JP2852210B2 - Unspecified speaker model creation device and speech recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an unspecified speaker model creating apparatus for creating a hidden Markov model (hereinafter, referred to as an HMM) of an unspecified speaker based on a plurality of hidden Markov models of a specific speaker. 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) has been used to create an HMM of an unspecified speaker based on a specific speaker model for learning. ) Are widely used (eg, Seiichi Nakagawa, “Speech Recognition by Stochastic Model”, p.
p. 55-64, IEICE, July 1988. ). In the first conventional example, the partial observation sequence {y ₁ , y ₂ , y ₃ ,
, Y _t }, 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 in state i at time t Using the backward probability of observing _{t + 3} ,..., y _r }, the HMM of the unspecified speaker is created by re-estimating and learning the parameters of the HMM.

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 fluctuations in acoustic features 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 from a large number of speakers. However, when dealing with such a large amount of data,
The enormous amount of calculation has become a problem even today, as the processing speed of computers is increasing.

In order to reduce the amount of calculation of such an unspecified speaker model, Kosaka et al. Have already developed a speaker clustering based on a specific speaker model and a CCL method (hereinafter referred to as a second conventional example) based on model synthesis. Proposal (Conventional Document 2, "Kosaka et al.," Method of Creating Unspecified Speaker Model Using Clustering Method ", Transactions of the Acoustical Society of Japan, 1-R-1
2, November 1994 ". ). In the method of the second conventional example, clustering is performed in units of the entire model set over all phonemes, on the assumption that the similarity of acoustic features of the speech of each speaker is equal in all acoustic spaces. I have. Specifically, after a specific speaker model that has been sufficiently learned is clustered by defining a distance between the models, an unspecific speaker model is created by combining the specific speaker models.

[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 specific speaker model are sufficiently learned. Otherwise, a high-performance model cannot be obtained, so that a lot of utterance data is required for each speaker. In addition, the mixed number of the mixed output Gaussian distribution always becomes the same in all the states of the HMM, and there is a problem that a useless parameter increases in a state in which the variation of the feature amount by the speaker is small.

A first object of the present invention is to solve the above problems, and it is not necessary for all 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 generating apparatus which can generate an unspecified speaker model even if the processing is performed, and which requires a small memory capacity of the processing device and can shorten the calculation time. A second object of the present invention, in addition to the first object, is that speech recognition can be performed using the created speaker-independent model, and the speech recognition rate is improved as compared with the conventional example. It is an object of the present invention to provide a voice recognition device capable of performing the above.

[0007]

According to a first aspect of the present invention, there is provided an unspecified speaker model generating apparatus for generating an unspecified speaker model based on a single Gaussian distribution hidden Markov model of a plurality of specified speakers. In an unspecified speaker model creating apparatus for creating a hidden Markov model of a mixed Gaussian distribution of speakers, an output Gaussian distribution of each state of a hidden Markov model of a single Gaussian distribution of a plurality of specific speakers is input for each state. And a model creating means for creating a hidden Markov model of a Gaussian mixture distribution of unspecified speakers by independently clustering and combining.

According to a second aspect of the present invention, there is provided the unspecified speaker model creating apparatus according to the first aspect, wherein the model creating means includes a plurality of input uttered voices of the specific speakers. By learning the output Gaussian distribution only for the state where the uttered voice data exists by a predetermined learning method using the same initial speaker hidden Markov model for a plurality of speakers based on the data, Learning means for creating a single speaker specific Gaussian distribution hidden Markov model; anda plurality of output Gaussian models based on a plurality of specific speaker single Gaussian distribution hidden Markov models created by the learning means. Clustering means for performing clustering on a plurality of clusters based on the distance between distributions such that the output Gaussian distribution is included in a shorter distance for each cluster; The hidden Markov model of a plurality of output Gaussian distributions in each cluster is synthesized with the hidden Markov model of a single Gaussian distribution of each state based on the hidden Markov model of the single Gaussian distribution clustered for each state by the taling means. Combining means for creating a hidden Markov model of a mixed Gaussian distribution of an unspecified speaker by mixing the hidden Markov model of a single Gaussian distribution of each state synthesized by the combining means. It is characterized by.

Further, according to a third aspect of the present invention, in the unspecified speaker model generating apparatus according to the second aspect, the clustering means includes a predetermined threshold value or more for each state. After extracting only the output Gaussian distribution learned with the data amount, clustering is performed.

Further, in the apparatus for creating an unspecified speaker model according to claim 4, in the apparatus for creating an unspecified speaker model according to claim 2 or 3, the clustering means includes: By repeating clustering until the average value of the distance between the center and each output Gaussian distribution becomes equal to or less than a predetermined distance, the clusters in each state are increased so that the variation in each output Gaussian distribution in each state is larger and the number of clusters is larger. The number is determined.

Further, according to a fifth aspect of the present invention, there is provided a speech recognition apparatus, comprising the steps of: hiding a mixed Gaussian distribution of an unspecified speaker based on a single Gaussian distribution of a single Gaussian distribution of a plurality of specific speakers; An unspecified speaker model creating apparatus according to any one of claims 1 to 4, which creates a Markov model.
Voice recognition means for voice recognition using a hidden Markov model of a mixture distribution of unspecified speakers created by the unspecified speaker model creating apparatus based on the speech signal of the input uttered speech sentence. It is characterized by the following.

[0012]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a speech recognition device according to one embodiment of the present invention. In particular, the speech recognition apparatus of the present embodiment uses a known maximum likelihood estimation method based on utterance voice data of a plurality of N specific speakers stored in the memory of the utterance voice data 30 of the specific 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. To generate a hidden Markov network (hereinafter referred to as an HM network) having a Gaussian mixture, and store the created HM network in a memory of the HM network 11. The speech recognition is performed with reference to the stored HM network.

This speech 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 a speech recognition process 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 A parser (hereinafter referred to as LR parser) 5.

FIG. 2 is a flowchart showing an unspecified speaker model creation process executed by the unspecified speaker model creation unit 31. In the creation process, first, in step S1, feature parameters of the uttered voice data are extracted based on the uttered voice data of a plurality of N specific speakers, and a plurality of N Mean value and variance of the output Gaussian distribution only for the state where data exists by the known maximum likelihood estimation method using the initial speaker model (each state 1 mixture) which is the same HM network for all speakers To create an HM network of N single speaker-specific Gaussian distributions.

Next, in step 2, as shown in FIG. 3, the HMs of the N single Gaussian distributions for the specific speaker are created.
After extracting only the output Gaussian distributions learned with a data amount equal to or larger than a preset threshold value for each state based on the network, as shown in FIG. 4, a known Bhattacharyya between the output Gaussian distributions is obtained. Based on the distance, clustering is performed on a plurality of clusters such that the output Gaussian distribution is included in a shorter distance for each cluster.
Here, the threshold value is set for the amount of learning data to be extracted in order to prevent the output Gaussian distribution having low reliability from affecting the clustering. Thereby, a highly reliable HM network 11 can be obtained, and the HM network 11 can be obtained.
By performing voice recognition using, the voice recognition can be performed at a higher voice recognition rate than the conventional example. In addition, in the clustering, clustering is performed in each state, and clustering is repeated until the average value of a known Bhattacharyya distance between the center of each cluster and each output Gaussian distribution becomes equal to or less than a predetermined distance, and thereby, in each state. 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 relatively large, while when the variation is small, the number K of clusters is set relatively small. In determining the number of clusters K, a maximum number of clusters Kmax and a minimum number of clusters Kmin may be set. Furthermore, when the amount of learning data is small, the number of clusters K is preferably set small.

Next, in step S3, a plurality of output Gaussian distributions in the cluster are converted into a single Gaussian distribution of each state as shown in FIG. 5 by using the result clustered for each state in step S2. Combine. The synthesis was performed in the same manner 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 combining method in step S3 will be described later in detail. Further, in step S4, an HM network of a mixed Gaussian distribution is created by mixing the synthesized single Gaussian distribution of all clusters for each state using a known speaker mixing method, and the HM network 11 Store in memory. The mixture ratio was a value proportional to the ratio of the sum of the learning data amounts of the output Gaussian distribution of the members of each cluster. That is, the larger the learning data amount of the members of each cluster, the larger the mixture ratio is set.

The average value μh _j and the variance Sh _j of each cluster used in step S3 are represented 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 respectively represent an average value and a variance when a plurality of Gaussian distributions are determined as a single Gaussian distribution. Here, μ _j ⁽ⁱ⁾ and S _j ⁽ⁱ⁾ are the natural number i-th HM
Represents the mean and variance of the output probability density function, which is a single Gaussian distribution at network state j. N _j ⁽ⁱ⁾ represents the number of samples in the state j of the ith 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 weighted with a larger weighting factor w _j ⁽ⁱ⁾ as the number of samples n _j ⁽ⁱ⁾ increases in accordance with 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 expressed efficiently. One HM network includes many phoneme environment dependent models. The HM network 11 is composed of a combination of states including a Gaussian distribution, and states are shared between individual phoneme environment-dependent models. Therefore, even when the number of data for parameter estimation is insufficient, a robust model can be created. This HM network 11 uses a successive state division method (Successive Sta
te Splitting: Hereinafter, referred to as SSS. ) Automatically. The above SSS determines the topology of the HM network,
Determination of abnormal noise clusters and estimation of Gaussian distribution parameters in each state are performed simultaneously. In the present embodiment, the parameters of the HM network include an output probability and a transition probability expressed by a Gaussian distribution. Therefore, at the time of recognition, it can be handled in the same way as a general HMM.

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

In FIG. 1, a uttered voice of a speaker is input to a microphone 1 and converted into a voice signal, and then input to a feature extracting unit 2. After performing A / D conversion on the input audio signal, the feature extraction unit 2 performs, for example, LPC analysis, and performs 34-dimensional feature parameters including logarithmic power, 16th-order cepstrum coefficient, Δlogarithmic power, and 16th-order Δcepstrum coefficient. Is extracted. The time series of the extracted feature parameters is input to the phoneme matching unit 4 via the buffer memory 3.

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

The phoneme matching unit 4 executes phoneme matching processing 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 the 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 likelihood calculation uses the forward path algorithm used in the normal HMM as it is.

On the other hand, a predetermined context-free grammar (C
FG) The LR table is created by automatically converting the database 20 as is well known 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. If there is syntactic ambiguity, the stack is split and the analysis of all candidates is processed in parallel. The LR parser 5 has an LR table 13
, And outputs phoneme prediction data to the phoneme matching unit 4. In response, the phoneme matching unit 4 performs matching by referring to 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, continuous speech recognition is performed, and the speech 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 uttered voice data 30 of the specific speaker, the HM network 11, and the LR table 13
And the context-free grammar database 20 are stored in, for example, a hard disk memory. Further, the phoneme matching unit 4, the LR parser 5, and the speaker-independent model creation unit 31 are configured by, for example, a digital computer.

In the above embodiment, the unspecified speaker model is created by the unspecified speaker model creation processing of FIG. 2, but a known balm-welch is applied to the HM network created by the creation processing. HM network may be created by re-learning using the learning algorithm described in (1).

[0028]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present inventor conducted the following experiment in order to confirm the effectiveness of the speech recognition apparatus shown in FIG. In the experiment, an HM network that effectively shared the state of a context-dependent phoneme HMM (for example, conventional literature 3 “Takami et al.,“ Automatic Generation of Hidden Markov Network by Sequential State Partitioning with Phoneme Context and Time ”), Telecommunications Information Technology Research Report, SP91-88, December 1991 "). The structure of the HM network is one speaker uttering 2620
The decision was made using the voice data of the words, and two types of models having a total number of states of 200 and 600 were created. 1 for each model
A silence model with state 10 mixed was added. The initial speaker model for the specific speaker model learning was a single distribution for each state except the silence model, and the initial values of the parameters were determined using the same speech data as the structure determination. Based on this initial speaker model, a speech recognition database of a natural utterance having a task of travel planning, which is owned by the present applicant (for example, see 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 the variance of the output Gaussian distribution by the maximum likelihood estimation method using the natural utterance data of 81 men in the above. However, since the data amount per person was as small as about 20 utterances, the variance was updated only when the value was larger than the initial parameter.
In this case, an uncharacteristic speaker model was created and a recognition experiment was performed using only male speakers. The recognition experiment was performed on nine men who were not included in the learning data selected from the same natural utterance database as that used for learning.

The speaker-independent model is created by the CCL method of the second conventional example using model-based clustering in units of the entire HM network and the method using the state-based clustering result of the HMM according to the present invention. Were compared by phoneme recognition experiments. However, according to the state-based clustering method according to the present invention, the state occupation data amount at the time of learning is 10 out of the output Gaussian distribution of each state of the specific speaker model.
Only frames and more were used. Furthermore, the model created by state-based clustering was compared with the recognition rate of the model retrained by the Balm-Welch learning algorithm using the model as an initial model. In addition,
The HMM using the state-based clustering method according to the present invention
Then, experiments were performed on the recognition rate of the model re-learned by the Balm-Welch learning algorithm. Table 1 shows the analysis conditions, the parameters used, 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 Period = 5 ms ───────────────────────────────── Parameters used 16th order LPC cepstrum + 16th order cepstrum + logarithmic power + Δ Logarithmic power 学習 Learning data 81 men-1 conversation for each speaker (total 1799 (Utterance) ───────────────────────────────── Unspecified speaker model evaluation data 9 males --- each speaker 1 conversation (11-29 utterances)

Tables 2 and 3 show the CCL method of the second conventional example (hereinafter abbreviated as model clustering in the tables) and the method by state-based clustering according to the present invention (hereinafter referred to as state-based clustering in the tables). ), The total number of output Gaussian distributions included in the HM network of each state and number of mixtures. In the case of the CCL method of the second conventional example, the number of mixture distributions is equal 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. Done 10
If the number of output Gaussian distributions trained with data of more than frames is less than the set number of mixtures, the number of distributions extracted is the number of mixture distributions in that state, so the total number of distributions is smaller than with model-based clustering ing.
However, this time, the degree of variation of the average value of the extracted output Gaussian distribution in each state is not considered in determining the number of mixtures. When using a free speech database where it is difficult to collect speech data considering phoneme balance, it is possible to prevent the increase of unnecessary parameters by designing the number of mixture distributions for each state. You can see that there is.

[0032]

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

[0033]

[Table 3] Total number of distributions of the unspecified speaker model-In the case of the HM network in the 601 state ────── Creation method / mixing 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 phoneme recognition experiments using an unspecified speaker model created by each method. The results in the table show the average values for nine men.

[0035]

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

[0036]

Table 5 Comparison of phoneme recognition rate (%) by model creation method-In case of HM network in 601 state ───────── Creation method / mixing number 3 5 10 15 ──────────────────────────────── ─── Balm Welch 67.6 67.8--─────────────────────────────────── Model Clustering 65.1 65.5 66.2 66.2 別 By state Clustering 67.8 67.9 67.8 67.8 別 By state Clustering 69.2 69.2 − − + balm-we Ji ───────────────────────────────────

Looking at the results in Tables 4 and 5 together with the results in Tables 2 and 3, the method using the state-based clustering according to the present invention under all conditions provides the CCL of the second conventional example.
The high recognition performance is shown with a smaller number of parameters than in the case of the method.
Considering the speed of actual recognition processing and the efficiency of speaker adaptation, the performance as an unspecified speaker model is considered to be better if high recognition performance is obtained with as few parameters as possible. It is shown that the state-based clustering method according to the present invention is an effective method for obtaining a high-performance model.

When looking at the relationship between the number of states of the HM network and the recognition performance, the HM network in the 601 state shows higher recognition performance than the HM network in the 201 state, which is the same as the CCL of the second conventional example. The same can be said for both the method and the state-based clustering method according to the present invention. This is 201
It is considered that the cause is that the phonological environment has not been sufficiently segmented and modeled yet. If the phonological environment is not subdivided enough to be subdivided,
The output Gaussian distribution of each state must simultaneously represent the variation of acoustic features due to both the phonological environment and the speaker environment, making it difficult to distinguish between phonological and speakeric, and the possibility of recognition errors. Is thought to be higher.

Further, as is apparent from Tables 4 and 5, when clustering is performed by the state-based clustering method according to the present invention and then re-learned by using the Balm-Welch learning algorithm, compared to other methods, A high phoneme recognition rate has been obtained.

Lastly, the time for creating the speaker-independent model will be described. It is reported that the second conventional CCL method disclosed in Conventional Document 2 requires only a few percent of the calculation time of the Balm-Welch learning algorithm. When the state-based clustering according to the present invention is used, the number of times of performing the clustering increases,
Although the calculation time increases as compared with the conventional CCL method of
Since this time is very small as compared with the learning time of the specific speaker model that occupies most of the time required for model creation, the CCL of the second conventional example is viewed in the whole time.
Similar to the method, a speaker-independent model can be created in a calculation time of about several percent of the learning algorithm of Balm-Welch.

As described above, according to the embodiment of the present invention, the input Gaussian distribution of each state of a single Gaussian HMM of a plurality of specific speakers is clustered independently for each state. The HMM of the mixed Gaussian distribution of unspecified speakers is created by combining with each other, so that all the parameters of each specific speaker model do not need to be learned, and the parameters learned for each speaker are different. You can also respond to the situation. Therefore, the present invention can be used for voice data of a speaker with a small number of utterances or data with different utterance contents for each speaker, such as free speech. Furthermore, the number of clusters to be divided for each state can be determined by using information on the variation of the average value of the output Gaussian distribution extracted from each specific speaker model and the amount of learning data for each state of the HMM. The number of mixture distributions can be determined for each state of the HMM in consideration of the amount of learning data and the degree of variation in acoustic characteristics between speakers. By performing voice recognition using the HMM of the unspecified speaker model, voice recognition can be performed at a higher voice recognition rate than in the conventional example.

[0042]

As described above in detail, according to the unspecified speaker model generating apparatus according to the first aspect of the present invention, based on the input hidden Markov model of a single Gaussian distribution of a plurality of specific speakers. In an unspecified speaker model generating apparatus for generating a hidden Markov model of a mixed Gaussian distribution of unspecified speakers, an output Gaussian distribution of each state of a single Gaussian hidden Markov model of a plurality of input specific speakers is input. Are clustered independently for each state and synthesized to create a hidden Markov model of a mixed Gaussian distribution of unspecified speakers. Specifically, the model creating means performs a predetermined learning method using the same initial speaker hidden Markov model for a plurality of speakers based on the input utterance voice data of the plurality of specific speakers. Learning means for creating a hidden Markov model of a plurality of single Gaussian distributions for a specific speaker by learning an output Gaussian distribution only for a state in which the uttered voice data exists, and learning means created by the learning means. Clustering into multiple clusters based on the distance between each output Gaussian distribution based on the hidden Markov model of multiple single speaker specific Gaussian distributions such that each cluster contains the output Gaussian distribution at a shorter distance Based on a hidden Markov model of a single Gaussian distribution clustered for each state by the clustering means. Combining means for combining a plurality of output Gaussian hidden Markov models in each cluster into a single Gaussian hidden Markov model for each state; and a single Gaussian distribution Hidden Markov model for each state synthesized by the combining means. And a mixing means for generating a hidden Markov model of a mixed Gaussian distribution of an unspecified speaker by mixing

That is, by extracting only the output Gaussian distribution learned from a number of specific speaker models and performing independent clustering in each state of the HMM, the magnitude of the variation of the feature amount and the amount of training data in each state are reduced. The number of clusters can be determined in consideration of this, and the optimal number of output Gaussian distributions can be determined for each state. Also, since only the output Gaussian distribution of each specific speaker model that has been learned can be selectively used, it is not necessary that all output Gaussian distributions of each specific speaker model have been learned, and the utterance per person It can be used effectively even for small databases. Also, since the parameter estimation is performed separately for each speaker, the amount of calculation is drastically reduced as compared with the first conventional method using the Balm-Welch learning algorithm in which learning is performed using all data at once. Becomes possible. Therefore, it is possible to greatly reduce the time for creating the unspecified speaker model.

According to a third aspect of the present invention, the clustering means extracts only an output Gaussian distribution learned with a data amount equal to or larger than a predetermined threshold value for each state. And then clustering. As a result, an optimal speaker-independent model with higher reliability can be created. 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 a fourth aspect of the present invention, the clustering means determines in advance the average value of the distance between the center of each cluster and the output Gaussian distribution in each state. By repeating the clustering until the distance becomes equal to or less than the set distance, the number of clusters in each state is determined such that the larger the variation of each output Gaussian distribution in each state, the larger the number of clusters. 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 distributions can be determined for each state. As a result, an optimal speaker-independent model with higher reliability can be created. 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 specific speakers. 5. The unspecified speaker model creating apparatus according to claim 1, wherein the unidentified speaker model creating apparatus creates an Hidden Markov Model based on the speech signal of an input uttered voice sentence. Using a hidden Markov model of a mixture distribution of unspecified speakers created by the above method. 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 the 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 the unspecified speaker model creation unit of FIG. 1;

FIG. 3 is a diagram illustrating a process of learning a specific speaker model and extracting an output Gaussian distribution in an unspecified speaker model generating process performed by the unspecified speaker model generating unit in FIG. 1;

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

FIG. 5 is a diagram illustrating a process of mixing a plurality of probability density functions for each cluster in an unspecified speaker model creation process performed by the unspecified speaker model creation unit of FIG. 1;

[Explanation of symbols]

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

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-4-125599 (JP, A) JP-A-63-257798 (JP, A) Proceedings of the Acoustical Society of Japan (September 1995) 3-2 -9, p. 123-124 (58) Field surveyed (Int. Cl. ⁶ , DB name) G10L 3/00 521 G10L 3/00 531 G10L 3/00 535 JICST file (JOIS)

Claims (5)

(57) [Claims] 1. An unspecified speaker model creating apparatus for creating a hidden Markov model of a mixed Gaussian distribution of unspecified speakers based on an input hidden Markov model of a single Gaussian distribution of a plurality of specific speakers, Hidden Markov model of mixed Gaussian distribution of unspecified speakers by clustering and combining output Gaussian distributions of each state of each state of Hidden Markov model of single Gaussian distribution of multiple specific speakers independently for each state An unspecified speaker model creation device, characterized by comprising a model creation means for creating a speaker model. 2. The model creating means according to a predetermined learning method using the same initial speaker hidden Markov model for a plurality of speakers based on the input utterance voice data of a plurality of specific speakers. A learning means for creating a hidden Markov model of a plurality of single Gaussian distributions for a specific speaker by learning an output Gaussian distribution only for a state in which the uttered voice data exists; Clustering into multiple clusters based on the distance between each output Gaussian distribution based on the hidden Markov model of multiple single speaker specific Gaussian distributions such that each cluster contains the output Gaussian distribution at a shorter distance Based on a Hidden Markov Model with a single Gaussian distribution clustered for each state by the clustering means. Combining means for combining a hidden Markov model of a plurality of output Gaussian distributions in each cluster into a hidden Markov model of a single Gaussian distribution of each state; and hiding a single Gaussian distribution of each state synthesized by the combining means. 2. The unspecified speaker model creating apparatus according to claim 1, further comprising mixing means for creating a hidden Markov model having a mixed Gaussian distribution of unspecified speakers by mixing the Markov models. 3. The clustering device according to claim 2, wherein the clustering means extracts only an output Gaussian distribution learned with a data amount equal to or larger than a predetermined threshold value for each state, and then performs clustering. Specific speaker model creation device. 4. The clustering means repeats clustering until the average value of the distance between the center of each cluster that has been clustered in each state and each output Gaussian distribution is equal to or less than a predetermined distance. 4. The speaker-independent model generation apparatus according to claim 2, wherein the number of clusters in each state is determined such that the number of clusters increases as the variation of the output Gaussian distribution increases. 5. A hidden Markov model of a mixed Gaussian distribution of unspecified speakers is created based on the input Hidden Markov Model of a single Gaussian distribution of a plurality of specific speakers.
And an unspecified speaker model created by the unspecified speaker model creating apparatus based on the input speech signal of the uttered speech sentence. A speech recognition apparatus, comprising: speech recognition means for performing speech recognition using a hidden Markov model of distribution.