JP2001125593A - Speech recognition method and device, and medium recorded with speech recognition program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech recognition device capable of reducing the calculation volume of likelihood calculation thereby increasing the processing speed. SOLUTION: The speech recognition device includes an auxiliary selection dictionary 4a generated from each partial space, a likelihood calculation part 3 which calculates an approximate likelihood of each distribution constituting the state of an HMM 'Hidden Markov Model) on the basis of an input vector and contents stored in the auxiliary selection dictionary 4a and calculates a likelihood of the input vector to a distribution which is extracted by using the approximate likelihood, and a dictionary search part 5 which performs voice recognition on the basis of the likelihood calculated by the calculation part 3. Since the likelihood calculation part 3 calculates the likehood of the input vector to only the distribution which is extracted by using the approximate likelihood, the calculation volume of likelihood calculation is reduced to increase the processing speed of speech recognition.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition technology, and more particularly, to a Hidden Markov Model (HMM).
Markov Model), a speech recognition method, and a medium recording a speech recognition program.

[0002]

2. Description of the Related Art In recent years, in information processing apparatuses such as personal computers and word processors, voice recognition apparatuses have been actively developed in order to enable input of sentences by voice. Among voice recognitions, voice recognition using an HMM, which can obtain high recognition accuracy even with a change in spectrum itself due to individual differences of speakers, has been actively researched and developed.

[0003] In speech recognition using HMM, a model of statistical features of speech obtained from a large amount of speech data is used. The speech recognition using the HMM is described in detail in Seiichi Nakagawa, “Speech Recognition by Stochastic Model,” and the like, so please refer to these reference books.

[0004]

In the development of a speech recognition apparatus using an HMM, many aim at recognition of large vocabulary, unspecified speakers and continuous speech, and speech which can be processed in real time only by software processing. In order to construct a recognition device, it is necessary to speed up speech recognition processing.

[0005] In a speech recognition apparatus using a continuous distribution HMM, the calculation of the output probability of the HMM is a process requiring a large amount of calculation as in the case of language space search. The calculation of the output probability of the HMM is based on the total number of distributions (= (number of HMMs) × (HMM
(The number of states in the state) × (the number of mixed distributions in the state)). Generally, in speech recognition using a continuous distribution HMM, if the learning data is sufficient, the larger the number of distributions, the higher the recognition rate but the amount of calculation. Conversely, the smaller the number of distributions, the lower the recognition rate but the amount of calculation. Therefore, it is important to reduce the amount of calculation by reducing the total number of distributions actually calculated without lowering the recognition rate.

A first object of the present invention is to provide a speech recognition apparatus, a speech recognition method, and a medium in which a speech recognition program is recorded, which can reduce the amount of calculation of likelihood calculation and improve the processing speed. That is.

A second object is to provide a speech recognition device, a speech recognition method, and a medium on which a speech recognition program is recorded, having high recognition performance.

[0008]

According to one aspect of the present invention, a speech recognition apparatus is provided based on a preselection dictionary created for each subspace, an input vector, and contents stored in the preselection dictionary. A likelihood calculating means for calculating the approximate likelihood of each distribution constituting the state of the HMM, and calculating the likelihood of the distribution selected using the approximate likelihood with the input vector;
Recognizing means for recognizing speech based on the likelihood calculated by the likelihood calculating means.

Since the likelihood calculating means calculates the likelihood with the input vector only for the distribution selected using the approximate likelihood, the amount of calculation of the likelihood calculation can be reduced, and the processing speed of speech recognition can be reduced. Can be improved.

[0010] According to another aspect of the present invention, a preselection dictionary is created for each feature space. Since the preliminary selection dictionary is created for each feature space, more detailed approximate likelihood can be calculated, and a speech recognition device with high recognition performance can be provided.

According to still another aspect of the present invention, the preliminary selection dictionary expresses a cluster representative distribution created for each subspace by a normal distribution, and stores an average value and a variance.

In the preliminary selection dictionary, since the cluster representative distribution created for each subspace is represented by a normal distribution, it is easy to calculate the approximate likelihood, and it is possible to further improve the processing speed.

According to still another aspect of the present invention, the preselection dictionary includes a multi-layer dictionary created for each subspace.

For example, the pre-selection dictionary includes a first-layer dictionary and a second-layer dictionary, calculates an approximate likelihood using the first-layer dictionary, and further uses the second-layer dictionary. By calculating the approximate likelihood in detail, the amount of calculation of the approximate likelihood can be reduced.

According to still another aspect of the present invention, the likelihood calculating means includes a detailed dictionary in which the output probability distribution of each distribution obtained by learning the speech data is stored, an input vector and a preselection dictionary. An approximate likelihood calculating means for calculating an approximate likelihood of each distribution constituting the HMM based on the stored contents, and a detailed likelihood based on the approximate likelihood calculated by the approximate likelihood calculating means. A distribution selecting means for selecting a distribution to be calculated, and a detailed likelihood calculating means for calculating a detailed likelihood of the distribution selected by the distribution selecting means using a detailed dictionary.

Since the detailed likelihood calculating means calculates the detailed likelihood of the distribution selected by the distribution selecting means using the detailed dictionary, it is possible to reduce the number of distributions for calculating the detailed likelihood, and to reduce the likelihood of speech recognition. Processing speed can be improved.

According to still another aspect of the present invention, the preliminary selection dictionary is created by classifying each distribution stored in the detailed dictionary for each feature space.

Since the preselection dictionary is created by classifying each distribution stored in the detailed dictionary for each feature space, a detailed approximate likelihood can be calculated, and a speech recognition apparatus with high recognition performance is provided. It is possible to do.

According to still another aspect of the present invention, the pre-selection dictionary classifies which cluster the audio data belongs to after each distribution stored in the detailed dictionary is classified for each subspace. It is created by calculating from the value of the feature parameter of the audio data belonging to the cluster.

The pre-selection dictionary classifies which cluster the audio data belongs to after each distribution stored in the detailed dictionary is classified for each subspace, and determines from the value of the characteristic parameter of the audio data belonging to each cluster. Since it is calculated and created, each distribution can be classified more accurately, and a speech recognition device with high recognition performance can be provided.

According to still another aspect of the present invention, the distribution selecting means selects a distribution having a high likelihood from the approximate likelihoods calculated by the approximate likelihood calculating means.

The distribution selecting means selects a distribution having an approximate likelihood having a large likelihood from the approximate likelihoods calculated by the approximate likelihood calculating means, so that a distribution requiring detailed likelihood calculation can be accurately determined. It becomes possible to select.

According to yet another aspect of the present invention, the distribution selecting means calculates a reference likelihood from the maximum likelihood of the subspace and a threshold, and selects a distribution having a higher approximate likelihood than the reference likelihood.

The distribution selecting means calculates a reference likelihood from the maximum likelihood of the subspace and a threshold value and selects a distribution having an approximate likelihood larger than the reference likelihood. It becomes possible to select accurately.

According to still another aspect of the present invention, a speech recognition method calculates an approximate likelihood of each distribution constituting an HMM based on an input vector and contents of a preselection dictionary created for each subspace. And calculating the likelihood of the distribution selected using the approximate likelihood with the input vector, and performing speech recognition based on the calculated likelihood.

Since the likelihood with the input vector is calculated only for the distribution selected using the approximate likelihood, the calculation amount of the likelihood calculation can be reduced, and the processing speed of speech recognition can be improved. Becomes

According to still another aspect of the present invention, a speech recognition program recorded on a computer-readable medium forms an HMM based on an input vector and contents of a preliminary selection dictionary created for each subspace. Calculating the approximate likelihood of each distribution to be calculated, calculating the likelihood of the distribution selected using the approximate likelihood with the input vector, and performing speech recognition based on the calculated likelihood. And

Since the likelihood with the input vector is calculated only for the distribution selected using the approximate likelihood, the calculation amount of the likelihood calculation can be reduced, and the processing speed of speech recognition can be improved. Becomes

[0029]

FIG. 1 is a block diagram showing a functional configuration of a speech recognition apparatus according to an embodiment of the present invention.
This speech recognition device converts an analog speech signal collected by a microphone into a digital signal, cuts out a speech section to be subjected to speech recognition, and outputs the speech section.
An audio analysis unit 2 that analyzes a digital audio signal output from the audio input unit 1 to extract and output an audio parameter (input vector), and an HMM based on the audio parameter output from the audio analysis unit 2. A likelihood calculation unit 3 for calculating the likelihood of each distribution constituting the state of the above, an HMM storage unit 4 for storing an HMM used for likelihood calculation by the likelihood calculation unit 3, a word dictionary 6, A dictionary search unit 5 that searches the word dictionary 6 based on the likelihood calculated by the likelihood calculation unit 3 and outputs a recognition result, and a display unit 7 that displays the recognition result output from the dictionary search unit 5 Including.

The input unit 1 includes a microphone 1a to which a voice is input by a user, and an A / A converter for converting a voice signal input through the microphone 1a from an analog signal to a digital signal.
D (Analog / Digital) converter 1b.

The HMM storage unit 4 stores a detailed dictionary unit 4b for storing each distribution constituting the state of the HMM learned from a large amount of voice data, and stores each distribution of the voice data stored in the detailed dictionary 4b for each feature space. And a preliminary selection dictionary 4a for storing a cluster representative distribution created by clustering. H
Learning of the MM is performed by the Baum-Welch algorithm. The details of the preliminary selection dictionary 4a and the detailed dictionary 4b will be described later.

The likelihood calculating unit 3 calculates the likelihood by comparing the acoustic parameters extracted by the acoustic analyzing unit 2 and the cluster representative distribution stored in the preliminary selection dictionary 4a, and calculates the index information for each distribution. An approximate likelihood calculation unit 3a that calculates the approximate likelihood of each distribution using the distribution, and a predetermined number of distributions from the top of the approximate likelihood calculated by the approximate likelihood calculation unit 3a, or larger than a predetermined reference likelihood. It includes a distribution selection unit 3b that selects a distribution of approximate likelihood, and a detailed likelihood calculation unit 3c that recalculates the likelihood of the distribution selected by the distribution selection unit 3b using the detailed dictionary 4b.

The dictionary search unit 5 calculates a score using the Viterbi algorithm for the likelihood of each distribution calculated by the likelihood calculation unit 3 and the words registered in the word dictionary 6, and calculates the calculated score. Is output as a recognition result.

In the above description, the speech feature vector series is represented by the HMM, but a standard pattern represented by a time series of frame vectors may be used instead of the HMM. When performing recognition using this standard pattern, acoustic parameters and H
Instead of the likelihood calculation unit 3 that calculates the likelihood by comparing each distribution constituting the state of the MM and calculates the likelihood, a configuration may be adopted in which the distance between the acoustic parameter and the frame vector of the standard pattern is calculated and evaluated. . The dictionary search unit 5 calculates a score using DP (Dynamic Programming) matching for the distance between the frame vectors of the standard pattern and the words registered in the word dictionary 6, and the calculated score is minimized. What is necessary is just to employ | adopt the structure which outputs a word as a recognition result.

FIG. 2 is a flowchart for explaining the processing procedure of the speech recognition apparatus according to the present embodiment.
First, when the user inputs voice through the microphone 1a,
This analog audio signal is converted into a digital signal by the A / D converter 1b (S11). The acoustic analysis unit 2 extracts an acoustic parameter (input vector) that is a feature of the audio signal using linear prediction analysis or the like (S12).
In the present embodiment, as the acoustic parameters, LPC (Linear Predictive Coding) cepstrum 1st to 16th order, LPCΔ cepstrum 1st to 16th order
Next, four types of characteristics, power and Δpower, are used.

Next, the approximate likelihood calculating section 3a compares the acoustic parameters extracted by the acoustic analyzing section 2 with the preliminary selection dictionary 4a created for each feature, and creates a subspace-based likelihood table. . Then, the approximate likelihood of each distribution is calculated by referring to the likelihood table for each subspace using the distribution number in the index table and the index information for the cluster for each feature (S13).

Here, a method of creating the preliminary selection dictionary 4a, the index table, and the likelihood table for each subspace will be described.

FIG. 3 is a diagram showing an example of the contents of the preliminary selection dictionary 4a, and FIG. 4 is a diagram showing an example of the index table and the likelihood table for each subspace. FIG. 5 is a flowchart for explaining a method of creating the preliminary selection dictionary 4a. First, the HMM stored in the detailed dictionary 4b
Are classified into M clusters for each subspace (S21). As a method of classifying each feature of the distribution into M clusters, for example, K-means
You can use the law.

Next, an index table indicating the relationship between each distribution and the cluster is created from the result of the clustering (S22). This index table is
It shows which cluster (1 to M) each feature of each distribution is classified into. For example, in FIG. 4, the distribution power of the distribution number “1” is classified into the cluster with the cluster number “2”, the LPC cepstrum is classified into the cluster with the cluster number “1”, and the Δ power is the cluster with the cluster number “1”. This indicates that the LPCΔ cepstrum is classified into the cluster with the cluster number “2”.

Next, a large amount of voice data (voice data used for generating the detailed dictionary 4b) is compared with each distribution in the detailed dictionary 4b, and the cluster to which the voice data belongs is classified again. (S23). That is, by recognizing the audio data using the contents of the detailed dictionary 4b and taking a matching path, a distribution with the highest likelihood is obtained for the input vector, and each cluster belongs to any cluster with reference to the index table. Or classify.

Then, the average value and the variance are calculated from the values of the acoustic parameters of the audio data belonging to each cluster to create a cluster representative distribution (S24). The preliminary selection dictionary shown in FIG. 3 indicates that each created cluster representative distribution is stored for each feature. Also, in order to easily create the preliminary selection dictionary 4a, step S2
The average value and the variance may be calculated from the result of the clustering created in 1 and may be used as a cluster representative distribution.

Although FIG. 3 shows a case where only one layer of the preliminary selection dictionary for each subspace is created, a multi-layer structure as shown in FIG. 6 may be used. In this case, for example, FIG.
Is assumed to be the second layer, and M2 cluster representative distributions of the second layer are clustered for each subspace to create M1 cluster representative distributions.
The correspondence with the layer is stored in the index table. With the pre-selection dictionary having such a configuration, the likelihood with each cluster representative distribution of the first layer is calculated based on the acoustic parameters of the recognition target, and a cluster having a high likelihood is selected, and this cluster indicates. Recalculate the likelihood with each cluster representative distribution of the second layer, and use the likelihood of the cluster representative distribution of the first layer as the approximate likelihood for the cluster of the second layer indicated by the unselected cluster. Thereby, the calculation amount of likelihood calculation can be reduced.

Next, the calculation of the likelihood table for each subspace shown in FIG. 4 and the extraction of the distribution for which the detailed likelihood needs to be calculated will be described with reference to FIGS. 7 (a) and 7 (b). . As shown in FIG. 7A, first, the approximate likelihood calculating unit 3a compares the acoustic parameter to be recognized with the distribution of cluster representatives stored in the preliminary selection dictionary 4a for each subspace to determine the likelihood. The calculation is performed, and a likelihood table for each subspace is created as shown in FIG. 4 (S31).

Next, the approximate likelihood calculating section 3a extracts the likelihood of each feature in the likelihood table for each subspace by referring to the index in the order of distribution number of the index table, and adds them to calculate the approximate likelihood. Calculate the degree (S3
2). As shown in FIG. 4, since the content of the likelihood table for each subspace is represented by a logarithm, the approximate likelihood can be calculated only by addition.

Next, the distribution selecting section 3b sorts the approximate likelihood of each distribution calculated by the approximate likelihood calculating section 3a, and extracts the top T distributions having a large approximate likelihood (S
33). Then, the detailed likelihood calculation unit 3c includes the distribution selection unit 3
b, an output probability distribution corresponding to the top T distributions extracted from the detailed dictionary 4b, and comparing the acoustic parameter to be recognized with the output probability distribution corresponding to the top T distributions to obtain a detailed likelihood. The calculation is performed (S34).

As another method, as shown in FIG. 7 (b), first, the approximate likelihood calculating unit 3a calculates the acoustic parameters to be recognized and the cluster representative stored in the preliminary selection dictionary 4a for each subspace. The likelihood is calculated by comparing with the distribution, and a likelihood table for each subspace is created as shown in FIG. 4 (S4).
1).

Next, the approximate likelihood calculation unit 3a extracts the likelihood of each feature in the likelihood table for each subspace by referring to the index in the order of distribution number of the index table, and adds them to calculate the approximate likelihood. Calculate the degree (S4
2).

Next, the distribution selection unit 3b calculates the maximum likelihood of each subspace, and calculates a reference likelihood (reference likelihood <maximum likelihood) from the maximum likelihood of each subspace and a predetermined threshold value. ) Is calculated. Then, the distribution selection unit 3b selects a distribution having an approximate likelihood larger than the reference likelihood (S43). Then, the detailed likelihood calculating unit 3c extracts an output probability distribution corresponding to the distribution selected by the distribution selecting unit 3b from the detailed dictionary 4b, and outputs an acoustic parameter to be recognized and an output probability distribution corresponding to the extracted distribution. To calculate the detailed likelihood (S
44).

Returning to the description of the flowchart shown in FIG. As described above, the distribution is selected with reference to the preliminary selection dictionary 4a and the detailed dictionary 4b (S14), and the detailed likelihood of the selected distribution is calculated (S15).

Next, the dictionary search unit 5 calculates a score using the Viterbi algorithm for the likelihood of each distribution calculated in step S15 and the model for each word registered in the word dictionary 6 ( S16) The word having the largest score is displayed on the display unit 7 as a recognition result (S17).

For example, the number of distributions N is 1500, the number of clusters M in each subspace is 50, and the number of distributions T for calculating detailed likelihood is T
Is 100, the likelihood calculation for a distribution of 50 (number of clusters M) +100 (distribution T) = 150, calculation of the approximate likelihood of each distribution, and selection of the top 100 distributions are required. Among them, the calculation of the approximate likelihood of each distribution and the top one
The time required to select the 00 distributions is short compared to the likelihood calculation. Therefore, in speech recognition using the conventional HMM, likelihood calculation corresponding to the number of distributions N = 1500 is required, whereas in the speech recognition apparatus according to the present embodiment, likelihood calculation for 150 distributions is not possible. This is necessary, and the likelihood calculation can be performed with a calculation amount of about 1/10 as compared with the related art.

When the likelihood calculation is performed using the pre-selection dictionary created by clustering in all the feature spaces, the amount of calculation required for the likelihood calculation is calculated by using the pre-selection dictionary in the present embodiment. Same as quantity. However, when a pre-selection dictionary created by clustering in all feature spaces is used, only M possible approximate likelihoods can be expressed, whereas in the pre-selection dictionary of the present embodiment, clustering is performed for each feature space. Therefore, M × M × M × M approximate likelihoods can be expressed. Therefore, more detailed approximate likelihood can be calculated, and the speech recognition performance is improved.

In the speech recognition apparatus according to the present embodiment, power, Δpower, LPC
Although four are used, the cepstrum and the LPCΔ cepstrum, other feature spaces may be used. Further, in each feature space, the number of clusters in each subspace has been described as being the same, but the number of clusters in different subspaces may be set for each feature space.

FIG. 8 is a diagram showing an example of the appearance of the speech recognition apparatus of the present invention. This voice recognition device includes a voice input unit 1, a computer main body 11, a graphic display device 1,
2, a magnetic tape device 13 on which a magnetic tape 14 is mounted;
Keyboard 15, mouse 16, CD-ROM (Compact
Disc-Read Only Memory (CD-R) 18
It includes an OM device 17 and a communication modem 19. The voice recognition program is supplied by a recording medium such as the magnetic tape 14 or the CD-ROM 18. The voice recognition program is executed by the computer main body 11, and an operator looks at the graphic display device 12 while watching the keyboard 1.
By operating the mouse 5 or the mouse 16, an instruction for voice recognition or the like is given. Further, the voice recognition program may be supplied to the computer main body 11 from another computer via a communication line and a communication modem 19.

FIG. 9 is a block diagram showing a configuration example of the speech recognition apparatus of the present invention. Computer body 1 shown in FIG.
1 is a CPU (Central Processing Unit) 20, R
OM (Read Only Memory) 21, RAM (Random Acc
ess Memory) 22 and a hard disk 23.
The CPU 20 includes a graphic display device 12, a magnetic tape device 13, a keyboard 15, a mouse 16, and a CD.
ROM device 17, communication modem 19, ROM 21, RA
The processing is performed while inputting / outputting data to / from the M22 or the hard disk 23. Magnetic tape 14 or CD-R
The voice recognition program recorded in the OM 18 is
0, the magnetic tape device 13 or the CD-ROM device 1
7 is temporarily stored in the hard disk 23. CP
The U20 performs voice recognition by appropriately loading a voice recognition program from the hard disk 23 into the RAM 22 and executing the program.

As described above, according to the speech recognition apparatus of the present embodiment, the approximate likelihood of each distribution is calculated using the preliminary selection dictionary 4a created for each subspace.
Since the distributions for which the output probabilities need to be calculated exactly are extracted and the output probabilities are calculated only for the extracted distributions,
It has become possible to reduce the time required for recognition. In addition, clustering is performed for each feature space, and a preliminary selection dictionary 4 is created.
Since a is created, the approximate likelihood calculating unit 3a can calculate a detailed approximate likelihood, and can improve recognition performance.

The embodiments disclosed this time are to be considered in all respects as illustrative and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating an outline of a functional configuration of a speech recognition device according to an embodiment of the present invention.

FIG. 2 is a flowchart illustrating a processing procedure of the voice recognition device according to the embodiment of the present invention.

FIG. 3 is a diagram illustrating an example of a preliminary selection dictionary 4a.

FIG. 4 is a diagram illustrating an example of an index table and a likelihood table for each subspace.

FIG. 5 is a flowchart for explaining creation of a preliminary selection dictionary 4a and an index table for each subspace.

FIG. 6 is a diagram showing a case where the preliminary selection dictionary 4a has a multilayer structure.

FIG. 7 is a flowchart for explaining creation of a likelihood table for each subspace and calculation of detailed likelihood.

FIG. 8 is a diagram illustrating an example of an appearance of a speech recognition device according to an embodiment of the present invention.

FIG. 9 is a diagram illustrating a schematic configuration of a speech recognition device according to an embodiment of the present invention.

[Explanation of symbols]

Reference Signs List 1 voice input unit, 1a microphone, 1b A / D converter, 2 acoustic analysis unit, 3 likelihood calculation unit, 3a approximate likelihood calculation unit, 3b distribution selection unit, 3c detailed likelihood calculation unit, 4
HMM storage unit, 4a preliminary selection dictionary, 4b detailed dictionary, 5 dictionary search unit, 6 word dictionary, 7 display unit, 11
Computer body, 12 Graphic display device, 13 Magnetic tape device, 14 Magnetic tape, 15
Keyboard, 16 mouse, 17 CD-ROM device, 18 CD-ROM, 19 communication modem, 20 C
PU, 21 ROM, 22 RAM, 23 hard disk.

Claims (11)

[Claims] An approximate likelihood of each distribution constituting a state of an HMM is calculated based on a preselection dictionary created for each subspace, an input vector and contents stored in the preselection dictionary, Likelihood calculating means for calculating the likelihood of the distribution selected using the approximate likelihood with the input vector; and recognition means for recognizing speech based on the likelihood calculated by the likelihood calculating means. And a voice recognition device. 2. The speech recognition device according to claim 1, wherein the preliminary selection dictionary is created for each feature space. 3. The speech recognition apparatus according to claim 1, wherein the preliminary selection dictionary expresses a cluster representative distribution created for each subspace as a normal distribution, and stores an average value and a variance. 4. The speech recognition device according to claim 1, wherein said preliminary selection dictionary includes a multi-layer dictionary created for each subspace. 5. The likelihood calculating means includes: a detailed dictionary in which an output probability distribution of each distribution obtained by learning speech data is stored; and the input vector and contents stored in the preliminary selection dictionary. Approximate likelihood calculating means for calculating the approximate likelihood of each distribution constituting the state of the HMM based on the approximate likelihood calculated by the approximate likelihood calculating means. 5. A distribution selecting unit for selecting a distribution, and a detailed likelihood calculating unit for calculating a detailed likelihood of the distribution selected by the distribution selecting unit by using the detailed dictionary. A voice recognition device according to any one of the claims. 6. The preliminary selection dictionary is created by classifying each distribution stored in the detailed dictionary for each feature space.
The speech recognition device according to claim 5. 7. The pre-selection dictionary, after each distribution stored in the detailed dictionary is classified for each subspace, classifies which cluster the audio data belongs to, The speech recognition device according to claim 5, wherein the speech recognition device is created by calculating from a value of a feature parameter. 8. The distribution selecting unit according to claim 5, wherein the distribution selecting unit selects a distribution having an approximate likelihood having a large likelihood from the approximate likelihoods calculated by the approximate likelihood calculating unit. The speech recognition device according to the above. 9. The method according to claim 5, wherein the distribution selecting means calculates a reference likelihood from the maximum likelihood of the subspace and a threshold, and selects a distribution having an approximate likelihood larger than the reference likelihood. The speech recognition device according to any one of the above. 10. calculating the approximate likelihood of each distribution constituting the state of the HMM based on the input vector and the contents of the preliminary selection dictionary created for each subspace; and selecting the approximate likelihood using the approximate likelihood. A speech recognition method comprising: calculating a likelihood of the obtained distribution with the input vector; and performing speech recognition based on the calculated likelihood. 11. A step of calculating an approximate likelihood of each distribution constituting the state of the HMM based on an input vector and contents of a preselection dictionary created for each subspace, and selecting the approximate likelihood using the approximate likelihood. A computer-readable recording medium recording a speech recognition program, comprising: calculating a likelihood of the obtained distribution with the input vector; and performing speech recognition based on the calculated likelihood.