environment-dependent phoneme HMM finally decomposed in repetition and the parameters of the one-sided environment-dependent phoneme HMM prepared in advance corresponding to the parameters. Calculate weighted parameters with weighting factors according to the number,
Environment-dependent phoneme HMM decomposed last in the above iteration
Is replaced with the calculated parameter to learn the environment-dependent phoneme HMM. According to the first aspect of the invention, since the learning method of the HMM is configured as described above, after learning the environment-dependent phoneme HMM, the learning sample of the environment-dependent phoneme HMM is used. The number is counted. If it is determined that the number of samples is not enough, the finally decomposed environment-dependent phoneme HM
M parameter and corresponding environment independent phoneme HM
A parameter of M is weighted and averaged by a weighting factor according to the number of learning samples to calculate a new parameter. Then, the parameters of the environment-dependent phoneme HMM are replaced with the new parameters. If there are a sufficient number of learning samples, replacement processing is not performed. According to the second aspect of the invention, unlike the first aspect of the invention, the parameters of the one-sided environment-dependent phoneme HMM prepared in advance and the parameters of the environment-dependent phoneme HMM decomposed at the end are the number of learning samples. New parameters are calculated by weighted averaging according to the corresponding weighting factors.

[0008]

DESCRIPTION OF THE PREFERRED EMBODIMENTS First Embodiment FIG. 1 is a flow chart of the processing contents of a learning method of an HMM according to a first embodiment of the present invention. The learning method of the HMM according to the present embodiment will be described with reference to this figure. In the learning method of the HMM according to the first embodiment, for example, the processes of steps 1 to 13 in FIG. 1 are performed using a program-controlled computer. First, when learning is started in step 1, speech signals of learning data (for example, words "akai" and "sakae" as word speech) are input in step 2, and the process proceeds to pre-processing of step 3. In the pre-processing of step 3, for example, the input analog voice signal is converted into a digital signal by analog / digital conversion, and LPC (Linear Predictive Codin) is performed.
g, extraction of LPC cepstrum by linear prediction coding) analysis, etc., each state S _i (i =
The speech feature parameters corresponding to 1,..., 4) are extracted, and the process proceeds to step 5.

The environment-independent phoneme HMM dictionary 4 stores, for example, HMMs of Japanese phonemes (about 30 to 40 types). So-called environment independence refers to the fact that the phoneme before and after the phoneme is unknown. This is expressed, for example, as follows. (*) A (*) (*) i (*) (*) u (*) (*) e (*) (*) o (*) (*) k (*) (*) s (*) ( *) T (*) · · · In step 5, referring to the phoneme string representation of the input word and the environment-independent phoneme HMM dictionary 4, the environment-independent phoneme HMM can be expressed, for example, by the following equation (1) Are connected to form a word HMM. HMM of word "akai" * (*) a (*) + (*) k (*) + (*) a (*) + (*) i (*) HMM of word "sakae" * (*) s ( *) + (*) A (*) + (*) k (*) + (*) a (*) + (*) e (*) (1) where "+": connection of HMM means. Next, in step 6, the parameter of the word HMM is estimated using the result of the pre-processing in step 3 and the word HMM configured in step 5 (learning of HMM). H
For estimation of the MM parameter, for example, the Baum-Welch (B-W) algorithm described in the aforementioned reference 2 is used. In this B-W algorithm, for example, the observation label sequence O = o
₁ , o ₂ ,..., O _T and state series I = i ₁ , i ₂ ,.
·, I _T , as in the following equation (2), forward variable α
Define _t (i) and the backward variable β _t (i). α _t (i) = Pr (o ₁ , o ₂ ,..., o _t , i _t = S _i ) β _t (i) = Pr (o _{t + 1} , o _{t + 2} ,..., o _T | i _t = S _i ) (2) And, state transition probability a _ij and label output probability b _j (k)
Is estimated as in the following equation (3).

[0010]

[Equation 1] After learning the word HMM in this manner, in step 7, the word HMM is decomposed into the environment-dependent phoneme HMM, for example, according to the following equation (4). Word HMM "akai"-> (O) a (k); (a) k (a); (k) a (i); (a) i (O) word HMM "sakae"-> (O) s (a) (S) a (k); (a) k (a); (k) a (e); (a) e (O) (4) These environment-dependent phonemes HMMs are environment-dependent phonemes HMM
Save in dictionary 8. At this time, HM called (a) k (a)
Since there are two M's, they are averaged as in the following equation (5) and stored in the environment-dependent phoneme HMM dictionary 8. New (a) k (a) = {(a) k (a) of word "akai" + (a) k (a) of word "sakae" 2/2 (5) where "+": Means "add" in arithmetic.

In step 10, whether the environment-dependent phoneme HMM has converged on a certain criterion (ie, whether the difference between the previous value and the current value of the environment-dependent phoneme HMM parameters is smaller than this criterion) If it does not converge, the environment-dependent phoneme HMM decomposed in step 7 is linked in step 9 as in the following equation (6) to reconstruct a word HMM, and learning of the word HMM in step 6 Then, the learning process and the decomposition process are repeated. HMM of word "akai" k (O) a (k) + (a) k (a) + (k) a (i) + (a) i (O) HMM of word "sakae" O (O) s ( a) + (s) a (k) + (a) k (a) + (k) a (e) + (a) e (O) (6) On the other hand, the result of the determination in step 10 If converged, the learning loop is completed, and it is determined in step 11 whether or not the number of learning samples used for learning the corresponding environment-dependent phoneme HMM is sufficient. That is, the number of learning samples n
Is equal to or greater than a predetermined threshold m, step 13
End learning at.

If the number of learning samples n is smaller than the threshold value m, at step 12, using the weighting factor γ shown in the following equation (7) for the parameters of the environment-dependent phoneme HMM,
Perform weighted averaging processing. γ = (1-n / m) (7) For example, in step 12, the environment-dependent phoneme HMM
Each state parameter of is a _ij (i = 1,..., 4; j =
1, ..., 5), b _j (k) (j = 1, ..., 4)
And each state parameter of the environment independent phoneme HMM is a _ij
0 (i = 1, ..., 4; j = 1, ..., 5), b _j
Assuming 0 (k) (j = 1, ..., 4), each state parameter a _ij , b _j (k) of the new environment-dependent phoneme HMM
Are respectively as in the following equation (8). a _ij = γ × a _ij + (1−γ) × a _ij 0 b _j (k) = γ × b _j (k) + (1−γ) × b _j 0 (k) (8) This The environment-dependent phoneme HMM dictionary 8 is updated with the state parameters a _ij and b _j (k) of the new environment-dependent phoneme HMM obtained in the same manner, and the learning is ended in step 13. As described above, the first embodiment has the following advantages. According to the first embodiment, in step 11, it is judged based on the threshold value m whether or not the number of learning samples n is sufficient. Then, in step 12, based on the threshold value m, a weighting factor γ corresponding to the number n of samples used in actual learning is determined. For this reason, it is not necessary to manually determine the weighting factor γ while looking at the experimental results, and rapid learning is possible, and the number n of learning samples is reflected in the weighting factor γ. Can be reduced.

Second Embodiment FIG. 3 is a flow chart of the processing contents of a learning method of an HMM according to a second embodiment of the present invention, and the processing common to the processing in FIG. It is done. In the first embodiment, the environment-independent phoneme HMM of the environment-independent phoneme HMM dictionary 4 is used in step 12 of FIG. 1. On the other hand, in the second embodiment, a step 12A is provided instead of the step 12 of the first embodiment,
According to this step 12A, the diphone and the phoneme HMM
A weighted averaging process of HMM parameters using one-sided environment-dependent phoneme (that is, diphone) HMM prepared in advance in the dictionary 4A is performed. That is, in the HMM learning method of the second embodiment, as in the first embodiment,
For example, using a program controlled computer,
The processes of steps 1 to 13 in FIG. 3 are performed. Processing until it is determined whether or not the number of learning samples used for learning the environment-dependent phoneme HMM is sufficient is performed in steps 1 to 11 similar to the first embodiment. If the number of learning samples is sufficient, the learning is finished as it is in step 13. If the number of learning samples is insufficient, the process proceeds to step 12A.

In step 12A, the environment-dependent phoneme HM is
(7) for the parameters of M, weighted averaging with the parameters of the one-sided environment-dependent phoneme HMM prepared in advance in the diphone and phoneme HMM dictionary 4A Do. For example, in step 12, each state parameter of the environment-dependent phoneme HMM is a
_ij (i = 1, ..., 4; j = 1, ..., 5), b _j
(K) (j = 1,..., 4), and each state parameter of the one-sided environment dependent phoneme HMM is a _ij 1 (i = 1,...
· 4; j = 1, ..., 5), b _j 1 (k) (j =
1, ..., 4), the new environment-dependent phoneme HM
Each state parameter a _ij , b _j (k) of M is as shown in the following equation (9). a _ij = γ × a _ij + (1−γ) × a _ij 1 b _j (k) = γ × b _j (k) + (1−γ) × b _j 1 (k) (9) The environment-dependent phoneme HMM dictionary 8 is updated with the state parameters a _ij , b _j (k) of the new environment-dependent phoneme HMM obtained in the same manner, and the learning is finished in step 13. As described above, the second embodiment has the following advantages (1) and (2).

(1) According to the second embodiment, since the weighting factor γ is calculated as in the first embodiment, it is not necessary to manually determine the weighting factor γ, and rapid learning is possible. It becomes. Moreover, since the number of learning samples is reflected in the weighting factor γ, it is possible to reduce the bias to the learning data. (2) In the second embodiment, since the one-sided environment-dependent phoneme HMM is used in the weighted averaging process of step 12A, there are the following advantages different from the first embodiment. Environment-independent phoneme HM used in the first embodiment
Since M does not consider the phonological environment before and after, the number of phonemes is very limited. Therefore, it is easy to complete the environment independent phoneme HMM dictionary 4. On the other hand, since the HMM parameters obtained by the environment independent phoneme HMM dictionary 4 are extremely generalized, there is no possibility that a difference will occur when learning the environment dependent phoneme HMM of each word. Absent. On the other hand, the single-sided environment-dependent phoneme HMM used in the second embodiment is in a state close to the environment in which each word is actually used, in order to consider either the phonological environment before or after. HMM parameters can be obtained. For this reason, when learning environment-dependent phoneme HMMs of individual words, the difference is less likely to occur. However, the combination of phonetic environments increases the number of phonemes, making it difficult to complete the diphone and phoneme HMM dictionary 4A. Therefore, when a one-sided environment-dependent phoneme HMM is prepared in advance for each phoneme constituting a word to be learned, the HMM according to the second embodiment
It is possible to use the learning method of

As described above, in the second embodiment, since the one-sided environment-dependent phoneme HMM prepared in advance is used in the weighted averaging process of step 12 A, smoothing of parameters can be achieved. It becomes possible, and more accurate speech recognition becomes possible. In addition, this invention is not limited to the said embodiment, A various deformation | transformation is possible. Examples of this variation include the following. (A) In the first and second embodiments, the learning method of the HMM is described using the four-state HMM shown in FIG. 2 as an example, but the present invention is similarly applicable to HMMs having other numbers of states. (B) In the first and second embodiments, the learning method of the HMM for the input word speech is described.
Even when speech of a phrase or a sentence is input, environment-dependent phoneme HMMs can be similarly learned.

[0017]

As described above in detail, according to the first invention, the environment-dependent phoneme HM which has not been sufficiently learned.
For M only, replace the parameter with a weighted average with a weighting factor according to the number of training samples. This makes it possible to speed up the learning process. Furthermore,
Without losing the advantages of the conventional environment-dependent phoneme HMM
Since bias to learning data can be reduced and parameters can be smoothed, highly accurate speech recognition is possible. According to the second invention, as in the first invention,
Environment-dependent phoneme H with a weighting factor according to the number of training samples
Since the averaging process of the MM parameters is performed, the learning process can be speeded up. Moreover, since this averaging process uses the parameters of the one-sided environment-dependent phoneme HMM, parameters close to the environment in which the individual words (or phrases or sentences) are actually used can be obtained. Therefore, as compared to the first invention, it is possible to reduce the bias to the learning data and to further smooth the parameters without losing the merits of the conventional environment-dependent phoneme HMM.
Furthermore, high-accuracy speech recognition is possible.

Brief Description of the Drawings

FIG. 1 is a flowchart of processing content of a learning method of an HMM according to a first embodiment of the present invention.

FIG. 2 is a view showing an example of the structure of a word HMM used in a conventional speech recognition method.

FIG. 3 is a flowchart of processing contents of a learning method of an HMM according to a second embodiment of the present invention.

[Description of the code]

4 Environment-independent phoneme HMM dictionary 4A diphone and environment-independent phoneme HMM dictionary 5 Step of configuration process of word HMM 6 step of learning process of word HMM 7 step of decomposition process of resolving word HMM into environment-dependent phoneme HMM 8 environment Dependent Phoneme HMM Dictionary 9 Concatenation of Environment Dependent Phoneme HMMs and Word H
Step of connected learning processing to reconstruct MM Step of convergence judgment processing of environment-dependent phoneme HMM Step of judgment processing of number of learning samples 12, 12A Step of weighted averaging processing of HMM parameters

Claims (2)

[Claim of claim] 1. When learning environment-dependent phoneme hidden Markov models, environment-independent phoneme hidden Markov models prepared in advance are connected and any one of a word, a sentence or a sentence is connected. A learning process for constructing a Hidden Markov model and learning any one hidden Markov model, and a decomposition process for decomposing the learning result into an environment-dependent phoneme hidden Markov model after the learning process; The decomposed environment-dependent phoneme hidden Markov
The environment-dependent phoneme hidden by repeating the learning process, the decomposition process, and the connecting process using the connecting process of reconnecting the model and creating the hidden Markov model of any one of a word, a clause or a sentence -In the learning method of hidden Markov models for learning Markov models, counting the number of learning samples used for learning the environment-dependent phoneme hidden Markov models, and if the number of learning samples is insufficient Only when judged, the parameters of the environment-dependent phoneme hidden Markov model finally decomposed in the repetition and the parameters of the environment-independent phoneme hidden Markov model corresponding to the parameters are the learning sample Calculate the weighted averaged parameter with the weighting factor according to the number, and the environment finally decomposed in the repetition It exists type phoneme Hidden-substitute the parameter Markov model parameters the calculated learning method of hidden Markov models, characterized by learning the environment-dependent phoneme Hidden Markov Model. 2. When learning environment-dependent phoneme hidden Markov models, environment-independent phoneme hidden Markov models prepared in advance are connected to form any one of words, phrases or sentences. A learning process for constructing a Hidden Markov model and learning any one hidden Markov model, and a decomposition process for decomposing the learning result into an environment-dependent phoneme hidden Markov model after the learning process; The decomposed environment-dependent phoneme hidden Markov
The environment-dependent phoneme hidden by repeating the learning process, the decomposition process, and the connecting process using the connecting process of reconnecting the model and creating the hidden Markov model of any one of a word, a clause or a sentence -In the learning method of hidden Markov models for learning Markov models, counting the number of learning samples used for learning the environment-dependent phoneme hidden Markov models, and if the number of learning samples is insufficient Only when it is judged that the parameters of the environment-dependent phoneme hidden Markov model finally decomposed in the repetition and the parameters of the one-sided environment-dependent phoneme hidden Markov model prepared in advance corresponding to the parameters And calculating a weighted average of weighting factors according to the number of learning samples, and It has been replaced by the parameters of the environment-dependent phoneme Hidden Markov model parameters the calculated learning method of hidden Markov models, characterized by learning the environment-dependent phoneme Hidden Markov Model decomposed.