JP2005221752A - Speech recognition apparatus, speech recognition method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To obtain a speech recognition result with high robustness while improving meaning adaptability. <P>SOLUTION: A 1st decoder 71 computes N or more most likelihood solutions and their likelihoods (score) by using a speech feature quantity extracted by a feature quantity extraction part 2, a phoneme HMM stored in a sound model storage part 3, and a word 2-gram model stored in a language model storage part 4. The maximum likelihood solution of highest order among the plurality of maximum likelihood solutions is selected by using a concept expression pattern model stored in a concept expression pattern model storage part 5 and a concept appearance pattern model stored in a concept appearance pattern model storage part 6 and regarded as a final recognition result. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a speech recognition device, a speech recognition method, and a program, and more particularly, to a speech recognition device, a speech recognition method, and a program that improve semantic adaptability.

  Commonly used speech recognition devices include not only acoustic models that statistically model speech features, but also language models that statistically model local concatenation of words as language constraints (for example, words n-gram model etc.) is used. However, using only the language constraints of the language model may result in recognition results that are syntactically or semantically ineligible as a whole.

Therefore, the top N maximum likelihood recognition candidates obtained from the conventional speech recognition apparatus are subjected to syntax analysis using a context-free grammar prepared in advance as grammar knowledge to determine grammatical suitability. In addition, a speech recognition apparatus has been proposed that improves recognition accuracy by preventing recognition candidates that are grammatically ineligible from being output (see, for example, Patent Document 1).
In addition, a language model generation device that generates a language model that accurately and analytically shows a sentence structure and obtains a high recognition rate at the time of recognition, and a speech recognition device that performs speech recognition using this language model have been proposed ( For example, see Patent Document 2.)

  In Patent Literature 2, a word (or word class) N-gram model is generated from learning text data in which a big marker is inserted using a conversion-driven machine translation device (TDML device). This N-gram model has local sentence structure constraints. Furthermore, a word pattern model of a word pattern concatenation relation consisting of words included in one subtree and a plurality of connected subtrees from a syntax tree obtained by parsing learning text data into which a bigram marker is inserted. Generate. The word pattern model models a relationship between words at a long distance that are related to a partial word string.

The technique described in Patent Document 2 cannot be modeled by using not only the conventional local word chain constraint but also the local sentence structure constraint and the conventional word N-gram model by using these models. Incorporating constraints between long-distance words, we aim to improve speech recognition accuracy.
JP 2000-293196 A JP 2002-268676 A

  Since the spoken language (spoken language) is not grammatically stricter than the originally written language, the user (speaker) does not always speak a grammatically adapted sentence. It is almost impossible to create a grammar that includes such grammatically fluctuating spoken words.

  On the other hand, the speech recognition apparatus of Patent Document 1 may reject recognition results that are grammatically ineligible. For this reason, if the speaker originally made a grammatically ineligible utterance, the speech recognition device, even if it was correctly recognized at the initial stage, the recognition result at the grammatical suitability determination stage. May be rejected.

  Furthermore, the above speech recognition apparatus may be correct if it is syntactically correct even if it is semantically unsuitable only by syntactic analysis using a context-free grammar. For example, the sentence “I want to go to Naka-ku in Chikusa-ku” is not relevant in terms of meaning, but is syntactically correct, so it may be recognized as a correct answer.

  In Patent Document 2, if a grammatically ineligible utterance is included in the learning text data, it is considered that a grammatically fluctuating spoken word is robust to some extent. However, since the recognition result candidate includes word replacement, insertion, omission, and the like due to misrecognition, syntax analysis is not always possible.

  Therefore, since the technique described in Patent Document 2 needs to perform syntax analysis and obtain a syntax tree, it cannot be said that the robustness with respect to recognition result candidates including many misrecognitions is high. In addition, similar to Patent Document 1, there are cases where a recognition result candidate that is syntactically correct but semantically ineligible is regarded as a correct answer.

  The present invention has been proposed to solve the above-described problems, and an object of the present invention is to provide a speech recognition apparatus, a speech recognition method, and a program with improved semantic adaptability and high robustness.

  The present invention relates to a feature quantity extracting means for extracting a feature quantity of a speech signal, an acoustic model storing means for storing an acoustic model obtained by statistically modeling the feature quantity, and a language model that statistically models a chain of words. Extracted by the feature model extracting means, a concept appearance pattern model storing means for storing a concept appearance pattern model having a plurality of appearance patterns representing a combination order of one or more concepts, and a feature quantity extracting means Based on the feature amount, the acoustic model stored in the acoustic model storage unit, the language model stored in the language model storage unit, and the concept appearance pattern model stored in the concept appearance pattern model, Recognition means for recognizing the voice represented by the voice signal.

  The feature amount extraction unit extracts a feature amount obtained as a result of speech analysis for the speech signal. The acoustic model storage unit stores an acoustic model obtained by statistically modeling the feature amount. The acoustic model is not limited to phoneme units, but may be subword units such as syllables or mora, for example.

  The language model storage means stores a language model obtained by statistically modeling a word chain. The “word” may be a phoneme, a word or a morpheme, or a combination thereof. In other words, the language model is a model of linguistic constraints.

  The concept appearance pattern model storage means stores a concept appearance pattern model. The concept appearance pattern model represents a plurality of appearance patterns representing the combination order of one or more concepts. Note that the order of combination of a plurality of concepts has semantic constraints, that is, has semantic suitability. Therefore, the appearance pattern does not matter how many words are present between the concepts, but always has a semantic adaptability to the combination order of the concepts. Each appearance pattern may be associated with a score.

  The recognizing unit recognizes the voice represented by the audio signal by using the acoustic model, the language model, and the concept appearance pattern model for the feature amount extracted by the feature amount extracting unit.

  Therefore, according to the invention, the voice represented by the voice signal based on the concept appearance pattern model having a plurality of appearance patterns representing the feature amount of the voice signal, the acoustic model, the language model, and the combination order of one or more concepts. By recognizing, the semantic suitability of speech recognition can be improved.

  In the present invention, the recognition means includes a concept appearance pattern model in which a score is associated with each appearance pattern for a plurality of recognition result candidates based on the feature quantity, the acoustic model, and the language model, and their scores. May be used to weight each score of the plurality of recognition result candidates, and the recognition result candidate of the highest score among the weighted scores may be output.

  The invention further includes an expression pattern model storage unit that stores an expression pattern model representing a plurality of expression patterns for each concept, and the recognition unit further uses the expression pattern model stored in the expression pattern model storage unit. The voice may be recognized. Here, the expression pattern includes a word representing a predetermined concept or a peripheral partial word string. By including not only a word representing a predetermined concept but also a partial word string around it, robustness can be enhanced without being affected by word replacement or omission due to misrecognition.

  Furthermore, in the invention, the recognition means includes a conceptual expression in which a score is associated with each expression pattern for a plurality of recognition result candidates based on the feature quantity, the acoustic model, and the language model, and their scores. By further using a pattern model, each score of the plurality of recognition result candidates may be weighted.

  According to the speech recognition apparatus, the speech recognition method, and the program according to the present invention, a concept appearance pattern model having a plurality of appearance patterns representing a combination of a feature amount of an audio signal, an acoustic model, a language model, and one or more concepts. By recognizing the voice represented by the voice signal based on the above, it is possible to improve the semantic suitability of the voice recognition.

  The principle of the present invention will be described first, and then the best mode for carrying out the present invention will be described in detail with reference to the drawings.

[Principle of the present invention]
The problem of speech recognition can be regarded as a problem of obtaining a word string W that maximizes (Equation 1).

However,
W: Word strings W ₁ , W ₂ , W ₃ ,..., W _M
O: Speech feature sequence O ₁ , O ₂ , O ₃ ,..., O _N
It is.

  That is, the word that maximizes the product of the appearance probability P (W) of the word sequence W obtained from the language model and the appearance probability P (O | W) of the speech feature amount sequence in the word sequence W obtained from the acoustic model. The problem is to obtain the column W.

According to the present invention, if the partial word sequence w _i ^{i + K} in the word sequence W is a predetermined concept C _j with respect to (Equation 1), the score Score (w _i ⁱ obtained from the concept expression pattern model is determined. ^{+ K} | C _j ) is added, and scores Score (C _1j , C _2j ,..., C _Lj ) obtained from the concept appearance pattern model based on the appearance pattern (concept array) in the word string W are added. That's it. Here, Score (w _i ^{i + K} | C _j ) is a score obtained based on the frequency of occurrence of the partial word string w _i ^{i + K} in the concept C _j . Score (C _1j , C _2j ,..., C _Lj ) (C _ij is any one of a plurality of concepts) is a score obtained based on the frequency at which the concepts C _ij appear simultaneously in one utterance. It is.

  Therefore, the problem of speech recognition according to the present invention can be regarded as a problem that maximizes (Equation 2).

  In addition, in (Formula 2), addition and multiplication are used for convenience, but the present invention is not limited to this. That is, according to the present invention, if the weight of the conventional recognition likelihood using only the acoustic model and the language model can be weighted using the score of the concept expression pattern model and the score of the concept appearance pattern model, the above addition is performed. In addition to subtraction and multiplication, subtraction or division may be used.

[Embodiment]
FIG. 1 is a block diagram showing a configuration of a speech recognition apparatus according to an embodiment of the present invention. In this embodiment, a voice recognition apparatus suitable for use in a destination task dialogue system will be described as an example. That is, the speech recognition apparatus according to the present embodiment inputs speech including at least one of “city name”, “address”, “store name”, and “business type” necessary for setting a destination. It recognizes the earth.

  The voice recognition apparatus includes a microphone 1 that generates a voice signal according to a speaker's utterance, a feature quantity extraction unit 2 that extracts a feature quantity (voice feature quantity) of the voice signal, and an acoustic model storage that stores an acoustic model. Unit 3, a language model storage unit 4 for storing language models, a concept expression pattern model storage unit 5 for storing concept expression pattern models, a concept appearance pattern model storage unit 6 for storing concept appearance pattern models, and speech features A decoder 7 that performs speech recognition using the quantity and each model, and an output unit 8 that outputs a recognition result to a predetermined system are provided.

  The microphone 1 generates a voice signal corresponding to the voice uttered by the speaker and supplies the voice signal to the feature amount extraction unit 2. The feature quantity extraction unit 2 analyzes the voice signal supplied from the microphone 1, cuts out a voice section divided by silence, and extracts a voice feature quantity.

  The acoustic model storage unit 3 stores, for example, a phoneme hidden Markov model (HMM) as an acoustic model. The phoneme HMM is a model that statistically represents a speech feature amount for each phoneme unit such as “A” and “I”. It may exist in units of syllables and words depending on the usage.

  The language model storage unit 4 stores, for example, a word 2-gram (bigram) model as a language model obtained by statistically modeling a word chain. The word 2-gram model is a language model in which the occurrence probability of a word depends only on the immediately preceding word, and is learned by learning text data, for example.

  The concept expression pattern model storage unit 5 stores a concept expression pattern model representing an expression pattern including a word that is a key for understanding the meaning / intention of an utterance or a peripheral partial word string.

  In the destination task dialogue system according to the present embodiment, it is important to identify words representing concepts such as the address, the store name, and the type of business being uttered. For example, “soba shop” included in the concept of “industry” has various expression patterns such as “soba shop”, “soba shop”, and “soba shop”. Similarly, the “ramen restaurant” included in the concept of “industry” has various expression patterns such as “ramen store”. Therefore, the concept expression pattern model is obtained by learning a plurality of expression patterns representing a certain concept from learning text data and modeling each expression pattern in the concept.

  FIG. 2 is a diagram illustrating the concept expression pattern model stored in the concept expression pattern model storage unit 5. In the present embodiment, the concept expression pattern model includes “expression pattern” in each concept of “city name (city)”, “detailed address (address)”, “store name (name)”, and “business type (type)”, and It is a model that represents the “score”.

  The “expression pattern” is, for example, manually extracted from the text data for learning used when learning the word 2-gram model. For example, as shown in FIG. 2, “concept” of “type of industry” includes “expression patterns” such as buckwheat shop, buckwheat shop, buckwheat shop, restaurant, restaurant, ramen shop, ramen shop, etc. There is. Each “expression pattern” is associated with a “score”.

  The “score” is a value based on the appearance probability of the “expression pattern” in the “concept”. For example, the score “0.152” of a restaurant is a value based on the appearance probability of “restaurant” within the concept of “business type”. Similarly, for other concepts, a plurality of expression patterns are prepared for one concept, and a score is associated with each expression pattern.

  The concept appearance pattern model storage unit 6 stores a concept appearance pattern model that models a statistical tendency of appearance patterns between concepts when one or more concepts appear in one utterance (one recognition unit). Has been.

  FIG. 3 is a diagram showing a concept expression pattern model stored in the concept appearance pattern model storage unit 6. The concept appearance pattern model is a model representing an “appearance pattern” that is an arbitrary combination order of the above four concepts and its “score”.

  The “appearance pattern” is created from text obtained by replacing the “expression pattern” appearing in the learning text data with a concept symbol (eg, [city] or [type]) indicating the corresponding “concept”. is there. For this reason, the “appearance pattern” has semantic restrictions in the order of the “concept” even if there are many meaningless words between the “concepts”. Each “appearance pattern” is associated with a “score”. The “score” is a value based on the concept co-occurrence probability.

  The decoder 7 includes a first decoder 71 that outputs a plurality of recognition result candidates, and a second decoder 72 that outputs a final recognition result from among the recognition result candidates.

  The first decoder 71 uses the speech feature amount extracted by the feature amount extraction unit 2, the phoneme HMM stored in the acoustic model storage unit 3, and the word 2-gram model stored in the language model storage unit 4. A plurality of (for example, N or more) maximum likelihood solutions and their likelihoods (scores) are calculated. The maximum likelihood solution is a so-called recognition result candidate in which acoustic models are connected and networked according to a language model. Note that the likelihood of the maximum likelihood solution obtained by the first decoder 71 (likelihood parameter L described later) is expressed by, for example, (Equation 3). Each parameter is the same as in (Equation 1).

  The second decoder 72 uses the concept expression pattern model stored in the concept expression pattern model storage unit 5 and the concept appearance pattern model stored in the concept appearance pattern model storage unit 6 to select from the plurality of maximum likelihood solutions. The highest likelihood solution is selected and this maximum likelihood solution is used as the final recognition result.

  FIG. 4 is a flowchart showing a processing procedure for the second decoder 72 to obtain a final recognition result.

  In step ST1, the second decoder 72 sets the top N most likely solutions and their scores in the maximum likelihood solution matrix C among the plurality of maximum likelihood solutions obtained by the first decoder 71. Then, after i = 1, the process proceeds to step ST2. Note that i is a natural number from 1 to N.

  In step ST2, the second decoder 72 extracts the i-th maximum likelihood solution from the maximum likelihood matrix C, sets this maximum likelihood solution as the maximum likelihood parameter R, and sets the likelihood (score) to the likelihood. The parameter L is copied and the process proceeds to step ST3. In the process of step ST3, the i-th maximum likelihood solution will be described using the following example.

  FIG. 5A is a diagram showing the i-th maximum likelihood solution in text display, FIG. 5B is a diagram for explaining processing when a conceptual expression pattern model is used for the i-th maximum likelihood solution, and FIG. It is a figure explaining a process when a concept appearance pattern model is used for the maximum likelihood solution.

  As shown in FIG. 5A, the i-th maximum likelihood solution in step ST2 is one of recognition result candidates obtained by the acoustic model and the language model, as in the conventional case. Note that “<” represents a head symbol and “>” represents a terminal symbol.

  In step ST3, the second decoder 72 matches the “expression pattern” in the concept expression pattern model with the longest match for all the partial word strings in the maximum likelihood parameter R. Then, a corresponding conceptual symbol is assigned to the matched partial word string, and the score of the corresponding “expression pattern” is added to the likelihood parameter L, and the process proceeds to step ST4.

  For example, in the maximum likelihood parameter R shown in FIG. 5A, as shown in FIG. 5B, “in the Chikusa-ku” matches one of the “expression patterns” in the concept “detailed address”. For this reason, a concept symbol [address] is assigned to “in the Chikusa-ku” and a score corresponding to this expression pattern (= Score (up to Chikusa-ku | [address])) is added to the likelihood parameter L. To do.

  Furthermore, “to Nakayoshi” matches one of the “expression patterns” in the concept “store name”. Therefore, a concept symbol [name] is given to “up to Nakayoshi”, and a score (= Score (| [name] up to Nakayoshi)) corresponding to this expression pattern is further added to the likelihood parameter L. At this time, the likelihood parameter L obtained by the second decoder 72 is expressed by, for example, (Expression 4).

  In step ST4, the second decoder 72 extracts the first symbol, the concept symbol, and the terminal symbol from the maximum likelihood parameter R, and sequentially matches one or more removed concept symbols with the “appearance pattern” of the concept appearance pattern model. . Then, the score of the “appearance pattern” that matches the combination order of the concept symbols is obtained, this score is added to the likelihood parameter L, and the process proceeds to step ST5.

  For example, the concept symbols [address] and [name] are sequentially removed from the text shown in FIG. Therefore, since the appearance pattern ([address] [name]) is matched with the combination order of the concept symbols, as shown in FIG. 5C, the corresponding score (= Score ([address], [name] ])) Is added to the likelihood parameter L. At this time, the likelihood parameter L obtained by the second decoder 72 is expressed by, for example, (Equation 5).

  In step ST5, the second decoder 72 updates the likelihood of the i-th maximum likelihood solution in the maximum likelihood matrix C to the value of the likelihood parameter L, and proceeds to step ST6.

  In step ST6, the second decoder 72 determines whether i = N. If the determination is affirmative, the process proceeds to step ST8. If the determination is negative, the process proceeds to step ST7.

  In step ST7, the second decoder 72 increments i (i = i + 1), and returns to step ST2. As a result, the second decoder 72 performs the processing from step ST2 to step ST5 on the N maximum likelihood solutions and their scores, and then proceeds to step ST8.

  In step ST8, the second decoder 72 sorts the N maximum likelihood solutions in the maximum likelihood matrix C in the order of the likelihoods, and outputs the most significant solution as the final recognition result. The final recognition result data is digital / analog converted and supplied to the output unit 8.

  As described above, the speech recognition apparatus according to the present embodiment cannot be recognized by the N-gram model by using the concept appearance pattern model having strong semantic constraints between concepts for a plurality of maximum likelihood solutions. The recognition accuracy can be improved in consideration of the meaning between the concepts of long distances.

  In addition, the speech recognition device uses a conceptual expression pattern model that summarizes semantically important keywords and surrounding words as expression patterns, thereby placing relatively strong grammatical restrictions on the partial word strings of utterances. The recognition accuracy can be improved.

  Note that the present invention is not limited to the above-described embodiments, and can be similarly applied to those whose design is changed within the scope of the matters described in the claims.

  FIG. 6 is a block diagram showing another configuration of the speech recognition apparatus.

  The voice recognition device includes a microphone 1 that obtains a voice signal, a feature quantity extraction unit 2 that extracts a voice feature quantity from the voice signal, an acoustic model storage unit 3 that stores an acoustic model, and a language model storage that stores a language model. Unit 4, a concept expression pattern model storage unit 5 that stores a concept expression pattern model, a concept appearance pattern model storage unit 6 that stores a concept appearance pattern model, and a decoder that performs speech recognition using the speech feature amount and each model 9 and an output unit 8 for outputting the recognition result to a predetermined system.

  The decoder 7 shown in FIG. 1 calculates the N maximum likelihood solutions using the acoustic model and the language model, and then finalizes the highest likelihood solution using the concept expression pattern model and the concept appearance pattern model. Output as recognition result. On the other hand, the decoder 9 shown in FIG. 6 outputs a final recognition result at a time using an acoustic model, a language model, a concept expression pattern model, and a concept appearance pattern model.

  Further, the concept expression pattern model may be incorporated in the language model (language model storage unit 4) in advance. Thereby, the speech recognition apparatus can use the concept expression pattern model as a language model. Further, if a finite state automaton (FSA) is created in advance from the above four models, only FSA may be used for decoding.

  Furthermore, the speech recognition apparatus does not use the concept expression pattern model and the concept appearance pattern model at the time of decoding, but creates another model that incorporates knowledge of these models offline in advance, and uses the other model. Good.

  Needless to say, the speech recognition apparatus is not limited to being applied to the destination task dialogue system.

  As shown in FIG. 2, the concept expression pattern model is not limited to the “concept pattern” centered on the body, but for example, the premise such as “I want to go”, “I want to know”, “I want you to tell me”, etc. The “concept pattern” may be used. Further, the concept appearance pattern model is not limited to the one shown in FIG.

  Further, a program for executing the processes from step ST1 to step ST8 described above may be installed in the computer. As a result, the computer functions as a voice recognition device and can execute the processes from step ST1 to step ST8 described above.

  Of course, the program may be transmitted via a communication line or may be recorded on a recording medium such as an optical disk, a magnetic disk, or a semiconductor memory.

It is a block diagram which shows the structure of the speech recognition apparatus which concerns on embodiment of this invention. It is a figure which shows the concept expression pattern model memorize | stored in the concept expression pattern model memory | storage part. It is a figure which shows the concept expression pattern model memorize | stored in the concept appearance pattern model memory | storage part. It is the flowchart which showed the process sequence for a 2nd decoder to obtain a final recognition result. (A) is a diagram in which the i-th maximum likelihood solution is displayed in text, (B) is a diagram for explaining the processing when the conceptual expression pattern model is used for the i-th maximum likelihood solution, and (C) is the i-th maximum likelihood solution. It is a figure explaining a process when a concept appearance pattern model is used for likelihood. It is a block diagram which shows the other structure of a speech recognition apparatus.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Microphone 2 Feature-value extraction part 3 Acoustic model memory | storage part 4 Language model memory | storage part 5 Concept expression pattern model memory | storage part 6 Concept appearance pattern model memory | storage part 7, 9 Decoder 8 Output part 71 1st decoder 72 2nd decoder

Claims (9)

A feature amount extraction means for extracting a feature amount of an audio signal;
Acoustic model storage means for storing an acoustic model obtained by statistically modeling features;
Language model storage means for storing a language model that statistically models a chain of words;
A concept appearance pattern model storage means for storing a concept appearance pattern model having a plurality of appearance patterns representing a combination order of one or more concepts;
The feature amount extracted by the feature amount extraction unit, the acoustic model stored in the acoustic model storage unit, the language model stored in the language model storage unit, and the concept appearance stored in the concept appearance pattern model Recognition means for recognizing the voice represented by the voice signal based on the pattern model;
A speech recognition device comprising: The recognition means uses a concept appearance pattern model in which a score is associated with each appearance pattern for a plurality of recognition result candidates based on the feature quantity, the acoustic model, and the language model, and their scores. The speech recognition apparatus according to claim 1, wherein each score of the plurality of recognition result candidates is weighted, and a recognition result candidate of the highest score among the weighted scores is output. An expression pattern model storage means for storing an expression pattern model representing a plurality of expression patterns for each concept;
The speech recognition apparatus according to claim 1, wherein the recognition unit recognizes speech by further using an expression pattern model stored in the expression pattern model storage unit. The recognition means further uses a concept expression pattern model in which a score is associated with each expression pattern for a plurality of recognition result candidates based on the feature quantity, the acoustic model, and the language model, and their scores. The voice recognition device according to claim 3, wherein each score of the plurality of recognition result candidates is weighted. A feature amount extraction step for extracting a feature amount of an audio signal;
The feature amount extracted in the feature amount extraction step, an acoustic model in which the feature amount is statistically modeled, a language model in which the word chain is statistically modeled, and a combination order of one or more concepts are displayed. A recognition step of recognizing a voice represented by the voice signal based on a concept appearance pattern model having a plurality of appearance patterns.
A speech recognition method comprising: In the recognition step, a concept appearance pattern model in which a score is associated with each appearance pattern is used for a plurality of recognition result candidates based on the feature quantity, the acoustic model, and the language model and their scores. The speech recognition method according to claim 5, wherein each score of the plurality of recognition result candidates is weighted, and a recognition result candidate of the highest score among the weighted scores is output. The speech recognition method according to claim 5, wherein in the recognition step, speech is recognized by further using an expression pattern model representing a plurality of expression patterns and their scores for each concept. In the recognition step, a concept expression pattern model in which a score is associated with each expression pattern is further used for a plurality of recognition result candidates based on the feature quantity, the acoustic model, and the language model and their scores. The voice recognition method according to claim 7, wherein each score of the plurality of recognition result candidates is weighted. On the computer,
A feature amount extraction step for extracting a feature amount of an audio signal;
The feature amount extracted in the feature amount extraction step, an acoustic model in which the feature amount is statistically modeled, a language model in which the word chain is statistically modeled, and a combination order of one or more concepts are displayed. A recognition step of recognizing a voice represented by the voice signal based on a concept appearance pattern model having a plurality of appearance patterns,
Voice recognition program that executes

Images