JP2003271179A - Method, device and program for speech recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognition method capable of performing speech recognition by speaker adaptation without needing a storage medium with large capacity for storing the acoustic model, etc., of a particular speaker and without having to identifying the particular speaker in advance, and to provide a device and a program therefor. <P>SOLUTION: This speech recognition device 1 using an acoustic model in which each phoneme of a speech signal to be an object of speech recognition is composed of a plurality of statistical distributions, a language model in which each inter-word relation is composed of a statistical distribution, and a word pronunciation dictionary for showing the relation between a word and a phoneme, is provided with a statistical distribution specification means 3a for specifying a statistical distribution with high output probability among a plurality of statistical distributions of the acoustic model on the basis of a result obtained by vocally recognizing speech uttered by a speaker desired to be adapted, and a speech recognition means 3b for recognizing the speech to be recognized on the basis of the statistical distribution specified by the statistical distribution specification means. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition method, a voice recognition device, and a voice recognition program which are adapted to a speaker generating a voice to be voice-recognized and recognized.

[0002]

2. Description of the Related Art Conventionally, as a method for performing speech recognition by so-called speaker adaptation, which is adapted to the tone of a speaker who emits a voice (habit of speaking), vocabulary often used, etc. There is something like this.

This is a method utilizing the MAP estimation method ("Spoken language processing", Kenji Kita et al., 65-68), which is a MAP estimation method (maximum posterior probability estimation method) when performing speaker adaptation of speech recognition. ) Is applied. In this method, the parameters of the acoustic model of the unspecified speaker are brought close to the parameters estimated by the utterance (speech data) of the speaker (specified speaker) to be adapted. By using this method, the acoustic model of the unspecified speaker can be adapted to the acoustic model of the specified speaker.

[0004]

However, in the conventional "method utilizing the MAP estimation method", it is necessary to prepare an acoustic model of a specific speaker for each speaker (specific speaker) to be adapted. When there are a plurality of speakers, the number of specific speaker acoustic models adapted from the unspecified speaker acoustic models increases as the number of specific speakers increases. Therefore, there is a problem that a large capacity of a storage medium such as a hard disk for storing the acoustic model of the specific speaker must be secured.

Further, when using the utterance data for each specific speaker, it is necessary to determine in advance who the specific speaker is, and there is a possibility that the determination of the speaker may be wrong. However, as a result, there is a problem that the voice recognition rate decreases.

Therefore, an object of the present invention is to solve the problems of the above-mentioned conventional techniques, and to utter each specific speaker without requiring a large-capacity storage medium for storing the acoustic model of the specific speaker. It is an object of the present invention to provide a voice recognition method, a voice recognition device, and a voice recognition program capable of performing voice recognition by speaker adaptation without distinguishing a specific speaker when using data.

[0007]

In order to achieve the above-mentioned object, the present invention has the following constitution. The speech recognition method according to claim 1, wherein each phoneme of the speech signal to be subjected to speech recognition is composed of a plurality of statistical distributions, and a language model in which the relationship between each word is expressed by probability. A speech recognition method using a pronunciation dictionary of words indicating the relationship between words and phonemes, wherein a plurality of statistical distributions of the acoustic model are based on the result of preliminarily speech recognition of a speech produced by a speaker to be adapted. From among the statistical distribution identification step of identifying a statistical distribution with a high output probability, and if the speaker utters,
A voice recognition step of recognizing a recognition target voice that is a recognition target based on the statistical distribution specified in the statistical distribution specifying step.

According to this method, first, in the statistical distribution specifying step, based on the voice recognition result of the voice uttered by the speaker who wants to be adapted, from among a plurality of statistical distributions of the acoustic model constituted by the statistical distribution. , A statistical distribution with a high output probability is specified. In this case, the voice uttered by the speaker is freely uttered by the intention of the speaker. A typical distribution is a normal distribution. Then, in the voice recognition step, the recognition target voice is recognized based on the statistical distribution specified from the plurality of statistical distributions. That is, in the statistical distribution specifying step, a statistical distribution having a high output probability (high usage frequency) is specified from one acoustic model targeting an unspecified speaker. That is, unlike the conventional case, it is not necessary to prepare a plurality of acoustic models for each speaker who wants to adapt. Then, in the voice recognition step, the recognition target voice of the specific speaker (the speaker to be adapted) is recognized based on the statistical distribution having a high output probability.

According to a second aspect of the present invention, there is provided a speech recognition method, in which each phoneme of a speech signal to be speech-recognized is composed of a plurality of statistical distributions, and a language in which the relation between each word is expressed by probability. A speech recognition method using a model and a pronunciation dictionary of words indicating a relationship between a word and a phoneme, in which voice content whose content is known in advance and voice produced by a speaker who wants to adapt this voice content Based on the correspondence result in which the, corresponding to, from among a plurality of statistical distribution of the acoustic model, a statistical distribution identification step of identifying a statistical distribution with a high output probability,
When the speaker utters, a voice recognition step of recognizing a recognition target voice that is a recognition target based on the statistical distribution specified in the statistical distribution specifying step is included.

According to this method, in the statistical distribution specifying step, the statistical distribution is calculated based on the correspondence result in which the utterance content whose content is known and the voice uttered by the speaker who wants to adapt the utterance content are associated with each other. A statistical distribution having a high output probability is specified from a plurality of statistical distributions of the acoustic model that is configured. Then, in the voice recognition step, the recognition target voice is recognized based on the specified statistical distribution. Note that the utterance content whose content is already known usually means that the utterance content is documented (text data).

According to another aspect of the speech recognition apparatus of the present invention, the acoustic model in which each phoneme of the speech signal to be recognized by speech is composed of a plurality of statistical distributions, and the relationship between each word is expressed by probability. A speech recognition device that uses a language model that exists and a pronunciation dictionary of words that indicate the relationship between a word and a phoneme, wherein the acoustic model is based on the result of preliminarily speech recognition of a voice uttered by a speaker to be adapted. From multiple statistical distributions of
When a speaker utters a statistical distribution specifying unit that specifies a statistical distribution having a high output probability, a voice recognition that recognizes a recognition target voice that is a recognition target based on the statistical distribution specified by the statistical distribution specifying unit. Means and are provided.

According to such a configuration, the statistical distribution specifying means outputs an output from a plurality of statistical distributions of the acoustic model formed by the statistical distribution based on the voice recognition result of the voice uttered by the speaker to be adapted. A statistical distribution with a high probability is identified.
In this case, the voice uttered by the speaker is freely uttered by the intention of the speaker. Then, the voice recognition means recognizes the recognition target voice based on the specified statistical distribution.

According to a fourth aspect of the present invention, there is provided a speech recognition apparatus in which a phoneme of a speech signal to be speech-recognized is composed of a plurality of statistical distributions and a language in which the relation between each word is expressed by probability. A speech recognition device that uses a model and a pronunciation dictionary of words indicating a relationship between a word and a phoneme, and a voicing content whose content is known in advance and a voice uttered by a speaker who wants to adapt this voicing content. Based on the correspondence result in which the acoustic model is associated, a statistical distribution identifying unit that identifies a statistical distribution having a high output probability from among a plurality of statistical distributions of the acoustic model, and the statistical distribution identifying unit when the speaker utters. Voice recognition means for recognizing a recognition target voice as a recognition target based on the statistical distribution specified in.

According to this structure, the statistical distribution specifying means calculates the statistical distribution based on the correspondence result in which the utterance content of which content is known and the voice uttered by the speaker who wants to adapt the utterance content are associated with each other. A statistical distribution having a high output probability is specified from a plurality of statistical distributions of the acoustic model that is configured.
Then, the voice recognition means recognizes the recognition target voice based on the specified statistical distribution.

A speech recognition program according to a fifth aspect of the present invention is a language in which each phoneme of a speech signal to be subjected to speech recognition is composed of a plurality of statistical distributions, and a relationship between each word is expressed by probability. A device using a model and a pronunciation dictionary of words indicating the relationship between words and phonemes is characterized by causing it to function as the following means. A means for operating the device is a statistical distribution that specifies a statistical distribution having a high output probability from among a plurality of statistical distributions of the acoustic model based on a result of previously recognizing a voice uttered by a speaker to be adapted. The specifying unit is a voice recognizing unit that recognizes a recognition target voice that is a recognition target based on the statistical distribution specified by the statistical distribution specifying unit when the speaker utters.

According to this configuration, the statistical distribution specifying means outputs the statistical distribution based on the speech recognition result of the voice uttered by the speaker to be adapted, from among the statistical distributions of the acoustic model constituted by the statistical distribution. A statistical distribution with a high probability is identified.
In this case, the voice uttered by the speaker is freely uttered by the intention of the speaker. Then, the voice recognition means recognizes the recognition target voice based on the specified statistical distribution.

According to a sixth aspect of the present invention, there is provided a speech recognition program, wherein an acoustic model in which each phoneme of a speech signal to be subjected to speech recognition is composed of a plurality of statistical distributions, and a language in which the relation between each word is expressed by probability. A device using a model and a pronunciation dictionary of words indicating the relationship between words and phonemes is characterized by causing it to function as the following means. The means for causing the device to function includes a plurality of statistical distributions of the acoustic model based on a correspondence result in which the utterance content whose content is already known and the voice uttered by the speaker who wants to adapt the utterance content are associated with each other. Among them, a statistical distribution specifying unit that specifies a statistical distribution having a high output probability, and when the speaker utters, recognizes a recognition target voice that is a recognition target based on the statistical distribution specified by the statistical distribution specifying unit. Voice recognition means for doing so.

According to this structure, the statistical distribution specifying means calculates the statistical distribution based on the correspondence result in which the utterance content of which the content is known and the voice uttered by the speaker who wants to adapt the utterance content are associated with each other. A statistical distribution having a high output probability is specified from a plurality of statistical distributions of the acoustic model that is configured.
Then, the voice recognition means recognizes the voice to be recognized based on the recorded statistical distribution.

[0019]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention will be described in detail below with reference to the drawings. In the description of this embodiment, the configuration of the voice recognition device will be described first, and then the operation of the voice recognition device will be described. Then, the waveform data at the time of voice recognition and the acoustic score will be supplementarily described.

(Structure of Speech Recognition Device) FIG. 1 is a block diagram of the speech recognition device. As shown in FIG. 1, the voice recognition device 1 includes a main control unit 3, a storage unit 5, a text input unit 7, a voice input unit 9, and a display output unit 11. The voice recognition device 1 is based on the voice recognition result of the voice uttered by the speaker in advance from the normal distribution of the acoustic model,
The normal distribution index with a high output probability value is recorded, and then the recorded normal distribution index is used to recognize the voice uttered by the speaker. In this embodiment, the voice recognition device 1 is composed of a general personal computer.

The main control section 3 is composed of a CPU, a main memory, etc., and controls the voice recognition apparatus 1, and comprises a speaker adaptation means 3a, a voice recognition means 3b, and a word string output means 3c. I have it. The speaker adaptation unit 3a stores the acoustic model stored in the storage unit 5 based on the raw voice data (voice data) input from the voice input unit 9 or the voice feature amount extracted from the raw voice data. Data 5a,
Speech recognition is performed using the language model data 5b and the pronunciation dictionary 5c, and the used acoustic model normal distribution index 5 is used.
a1 is generated.

That is, the speaker adaptation means 3a generates a normal distribution index of an acoustic model corresponding to a speaker (speaker to be adapted) who emits a voice to be voice-recognized, and uses this normal acoustic model distribution. Index 5a1
Is a normal distribution having a high output probability in the input raw voice data or voice feature amount.

If the utterance content is known (for example, if it is converted into text data in advance), the speaker adaptation means 3a outputs the utterance content and the raw input data from the voice input unit 9. Based on a correspondence result in which voice data (voice data) or voice feature amounts extracted from raw voice data are associated with each other, a used acoustic model normal distribution index 5a1 is generated for a high output probability. The speaker adaptation means 3a corresponds to the statistical distribution specifying means described in the claims.

The voice recognition means 3b is a speaker adaptation means 3a.
Acoustic model normal distribution index 5a generated in
1 is used to calculate the acoustic score and the language score (using the language model data 5b) of the recognition target voice that is the voice recognition target input from the voice input unit 9. The acoustic score is a numerical value calculated from the acoustic model in the hidden Markov model based on the comparison result of the waveform patterns of the recognition target voice input from the voice input unit 9. The language score is a numerical value of a language model that calculates the connection between words. The higher of these acoustic score and language score becomes the speech recognition result.

The word string output means 3c is a voice recognition means 3b.
Based on the score of the acoustic model and the score of the language model calculated in, the output word string which is the speech recognition result is transmitted. The storage unit 5 is composed of a large-capacity hard disk or the like, and stores acoustic model data 5a, language model data 5b, pronunciation dictionary 5c, adapted voice data 5d, recognition voice data 5e, and text data 5f. To do.

The acoustic model data 5a is a collection of acoustic model data in the hidden Markov model of an unspecified speaker composed of a plurality of normal distributions. The language model data 5b has a statistical distribution represented by n-gram. In the pronunciation dictionary 5c, the reading of this word is written for each word. The adapted voice data 5d is voice data input from the voice input unit 9 and uttered by the speaker to be adapted. The recognized voice data 5e is stored in the voice input unit 9
It is the voice data which is the target of the voice recognition input from. The text data 5f is text data of the utterance content input from the text input unit 7 when the utterance content of the voice data input from the adapted voice data 5d is known. The acoustic model data 5a corresponds to the acoustic model described in the claims, and the language model data 5b corresponds to the language model described in the claims.

Of these, the adapted voice data 5d, the recognized voice data 5e, and the text data 5f strictly indicate the storage area in the storage unit 5, and at the time of adaptation, that is, the speaker who wants to adapt emits a voice. Nothing is stored in the adapted voice data 5d or the text data 5f, and nothing is stored in the recognized voice data 5e at the time of recognition, that is, until the adapted speaker emits a recognition target voice. Not not. However, in the case of real-time voice recognition, it is not necessary to store data in the storage area of the recognized voice data 5e.

The text input unit 7 is composed of a keyboard, a mouse, etc., and the adapted voice data 5d of the storage unit 5 is used.
When the utterance content of the voice data to be stored in is known, the utterance content is input as text data of the utterance content, and is used for operating the voice recognition device 1.

The voice input unit 9 is composed of a speaker or the like, and inputs the voice (voice data) uttered by the speaker to be adapted and the voice uttered by the speaker during voice recognition. The display output unit 11 is configured by a display or the like, and displays an output word string that is a voice recognition result generated by the word string output unit 3c of the main control unit 3.

(Operation of Speech Recognition Device) Next, the operation of the speech recognition device 1 will be described with reference to FIG. In the operation of the voice recognition device 1, the speaker adaptation time and the voice recognition time are described as a series of operations. First, it is determined whether or not the text data of the voice uttered by the speaker who wants to be adapted is stored in the text data 5f of the storage unit 5 (S1).
When it is not determined that the text data is stored (S1, No), the operation of the voice recognition device 1 is as follows.
This corresponds to the operation in the case where the utterance content uttered by the speaker to be adapted is unknown (S2 to S6 described later: correspond to the method of claim 1). If it is determined that the text data is stored (S1, Yes), the operation corresponds to the case where the utterance content of the speaker to be adapted is known (S7, S3 to S6 described below: Claim). 2)).

As an operation at the time of adaptation when the utterance content is unknown, the voice (voice data) of the speaker to be adapted is input from the voice input unit 9 and stored in the adapted voice data 5d of the storage unit 5 ( S2). Then, the inputted voice (voice data) of the speaker to be adapted is recognized by using the acoustic model data 5a, the language model data 5b, and the pronunciation dictionary 5c. The speaker adaptation means 3 of the main control unit 3 is based on the speech recognition result (speech recognition result at the time of adaptation) that has been recognized.
From the acoustic model normal distribution data 5a stored in the storage unit 5 at a, the used acoustic model normal distribution index 5a1
Is generated (S3). If the input voice (voice data) is extremely small in the adaptation operation, the acoustic model data 5a is used as it is in the recognition operation.

Next, as a recognition operation, the recognition target voice (recognition target voice data) uttered by the adapted speaker is input from the voice input unit 9 and the recognition voice data 5e of the storage unit 5 is inputted.
(S4). Then, the voice recognition means 3b of the main controller 3 uses the acoustic model normal distribution index 5a.
1, the acoustic model data 5a, the language model data 5b, and the pronunciation dictionary 5c are used, and the acoustic score and language score of the recognition target speech are calculated based on these. The one with the highest acoustic score and language score (becomes a speech recognition result) is output to the word string output means 3c (S5). Then, the word string output means 3c sends the output word string, which is the speech recognition result, to the display output unit 11 (S).
6).

According to the operation of the voice recognition apparatus 1, the speaker adaptation means 3a produces a plurality of normal distributions of the acoustic model data 5a based on the result of the adaptive voice recognition of the voice uttered by the speaker to be adapted. Among these, the used acoustic model normal distribution index 5a1 having a high output probability is stored in the storage unit 5.
Then, the voice recognition means 3b recognizes the recognition target voice input from the voice input unit 9 by using the used acoustic model normal distribution index 5a1, the acoustic model data 5a, the language model data 5b, and the pronunciation dictionary 5c.

Therefore, since the used acoustic model normal distribution index 5a1 is generated from the acoustic model data 5a for each speaker to be adapted, an acoustic model is prepared and stored for each speaker to be adapted. There is no need, and the storage capacity of the storage unit 5 can be suppressed.

In addition, in the case of normal adaptation, when there are a plurality of speakers, it is necessary to collect the utterances of the speakers for each speaker. However, in the voice recognition device 1, the sentences uttered by a plurality of speakers are collectively treated as adapted data, and the used acoustic model normal distribution index 5a is used.
1 can be created, and voice recognition can be performed by using the used acoustic model normal distribution index 5a1 without determining the speaker during voice recognition.

Now, here, the text data 5 in the storage unit 5
The operation when it is determined that the text data of the voice uttered by the speaker adapted to f is stored (S1, YES), that is, the utterance content of the speaker who wants to be adapted is known will be described. .

As an operation at the time of adaptation, the voice data (voice waveform data) produced by the speaker to be adapted is converted into the voice input unit 9.
Is input from and is stored in the adapted voice data 5d of the storage unit 5 (S7). And the stored adapted voice data 5
A correspondence result in which d and the utterance content (text data 5f) are associated by the speaker adaptation means 3a is obtained. Based on the correspondence result, the speaker adaptation means 3a of the main control unit 3
Then, the used acoustic model normal distribution index 5a1 is generated from the acoustic model data 5a stored in the storage unit 5 (S8). After that, the operations S4 to S6 at the time of recognition are executed.

According to the operation of the voice recognition apparatus 1, the speaker adaptation means 3a produces a sound based on the correspondence result in which the waveform data of the voice uttered by the speaker to be adapted is associated with the text data 5f. Of the plurality of normal distributions of the model data 5a, the one with a high output probability is stored in the storage unit 5 as the used acoustic model normal distribution index 5a1. Then, the voice recognition means 3b recognizes the recognition target voice input from the voice input unit 9 by using the used acoustic model normal distribution index 5a1, the acoustic model data 5a, the language model data 5b, and the pronunciation dictionary 5c.

Therefore, since the used acoustic model normal distribution index 5a1 is generated from the acoustic model data 5a for each speaker to be adapted, an acoustic model is prepared and stored for each speaker to be adapted. There is no need, and the storage capacity of the storage unit 5 can be suppressed.

Further, in the case of normal adaptation, when there are a plurality of speakers, it is necessary to group the utterances of the speakers for each speaker. However, in the voice recognition device 1, the sentences uttered by a plurality of speakers are collectively treated as adapted data, and the used acoustic model normal distribution index 5a is used.
1 can be created, and voice recognition can be performed by using the used acoustic model normal distribution index 5a1 without determining the speaker during voice recognition.

(Regarding Speech Waveform and Phoneme Score Calculation During Speech Recognition) Next, the speech waveform and phoneme score calculation during speech recognition will be described with reference to FIG. 3 (see FIG. 1 as appropriate). In FIG. 3, the speech waveform data (the speech waveform data input from the speech input unit 9 at the time of speech recognition) which the speaker who wants to adapt utters "next" at the time of speech recognition, and the speech waveform data (at the time of adaptation) are adapted ( Prioritized)
It shows a relationship with the used acoustic model normal distribution index 5a1.

FIG. 3 (a) shows the relationship when the utterance content of the voice uttered by the speaker who wants to adapt during adaptation is unknown, and FIG. 3 (b) shows the speaker who wants to adapt during adaptation. The relationship is shown when the utterance content of the uttered voice is known. That is, in the case of FIG. 3A, what is compared with the voice waveform data input at the time of speech recognition is the acoustic feature of each phoneme that forms the phoneme sequence of the result of speech recognition at the time of adaptation (speech recognition result at adaptation). In the case of FIG. 3B, what is compared with the speech waveform data input at the time of speech recognition is the acoustic feature pattern of each phoneme forming a phoneme string (or word string) given in advance in the case of FIG. 3B. .

That is, which acoustic feature pattern was used for these phoneme strings during adaptation is the utilized acoustic model normal distribution index 5a1 (in the figure ...
)), And at the time of voice recognition, the acoustic feature pattern is compared with the place corresponding to each phoneme of the voice waveform data, and voice recognition is performed.

Although the present invention has been described based on the embodiment, the present invention is not limited to this. It is possible to regard each component of the voice recognition device 1 as a voice recognition method in which it is regarded as one process (step), and in this case, the same effect as that of the voice recognition device 1 can be obtained. Further, the processing of each component of the voice recognition device 1 can be regarded as a voice recognition program written in a general-purpose computer language, and in this case, the same effect as that of the voice recognition device 1 can be obtained.

[0045]

According to the inventions of claims 1, 3, and 5,
In the statistical distribution specifying step (statistical distribution specifying means), based on the voice recognition result of the voice uttered by the speaker who wants to be adapted, the statistics with a high output probability are selected from among the statistical distributions of the acoustic model configured by the statistical distribution. Since the distribution is specified and the voice is recognized in the voice recognition step (voice recognition means), it is not necessary to prepare and store an acoustic model for each speaker to be adapted, and the storage capacity of the storage medium can be suppressed. it can.

According to the present invention, the statistical distribution specifying step (statistical distribution specifying means) is based on the correspondence result in which the voice uttered by the speaker to be adapted is associated with the text data. Then, a statistical distribution with a high output probability is specified from among a plurality of statistical distributions of the acoustic model composed of statistical distributions, and the speech recognition step (speech recognition means) performs speech recognition. Since it is not necessary to prepare and store the acoustic model in the storage medium, the storage capacity of the storage medium can be suppressed.

[Brief description of drawings]

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

FIG. 2 is an operation of the voice recognition device shown in FIG.

FIG. 3 is an explanatory diagram illustrating a speech waveform and a phoneme score calculation during speech recognition.

[Explanation of symbols] 1 Speech recognition device 3 Main control section 3a Speaker adaptation means (statistical distribution identification means) 3b voice recognition means 3c word string output means 5 memory 5a Acoustic model data 5b Language model data 5c Pronunciation dictionary 7 Text input section 9 Voice input section 11 Display output section

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Kaho Onoe             1-10-11 Kinuta, Setagaya-ku, Tokyo, Japan             Broadcasting Association Broadcast Technology Institute F-term (reference) 5D015 GG04 HH23

Claims (6)

[Claims] 1. An acoustic model in which each phoneme of a speech signal to be subjected to speech recognition is composed of a plurality of statistical distributions, a language model in which a relationship between each word is expressed by probability, and a word and a phoneme. A speech recognition method using a pronunciation dictionary of words indicating a relationship, wherein based on a result of speech recognition of a voice uttered by a speaker to be adapted in advance, output is made from a plurality of statistical distributions of the acoustic model. A statistical distribution identification step of identifying a statistical distribution with a high probability, and a voice recognition step of recognizing a recognition target speech to be a recognition target based on the statistical distribution identified in the statistical distribution identification step when the speaker utters And a speech recognition method comprising: 2. An acoustic model in which each phoneme of a speech signal to be subjected to speech recognition is composed of a plurality of statistical distributions, a language model in which a relationship between each word is expressed by probability, and a word and a phoneme. A speech recognition method using a pronunciation dictionary of words indicating a relationship, which is based on a correspondence result in which the utterance content whose content is already known and the voice uttered by the speaker who wants to adapt the utterance content are associated with each other. Then, from among the plurality of statistical distributions of the acoustic model,
A statistical distribution specifying step of specifying a statistical distribution having a high output probability, and when the speaker utters, a voice recognition for recognizing a recognition target voice as a recognition target based on the statistical distribution specified in the statistical distribution specifying step. A voice recognition method, comprising: 3. An acoustic model in which each phoneme of a speech signal to be subjected to speech recognition is composed of a plurality of statistical distributions, a language model in which a relationship between each word is expressed by probability, and a word and a phoneme. A voice recognition device that uses a pronunciation dictionary of words indicating a relationship, and outputs from a plurality of statistical distributions of the acoustic model based on a result of voice recognition of a voice uttered by a speaker to be adapted in advance. A statistical distribution specifying unit that specifies a statistical distribution with a high probability, and a voice recognition unit that recognizes a recognition target voice that is a recognition target based on the statistical distribution specified by the statistical distribution specifying unit when the speaker utters. A voice recognition device comprising: 4. An acoustic model in which each phoneme of a speech signal to be subjected to speech recognition is composed of a plurality of statistical distributions, a language model in which a relationship between each word is expressed by probability, and a word and a phoneme. A voice recognition device that uses a pronunciation dictionary of words indicating a relationship, based on a correspondence result in which the utterance content whose content is already known and the voice uttered by the speaker who wants to adapt this utterance content are associated with each other. Then, from among the plurality of statistical distributions of the acoustic model,
A statistical distribution specifying unit that specifies a statistical distribution having a high output probability, and when the speaker utters, a voice recognition that recognizes a recognition target voice that is a recognition target based on the statistical distribution specified by the statistical distribution specifying unit. A voice recognition device comprising: 5. An acoustic model in which each phoneme of a speech signal to be subjected to speech recognition is composed of a plurality of statistical distributions, a language model in which the relationship between each word is expressed by probability, and a word and a phoneme. A device that uses a pronunciation dictionary of words indicating a relationship, based on the result of pre-voice recognition of a voice uttered by a speaker who wants to be adapted, from among a plurality of statistical distributions of the acoustic model, a statistic having a high output probability. Statistical distribution specifying means for specifying a distribution, when the speaker utters, based on the statistical distribution specified by the statistical distribution specifying means, to function as a voice recognition means for recognizing a recognition target voice to be a recognition target. A speech recognition program characterized by. 6. An acoustic model in which each phoneme of a speech signal to be speech-recognized is composed of a plurality of statistical distributions, a language model in which the relationship between each word is expressed by probability, and a word and a phoneme. A device using a pronunciation dictionary of words indicating a relationship is used in advance, based on a correspondence result in which a voicing content whose content is already known and a voice uttered by a speaker who wants to adapt the voicing content are associated with each other. From the multiple statistical distributions of the model,
Statistical distribution specifying means for specifying a statistical distribution having a high output probability, and when the speaker utters, a voice recognition means for recognizing a recognition target voice as a recognition target based on the statistical distribution specified by the statistical distribution specifying means. A speech recognition program characterized by causing it to function as.