JPH0981178A - Unspecified speaker model generating device and voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To always generate an unspecified speaker model by independently clustering the output Gaussian distribution of each state of the hidden Markov model of single Gaussian distribution of plural specific speakers and synthesizing them. SOLUTION: The unspecified speaker model generation section 31 learns the output Gaussian distribution only for the state, in which data exist, based on the uttered voice data of plurality N specific speakers stored in the memory of uttered voice data 30 of the specific speakers, extracts only the parameters of the learned output Gaussian distributions among the specific speaker's models and performs clustering for every state corresponding to the hidden Markov model(HMM). Then, synthesis and mixing are performed to generate a hidden Markov network (an HM network) 11 of the mixed Gaussian distributions and it is stored in the memory of the network 11. Then, voice recognition is performed by referring to the network 11. In other words, the output Gaussian distribution of each state of the hidden Markov model of a single Gaussian distribution of plural specific speakers is independently clustered for every state and an HMM is generated.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device for creating an unspecified speaker model, which creates a hidden Markov model for an unspecified speaker (hereinafter referred to as an HMM) based on hidden Markov models for a plurality of specified speakers. And a speech recognition device using the unspecified speaker model creation device.

[0002]

2. Description of the Related Art Conventionally, a Baum-Welch learning algorithm (hereinafter referred to as a first conventional example) for creating an HMM of an unspecified speaker based on a specific speaker model for learning. ) Is widely used (for example, Seiichi Nakagawa, “Speech recognition by probabilistic model”, p.
p. 55-64, Institute of Electronics, Information and Communication Engineers, published in July 1988. ). In the first conventional example, the partial observation sequence {y ₁ , y ₂ , y ₃ , between the time 1 and the time t in the HMM is used.
, Y _t } after observation, the forward probability of being in state i at time t, and the partial observation sequence {y _{t + 1} , y _{t + 2} , y from time t + 1 to the end at state t at time t _The backward probability of observing _{t + 3} , ..., Y _r } is used to re-estimate and learn the parameters of the HMM to create the HMM of the unspecified speaker.

Using the method of the first conventional example, it is desirable to learn a model with voice data of a large number of speakers in order to cope with variations in acoustic feature amounts of voices of various speakers. It tends to be large, and it is desirable to learn the model with a large amount of voice data by a large number of speakers. However, when dealing with such a large amount of data,
The enormous amount of calculation is a problem even now, as the processing speed of computers is increasing.

In order to reduce the calculation amount of such an unspecified speaker model, the CCL method (hereinafter referred to as a second conventional example) by speaker clustering and model synthesis by the specific speaker model has already been used by Kosaka et al. Proposed (Conventional document 2 "Kosaka et al.," Independent speaker model creation method using clustering method ", Proceedings of Acoustical Society of Japan, 1-R-1
2, November 1994 ”. ). In the method of the second conventional example, clustering is performed with the entire model set covering all phonemes as a unit under the assumption that the acoustic characteristics of the voices of the speakers are similar in all acoustic spaces. There is. Specifically, a well-trained specific speaker model is clustered by defining the distance between the models, and then each specific speaker model is synthesized to create an unspecified speaker model.

[0005]

In the method of the second conventional example, it is possible to create an unspecified speaker model with a small amount of calculation, but all parameters of the specified speaker model are sufficiently learned. If there is no such model, a good model cannot be obtained, and a large amount of vocal data is required for each speaker. Further, the number of mixed output Gaussian distributions is always the same in all the states of the HMM, and there is a problem that unnecessary parameters increase in a state where there is little variation in the feature amount depending on the speaker.

The first object of the present invention is to solve the above problems and it is not necessary for all the parameters of each specific speaker model to be learned, and the parameters learned for each speaker are different. It is an object of the present invention to provide an unspecified speaker model creation device that can create an unspecified speaker model even when the above is present, and that requires less memory capacity of the processing device and shortens the calculation time. A second object of the present invention is, in addition to the above first object, that speech recognition can be performed using the created unspecified speaker model, and the speech recognition rate is improved compared to the conventional example. It is to provide a voice recognition device that can perform.

[0007]

According to a first aspect of the present invention, there is provided a device for creating a speaker-independent model according to claim 1, which is based on a hidden Gaussian Markov model of a single Gaussian distribution of a plurality of specified speakers. In an unspecified speaker model creating device that creates a hidden Markov model of a mixed Gaussian distribution of speakers, the output Gaussian distribution of each state of the input hidden Markov model of a single Gaussian of a plurality of specific speakers is input for each state. It is characterized by comprising a model creating means for creating a hidden Markov model of a mixed Gaussian distribution of an unspecified speaker by independently clustering and synthesizing.

The unspecified speaker model creating device according to claim 2 is the unspecified speaker model creating device according to claim 1, wherein the model creating means inputs the uttered voices of the specified speakers. Based on the data, by learning the output Gaussian distribution only for the state in which the uttered voice data exists by a predetermined learning method using the same initial speaker hidden Markov model for a plurality of speakers, Each output Gaussian based on the learning means for creating a single hidden Gaussian Markov model for a specific speaker and the plurality of hidden Gaussian hidden Markov models for a specific speaker created by the learning means Clustering means for performing clustering on a plurality of clusters so that the output Gaussian distribution is included in a shorter distance for each cluster based on the distance between the distributions; The hidden Markov model with multiple output Gaussian distributions in each cluster is synthesized into the hidden Markov model with single Gaussian distribution in each cluster, based on the hidden Markov model with single Gaussian distribution clustered for each state by the Tulling means. A synthesizing means and a synthesizing means for creating a hidden Markov model of a mixed Gaussian distribution of an unspecified speaker by mixing the hidden Markov models of the single Gaussian distributions of the respective states synthesized by the synthesizing means. Is characterized by.

Further, the speaker-independent speaker model creating apparatus according to claim 3 is the speaker-independent speaker model creating apparatus according to claim 2, wherein the clustering means is equal to or more than a threshold value set in advance for each state. It is characterized in that only the output Gaussian distribution learned with the data amount is extracted and then clustered.

Still further, in the speaker-independent speaker model producing apparatus according to claim 4, in the speaker-independent speaker model producing apparatus according to claim 2 or 3, the clustering means is provided for each cluster clustered in each state. By repeating the clustering until the average value of the distance between the center and each output Gaussian distribution becomes less than or equal to the predetermined distance, the clusters in each state are increased so that the larger the variation of each output Gaussian distribution in each state is, the larger the number of clusters becomes. Characterized by determining the number.

According to a fifth aspect of the speech recognition apparatus of the present invention, based on the input hidden Gaussian Markov model of a plurality of specific speakers, a hidden Gaussian mixture of unspecified speakers is hidden. An unspecified speaker model creation device according to any one of claims 1 to 4, which creates a Markov model,
Based on the voice signal of the input uttered voice sentence, using a hidden Markov model of the mixture distribution of the unspecified speakers created by the above-described unspecified speaker model creating device, a voice recognition means for recognizing the voice is provided. It is characterized by

[0012]

DETAILED DESCRIPTION OF THE INVENTION Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a voice recognition device according to an embodiment of the present invention. The speech recognition apparatus according to the present embodiment uses, in particular, a known maximum likelihood estimation method based on the uttered voice data of a plurality of N specified speakers stored in the memory of the uttered voice data 30 of the specified speaker. The output Gaussian distribution is learned only for existing states, only the parameters of the output Gaussian distribution learned from the specific speaker model are taken out, clustering is performed for each corresponding state of the HMM, and then synthesis and mixing are performed. A hidden Markov network (hereinafter referred to as an HM network) having a mixed Gaussian distribution is created by using the unspecified speaker model creating unit 31 that stores the created HM network in the memory of the HM network 11. It is characterized in that the voice recognition is performed by referring to the stored HM network.

This voice recognition device comprises a microphone 1 and
Feature extraction unit 2, buffer memory 3, phoneme matching unit 4
And a phoneme context-dependent LR that executes speech recognition processing by referring to the memory of the LR table 13 stored in the memory, which is created based on a predetermined context-free grammar stored in the memory of the context-free grammar database 20. And a parser (hereinafter referred to as LR parser) 5.

FIG. 2 is a flowchart showing an unspecified speaker model creating process executed by the unspecified speaker model creating unit 31. In the creation process, first, in step S1, the characteristic parameters of the uttered voice data are extracted based on the uttered voice data of a plurality of N specific speakers, and based on the extracted characteristic parameters, Mean value and variance of output Gaussian distribution only for states in which data exist by a known maximum likelihood estimation method using an initial speaker model (mixture of states 1) that is the same HM network for all speakers. By learning N, a single Gaussian HM network for N specific speakers is created.

Next, in step 2, as shown in FIG. 3, the HMs of N single Gaussian distributions for specific speakers are created.
After extracting only the output Gaussian distribution learned with a data amount equal to or more than a preset threshold value for each state based on the network, as shown in FIG. 4, a known buttercha (Bhattacharya) between the output Gaussian distributions is obtained. Based on the distance, clustering is performed on a plurality of clusters so that the output Gaussian distribution is included in a shorter distance for each cluster.
Here, the threshold value is set for the learning data amount to be taken out so that the unreliable output Gaussian distribution does not adversely affect the clustering. As a result, a highly reliable HM network 11 can be obtained, and the HM network 11
By performing voice recognition using, it is possible to perform voice recognition with a higher voice recognition rate than in the conventional example. Further, in the clustering, in each state, clustering is repeated in each state until the average value of known Bhattacharya distances between the center of each cluster and each output Gaussian distribution becomes equal to or less than a predetermined distance. The number of clusters K is determined according to the variation of the output Gaussian distribution of each member. Here, when the variation is large, the number of clusters K is set to be relatively large, while when the variation is small, the number of clusters K is set to be relatively small. In determining the number of clusters K, the maximum number of clusters Kmax and the minimum number of clusters Kmin may be set. Further, when the learning data amount is small, preferably the number of clusters K is set small.

Next, in step S3, using the results of clustering for each state in step S2, as shown in FIG. 5, a plurality of output Gaussian distributions in the cluster are converted into a single Gaussian distribution in each state. To synthesize. The synthesis was performed by the same method as the method of the conventional document 2 except that the total number of output Gaussian distributions and the clustering result were different for each state. The composition method of step S3 will be described in detail later. Further, in step S4, the combined single Gaussian distributions of all the clusters for each state are mixed using a known speaker mixing method to create an HM network of mixed Gaussian distributions, and the HM network 11 Store in memory. The mixing ratio was set to a value proportional to the ratio of the total learning data amount of the output Gaussian distribution of the members of each cluster. That is, the larger the learning data amount of each cluster member, the larger the mixing ratio is set.

The average value μh _j and the variance Sh _j after combination in each cluster used in step S3 are expressed by the following equations 1 and 2. The weight coefficient w _j ⁽ⁱ⁾ is expressed by the following equation 3.

[0018]

[Equation 1]

[Equation 2]

(Equation 3)

Equations 1 and 2 represent the average value and the variance when a plurality of Gaussian distributions are regarded as a single Gaussian distribution. Here, μ _j ⁽ⁱ⁾ and S _j ⁽ⁱ⁾ are the natural number i-th HM
It represents the mean and variance of the output probability density function which is a single Gaussian distribution in the state j of the network. Further, n _j ⁽ⁱ⁾ represents the number of samples in the state j of the i-th HM network. That is, as is clear from Equation 1, the average value μh _j and the variance Sh _j after the synthesis are respectively the average value μ _j and the variance S _j before the synthesis.
Is calculated with a larger weighting factor w _j ⁽ⁱ⁾ as the number of samples n _j ⁽ⁱ⁾ increases according to the number of samples n _j ⁽ⁱ⁾ in each state.

In this embodiment, the HM network 11 is used as a statistical phoneme model set for speech recognition. The HM network 11 is a phoneme environment-dependent model that is efficiently expressed. One HM network contains many phoneme environment dependent models. The HM network 11 is composed of a combination of states including a Gaussian distribution, and the states are shared between individual phoneme environment dependent models. Therefore, a robust model can be created even when the number of data for parameter estimation is insufficient. This HM network 11 uses the sequential state division method (Successive Sta
te Splitting: Hereinafter referred to as SSS. ) Is used to automatically create. In the above SSS, determination of the topology of the HM network,
Simultaneous determination of abnormal sound clusters and estimation of parameters of Gaussian distribution in each state. In the present embodiment, the output probability and the transition probability represented by the Gaussian distribution are included as the parameters of the HM network. Therefore, at the time of recognition, it can be handled like a general HMM.

Next, the SSS-LR (left-to-right right most) using the speech recognition method of this embodiment described above is used.
(Type) An unspecified speaker continuous speech recognition device will be described. This device uses a phoneme environment-dependent efficient HMM representation format called HM network 11 stored in a memory. Further, in the above SSS, the likelihood is compared with the stochastic model in which the temporal transition of the speech parameter is expressed by the stochastic transition between the stochastic stationary signal sources (states) assigned in the phoneme feature space. The model refinement is performed sequentially by repeating the operation of dividing each state in the context direction or the time direction based on the maximization criterion.

In FIG. 1, the vocalized voice of the speaker is input to the microphone 1 and converted into a voice signal, and then input to the feature extraction unit 2. The feature extraction unit 2 performs, for example, LPC analysis after A / D conversion of the input voice signal, and a 34-dimensional feature parameter including logarithmic power, 16th-order cepstrum coefficient, Δ logarithmic power, and 16th-order Δ cepstrum coefficient. To extract. The time series of the extracted characteristic parameters is input to the phoneme matching unit 4 via the buffer memory 3.

HM in the memory connected to the phoneme collation unit 4
The network 11 is represented as a plurality of networks in which each state is a node, and each state has the following information. (A) State number (b) Acceptable context cluster (c) List of preceding states and succeeding states (d) Parameter of output probability density distribution (e) Probability of self-transition and transition to subsequent states

The phoneme matching unit 4 executes a phoneme matching process in response to a phoneme matching request from the phoneme context dependent LR parser 5. Then, the likelihood for the data in the phoneme matching section is calculated using the unspecified speaker model, and the value of this likelihood is returned to the LR parser 5 as a phoneme matching score. Since the model used at this time is equivalent to the HMM, the forward path algorithm used in the normal HMM is used as it is for the calculation of the likelihood.

On the other hand, a predetermined context-free grammar (C
The FG) database 20 is automatically converted as known to create an LR table and stored in the memory of the LR table 13. The LR parser 5 refers to the LR table 13 and processes the input phoneme prediction data from left to right without backtracking. In the case of syntactic ambiguity, the stack is split and parsing of all candidates is processed in parallel. The LR parser 5 has an LR table 13
The next phoneme is predicted and the phoneme prediction data is output to the phoneme matching unit 4. In response to this, the phoneme collation unit 4 collates by referring to the information in the HM network 11 corresponding to the phoneme, returns the likelihood to the LR parser 5 as a speech recognition score, and sequentially connects the phonemes. As a result, the continuous voice is recognized, and the voice recognition result data is output. When a plurality of phonemes are predicted in the continuous speech recognition, the existence of all of them is checked, and a pruning is performed by using a beam search method to leave a partial tree having a high likelihood of partial speech recognition. To achieve high-speed processing.

In the above embodiment, the voice data 30 of the specific speaker, the HM network 11, and the LR table 13 are used.
And the context-free grammar database 20 are stored in, for example, a hard disk memory. The phoneme collation unit 4, the LR parser 5, and the unspecified speaker model creation unit 31 are composed of, for example, a digital computer.

In the above embodiment, the unspecified speaker model is created by the unspecified speaker model creating process of FIG. 2. However, the known balm welch for the HM network created by the creating process is used. The learning algorithm may be used to re-learn to create an HM network.

[0028]

EXAMPLE The present inventor conducted an experiment as described below in order to confirm the effectiveness of the speech recognition apparatus of FIG. In the experiment, an HM network that effectively shares the state of a context-dependent phoneme HMM (see, for example, Document 3 “Takami et al.,“ Automatic Generation of Hidden Markov Network by Sequential State Division with respect to Phoneme Context and Time ”], Electronic See IEICE Technical Report, SP91-88, December 1991.). The structure of HM network is 2620 uttered by one speaker.
Two types of models with total state numbers of 200 and 600 were created by making decisions using word voice data. 1 for each model
A silent model of state 10 mix was added. The initial speaker model for learning the specific speaker model was a single distribution in each state except for the silent model, and the initial values of the parameters were determined with the same speech data as the structure determination. Based on this initial speaker model, a speech recognition database of the spontaneous utterance owned by the applicant of the present patent and having a task of travel planning (see, for example, conventional document 4 “T. Morimoto et al.,“ A ”).
Speech and Language Database
se for Speech Translation
Research ”, Proc. Of ICSLP '
94, pp. 1791-1794, 1994 '').
The specific speaker model for 81 persons was created by learning the average value and variance of the output Gaussian distribution by the maximum likelihood estimation method using natural speech data of 81 males in the above. However, since the amount of data per person is as small as about 20 utterances, the variance is updated only when the value is larger than the initial parameter.
This time, we made an uncharacteristic speaker model and conducted a recognition experiment using only male speakers. The recognition experiment was performed on 9 men who were not included in the learning data selected from the same natural speech database as that used for learning.

The unspecified speaker model is created by the CCL method of the second conventional example using model-based clustering with the entire HM network as a unit and the method using the result of HMM state-based clustering according to the present invention. We compared the performance of the two by phoneme recognition experiments. However, in the state-based clustering method according to the present invention, the state occupancy data amount at the time of learning is 10 out of the output Gaussian distribution of each state of the specific speaker model.
Only more than the frame was used. Furthermore, we also compared the recognition rate of the model retrained by the learning algorithm of Balm-Welch with the model created by state clustering as the initial model. In addition,
The HMM according to the clustering method according to the present invention
After creating, we also conducted an experiment on the recognition rate of the model retrained by the learning algorithm of Balm Welch. Table 1 shows the analysis conditions, the use parameters, and the learning / recognition data, which are the experimental conditions.

[0030]

[Table 1] Experimental conditions ───────────────────────────────── Analysis conditions Sampling frequency = 12 KHz Hamming window = 20 ms Frame Cycle = 5ms ───────────────────────────────── Used parameter 16th-order LPC cepstrum + 16th-order Δ cepstrum + logarithmic power + Δ Logarithmic power ───────────────────────────────── Learning data 81 males --- 1 conversation for each speaker (total 1799) Utterance) ───────────────────────────────── Unspecified speaker model evaluation data 9 males − − each speaker 1 conversation (11 to 29 vocalizations) ─────────────────────────────────

Tables 2 and 3 show the CCL method of the second conventional example (hereinafter abbreviated as model clustering in the table) and the method by state-based clustering according to the present invention (hereinafter, state-based clustering in the table). The total number of output Gaussian distributions included in the HM network of each state and mixture number created in (1) is shown. In the case of the CCL method of the second conventional example, the number of mixture distributions is the same for all states except for the silent model, but in the case of state-based clustering according to the present invention, each state is extracted from the specific speaker model. Was done 10
If the number of output Gaussian distributions learned with data over a frame is less than the set number of mixed distributions, the number of extracted distributions will be the number of mixed distributions in that state, so the total number of distributions will be smaller than that with model-based clustering ing.
However, this time, the degree of variation in the average value of the extracted output Gaussian distribution in each state is not considered in the determination of the number of mixtures. In this way, when using a free-speech speech database in which it is difficult to collect speech data considering phoneme balance, it is possible to prevent unnecessary parameter increase by designing the number of mixture distributions for each state. I know there is.

[0032]

[Table 2] Total number of distributions of unspecified speaker model-In case of HM network in 201 state ────────────────────────────── ────── Preparation method / mixture number 5 10 15 20 ──────────────────────────────────── Model clustering 1010 2010 3010 4010 ─────────────────────────────────── State clustering 979 1903 2798 3678 ── ──────────────────────────────────

[0033]

[Table 3] Total number of distributions of unspecified speaker model-601 state HM network ────────────────────────────── ────── Preparation method / mixture number 3 5 10 15 ───────────────────────────────────── Model clustering 1810 3010 6010 9010 ──────────────────────────────────── Clustering by state 1617 2540 4614 6447 ── ──────────────────────────────────

Tables 4 and 5 show the results of the phoneme recognition experiment using the unspecified speaker model created by each method. The results in the table show average values for 9 men.

[0035]

[Table 4] Comparison of phoneme recognition rates (%) by the model creation method-In the case of HM network in 201 state ─────────────────────────── ───────── Preparation method / mixture number 5 10 15 20 ──────────────────────────────── ─── Balm Welch 65.9 66.8 − − ──────────────────────────────────── Model Clustering 62.2 62.5 63.3 63.2 ─────────────────────────────────── By status Clustering 63.6 64.1 64.0 64.5 ─────────────────────────────────── By status Clustering 68.0 68.6 − − + balm woo Ruchi ───────────────────────────────────

[0036]

[Table 5] Comparison of phoneme recognition rates (%) according to the model creation method-In the case of 601 state HM network ─────────────────────────── ───────── Preparation method / mixture number 3 5 10 15 ───────────────────────────────── ─── Balm Welch 67.6 67.8 − − ──────────────────────────────────── Model Clustering 65.1 65.5 66.2 66.2 ──────────────────────────────────── Clustering 67.8 67.9 67.8 67.8 ─────────────────────────────────── Clustering 69.2 69.2 − − + Balm Weh Ji ───────────────────────────────────

Looking at the results of Tables 4 and 5 together with the results of Tables 2 and 3, the method by state-based clustering according to the present invention is the CCL of the second conventional example under all conditions.
High recognition performance is shown with a smaller number of parameters than in the case of the method, and the difference in recognition rate is larger in the case of the 601 state than in the case of 201 states of the HM network.
Considering the speed of actual recognition processing and the efficiency of speaker adaptation, it is considered that the performance as an unspecified speaker model is better when high recognition performance is obtained with as few parameters as possible. It is shown that the method according to the state-based clustering according to the present invention is an effective method for obtaining a model with good performance.

When the relationship between the number of states of the HM network and the recognition performance is examined, the HM network in the 601 state shows higher recognition performance than the HM network in the 201 state, which is the CCL of the second conventional example. The same applies to both the method and the state-based clustering method according to the present invention. This is 201
This is probably because the phonological environment has not been sufficiently segmented and modeled in the state. If the phonological environment is not state-divided so that it is sufficiently subdivided,
The output Gaussian distribution of each state must represent changes in acoustic features due to both phonological and speaker environments at the same time, making it difficult to distinguish between phonological and speaker characteristics, and possible recognition errors. Is expected to increase.

Further, as is clear from Tables 4 and 5, when clustering is performed by the clustering method according to the present invention and then relearning is performed using the learning algorithm of Balm-Welch, compared to other methods, A high phoneme recognition rate is obtained.

Finally, the time for creating the unspecified speaker model will be described. It has been reported that the CCL method of the second conventional example disclosed in the conventional document 2 requires only a few percent of the calculation time of the learning algorithm of Balm-Welch. When the state-based clustering according to the present invention is used, the number of times of clustering increases,
Although the calculation time is increased as compared with the conventional CCL method of
This time is much smaller than the learning time of the specific speaker model, which occupies most of the time required to create the model. Therefore, when viewed from the entire time, the CCL of the second conventional example is used.
Similar to the method, it is possible to create an independent speaker model with a calculation time of several percent of the learning algorithm of Balm-Welch.

As described above, according to the embodiment of the present invention, the output Gaussian distributions of the respective states of the input single Gaussian HMMs of a plurality of specific speakers are clustered independently for each state. Since an HMM having a mixed Gaussian distribution of unspecified speakers is created by synthesizing by using all the parameters, it is not necessary to learn all the parameters of each specific speaker model, and the parameters learned for each speaker are different. You can deal with the case. Therefore, the present invention can be used for voice data of a speaker having a small number of utterances and data having different utterance contents for each speaker, such as free utterance voice. Furthermore, the number of clusters to be divided for each state can be determined by using the variation of the average value of the output Gaussian distribution extracted from each specific speaker model and the information of the learning data amount for each state of the HMM. It is possible to determine the number of mixture distributions in consideration of the learning data amount and the degree of variation of acoustic characteristics between speakers for each state of the HMM. By performing voice recognition using the HMM of the unspecified speaker model, it is possible to perform voice recognition with a higher voice recognition rate than in the conventional example.

[0042]

As described in detail above, according to the speaker-independent model creating apparatus according to claim 1 of the present invention, based on the input hidden Gaussian Markov model of a plurality of specific speakers. , The output Gaussian distribution of each state of the input single Gaussian hidden Markov model of a plurality of specified speakers in the device for creating a hidden Markov model of the mixed Gaussian distribution of the unspecified speaker Modeling means for creating a hidden Markov model of a mixed Gaussian distribution of an unspecified speaker by independently clustering and synthesizing each state. Specifically, the model creating means uses a predetermined learning method by using the same initial speaker hidden Markov model for a plurality of speakers, based on the input vocal data of a plurality of specific speakers. Learning means for creating a hidden Markov model of a plurality of single Gaussian distributions for specific speakers by learning the output Gaussian distribution only for the state in which the uttered voice data exists, and the learning means Based on the Hidden Markov Model of multiple single-Gaussian distributions for specific speakers, based on the distance between each output Gaussian distribution, clustering into multiple clusters so that each cluster contains the output Gaussian distribution in a shorter distance Based on the clustering means for performing and the hidden Markov model of the single Gaussian distribution clustered for each state by the above clustering means. Synthesizing means for synthesizing multiple output Gaussian hidden Markov models in each cluster into a single Gaussian hidden Markov model in each state, and a single Gaussian hidden Markov model in each state synthesized by the above synthesizing means And mixing means for creating a hidden Markov model of a Gaussian mixture mixture of unspecified speakers.

That is, by taking out only the output Gaussian distribution learned from a large number of specific speaker models and performing clustering independently in each state of the HMM, the magnitude of variation of the feature amount and the learning data amount in each state can be determined. The number of clusters can be determined in consideration, and the optimum number of output Gaussian distributions can be determined for each state. Also, since only the output Gaussian distributions trained for each specific speaker model can be selectively used, it is not necessary for all output Gaussian distributions for each specific speaker model to be trained. It can be effectively used for databases with a small amount. In addition, since the parameters are estimated separately for each speaker, the amount of calculation is drastically reduced as compared with the method using the learning algorithm of Balm-Welch of the first conventional example in which all data are used at one time for learning. Is possible. Therefore, the time required to create the unspecified speaker model can be significantly shortened.

Further, according to the speaker-independent speaker model creating apparatus of the third aspect, the clustering means extracts only the output Gaussian distribution learned with the data amount equal to or more than the preset threshold value for each state. Then cluster. This makes it possible to create an optimal speaker-independent model with higher reliability. Therefore, by performing voice recognition using the unspecified speaker model, voice recognition can be performed at a higher voice recognition rate than in the conventional example.

Furthermore, according to the speaker-independent model creating apparatus of the fourth aspect, the clustering means determines in advance the average value of the distances between the centers of the clusters clustered in each state and the output Gaussian distributions. The number of clusters in each state is determined such that the larger the variation of each output Gaussian distribution in each state is, the larger the number of clusters is by repeating the clustering until the distance becomes equal to or less than the distance. Therefore, the number of clusters can be determined in consideration of the variation of each output Gaussian distribution in each state, and the optimal number of output Gaussian distribution can be determined for each state. This makes it possible to create an optimal speaker-independent model with higher reliability. Therefore, by performing voice recognition using the unspecified speaker model, voice recognition can be performed at a higher voice recognition rate than in the conventional example.

According to the speech recognition apparatus of the fifth aspect of the present invention, the mixed Gaussian distribution of unspecified speakers is based on the input hidden Markov model of a single Gaussian distribution of a plurality of specified speakers. 5. The unspecified speaker model creating device according to claim 1, which creates the hidden Markov model, and the unspecified speaker model creating device based on a voice signal of an input uttered voice sentence. And a voice recognition means for recognizing a voice using the hidden Markov model of the mixture distribution of the unspecified speakers created by. Therefore, by performing voice recognition using the unspecified speaker model, voice recognition can be performed at a higher voice recognition rate than in the conventional example.

[Brief description of drawings]

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

FIG. 2 is a flowchart showing an unspecified speaker model creation process executed by an unspecified speaker model creation unit in FIG.

FIG. 3 is a diagram showing a learning process of a specific speaker model and a process of extracting an output Gaussian distribution in the non-specific speaker model creating process executed by the unspecified speaker model creating unit of FIG. 1;

FIG. 4 is a diagram showing a process of clustering an output Gaussian distribution for each state in the unspecified speaker model creating process executed by the unspecified speaker model creating unit in FIG. 1;

5 is a diagram showing a process of mixing a plurality of probability density functions for each cluster in the unspecified speaker model creating process executed by the unspecified speaker model creating unit of FIG. 1. FIG.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Microphone, 2 ... Feature extraction part, 3 ... Buffer memory, 4 ... Phoneme matching part, 5 ... LR parser, 11 ... Hidden Markov network (HM network), 13 ... LR table, 20 ... Context-free grammar database, 30 ... Speech data of specific speaker, 31 ... Unspecified speaker model creation unit.

Claims (5)

[Claims] 1. An unspecified speaker model creating device for creating a hidden Markov model of a mixed Gaussian distribution of unspecified speakers, based on the input hidden Markov models of a single Gaussian distribution of specified speakers. Hidden Markov model of mixed Gaussian distributions of unspecified speakers by individually combining output Gaussian distributions of each state of input specific Gaussian distributions of specific speakers An unspecified speaker model creating apparatus comprising a model creating unit for creating a speaker. 2. The model creating means uses a predetermined learning method by using the same initial speaker hidden Markov model for a plurality of speakers, based on the input voice data of a plurality of specific speakers. Learning means for creating a hidden Markov model of a plurality of single Gaussian distributions for specific speakers by learning the output Gaussian distribution only for the state in which the voiced speech data exists, and the learning means Based on the Hidden Markov Model of multiple single-Gaussian distributions for specific speakers, based on the distance between each output Gaussian distribution, clustering into multiple clusters so that each cluster contains the output Gaussian distribution in a shorter distance Based on a single Gaussian hidden Markov model clustered for each state by the above clustering means. Then, combining means for synthesizing hidden Markov models with multiple output Gaussian distributions in each cluster into hidden Markov models with single Gaussian distributions for each state, and hiding the single Gaussian distributions for each state synthesized by the above combining means The unspecified speaker model creating device according to claim 1, further comprising: a mixing unit that creates a hidden Markov model of a Gaussian mixture mixture of unspecified speakers by mixing the Markov models. 3. The clustering means extracts the output Gaussian distribution learned with an amount of data equal to or more than a preset threshold value for each state, and then performs clustering. Specific speaker model creation device. 4. The clustering means repeats the clustering until the average value of the distances between the centers of the clusters clustered in each state and the output Gaussian distributions becomes equal to or less than a predetermined distance, thereby making each cluster in each state. 4. The apparatus for creating an unspecified speaker model according to claim 2, wherein the number of clusters in each state is determined so that the larger the variation of the output Gaussian distribution is, the larger the number of clusters is. 5. A hidden Markov model of a mixed Gaussian distribution of unspecified speakers is created on the basis of the inputted hidden Markov model of a single Gaussian distribution of specified speakers.
1 to 4 and a mixture of the unspecified speakers created by the unspecified speaker model creating device based on the voice signal of the input uttered voice sentence. A voice recognition device comprising: a voice recognition means for recognizing a voice using a hidden Markov model of distribution.