JP2010175765A - Speech recognition device and speech recognition program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognition device for preventing reoccurrence of recognition error by performing feed-back of correction of speech recognition error performed by an operator. <P>SOLUTION: The speech recognition device 1 includes: a cache storage means 211 for storing words of corrected correction word rows in a predetermined word number part; a cache score calculation means 212 for calculating probability value allowing the words stored to the cache storage means 211 to appear in the cache storing means 211 as a cache score; a language score corrected means 221 for generating a corrected language score adding the cache score being an appearance probability value of the words stored in the cache storage means 211 to the language score being an appearance probability of the word obtained from a language model 12a; and a search means 232 for searching the word rows in which a connection probability value is maximum from the language model 12a as speech recognition result on the basis of the corrected language score. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a speech recognition apparatus and speech recognition program that perform speech recognition using a language model.

  At present, in order to perform broadcasting with subtitles added in a live broadcast program such as news, a voice recognition device is used to convert voice of a news manuscript or the like into character data in real time to create subtitles. The character data recognized by this voice recognition apparatus generally contains several percent of errors. Therefore, an operator who has heard the voice visually detects an error in the recognition result of the voice recognition device for the voice, and corrects it through the keyboard by the correction device, thereby generating and outputting a correct voice recognition result in real time. A system is disclosed (for example, see Patent Document 1).

  On the other hand, as a technique for improving the recognition accuracy of voice recognition, a technique in which a cache is applied to a language model is disclosed (for example, see Non-Patent Document 1). According to this technology, by using the local nature of the word that the word used immediately before is easy to use again, the language model is weighted in consideration of the appearance interval of the word in the cache. , The appearance probability of words that appear more recently in the cache is increased, and the accuracy of speech recognition is increased.

JP 2005-49655 A

Yu Yamashita et al., “Effects of language model adaptation considering distances between identical words”, Acoustical Society Autumn Meeting, 3-1-7, pp. 91-94, September 2008

  The above-described system in which the operator corrects the recognition error of the speech recognition apparatus in real time can output a correct recognition result by manually correcting the recognition error of the speech recognition apparatus. However, in this system, since the correction made by the operator is not reflected in the voice recognition apparatus, the voice recognition apparatus repeats the same error. Therefore, there is a problem that the operator must perform the same correction many times.

  On the other hand, in the speech recognition technology to which the cache is applied, the recognition accuracy can be improved in the language model being used, but when a recognition error occurs, the result of correcting the error cannot be applied to the next recognition. For this reason, the conventional speech recognition using the cache has a problem that it is impossible to prevent the recurrence of errors, as in the above-described system.

  The present invention has been made in view of the above problems, and a speech recognition apparatus and a speech that can prevent a recurrence of a recognition error by feeding back a correction of a speech recognition error performed by an operator. The purpose is to provide a recognition program.

  The present invention has been made to achieve the above object. First, the speech recognition apparatus according to claim 1 utters a speech signal by a speech recognition apparatus using a pronunciation dictionary, an acoustic model, and a language model. In a speech recognition apparatus in a speech recognition system in which an operator corrects a recognition error of a character string obtained by recognition using a correction device and outputs it as a speech recognition result, a cache storage unit, a cache score calculation unit, a language The score correction means and the search means are provided.

  In such a configuration, the speech recognition apparatus inputs the corrected character string corrected by the correction device, and stores the words constituting the corrected character string in the cache storage unit for the predetermined number of words. As a result, the most recently corrected word is always stored in the cache storage means.

  Then, the speech recognition apparatus calculates, as a cache score, a probability value that the word appears in the cache storage unit with respect to the word stored in the cache storage unit by the cache score calculation unit. Furthermore, the speech recognition apparatus adds the cache score, which is the appearance probability value of the word stored in the cache storage means, to the language score, which is the appearance probability value of the word obtained from the language model, by the language score correction means. Thus, a corrected language score obtained by correcting the language score of the word obtained from the language model is generated. As a result, the weight for the most recent word input from the correction device is added to the language score, and the language score is set higher for that word.

  Then, the speech recognition apparatus searches the language model as a speech recognition result for the word string having the maximum connection probability value based on the corrected language score generated by the language score correcting unit. As a result, the language score is increased for the most recent word input from the correction device, and the probability that the most recent word is selected in the word string having the maximum connection probability value is increased.

  A voice recognition device according to a second aspect of the present invention is the voice recognition device according to the first aspect, comprising an IDF value storage means.

In such a configuration, the speech recognition apparatus stores an IDF (Inverse Document Frequency) value, which is a scale indicating how many specific documents appear in a plurality of documents, in the IDF value storage unit for each specific word. Store them in association with each other.
Then, the speech recognition apparatus adds the IDF value stored in the IDF value storage unit associated with the word corresponding to the cache score as a weighting value to the cache score by the cache score calculation unit. By using the IDF value in this way, a greater weight is added to a word such as a keyword that appears only in a specific document.

  Furthermore, the speech recognition device according to claim 3 is the speech recognition device according to claim 1 or 2, wherein the corrected character string includes a word corrected by the operator and a word not corrected. The cache score calculation means adds a predetermined weighting value to the cache score for the corrected word corrected by the operator in the corrected character string input from the correction device. .

  In such a configuration, since the weight is added to the cache score of the word corrected by the operator in the voice recognition device, the language score is set higher for the corrected word. Whether or not the word is corrected can be obtained from the correction device. Alternatively, the voice recognition device may obtain the character string before correction and the character string after correction from the correction device, and determine by comparing them.

  A voice recognition device according to a fourth aspect of the present invention is the voice recognition device according to any one of the first to third aspects, further comprising known word conversion means.

  In such a configuration, the speech recognition apparatus converts an unknown word not registered in the language model into a known word registered in the language model for each word of the corrected character string input from the correction apparatus by the known word conversion unit. Decompose. As a result, unknown words that are not registered in the language model are converted into known words for which probability values are set in the language model and stored in the cache storage means.

  The speech recognition device according to claim 5 is the speech recognition device according to any one of claims 1 to 3, further comprising an external dictionary storage unit and an unknown word substitution unit. did.

In such a configuration, the speech recognition apparatus stores an external dictionary, which is a second pronunciation dictionary having a larger number of registered words than the pronunciation dictionary, in the external dictionary storage unit. Then, the voice recognition device acquires, from the external dictionary, the pronunciation of the unknown word that is not registered in the language model for each word of the corrected character string input from the correction device by the unknown word substitution unit, and the unknown word is The probability value to appear is set in place of a probability value registered in advance as an unknown word connection probability value in the language model. As a result, the unknown word probability value registered in advance in the language model is assigned and assigned to the unknown word input from the correction device.
Then, the speech recognition apparatus searches for a word string having the maximum connection probability value based on the connection probability value in the language model and the connection probability value of the substituted unknown word by the search means.

  Furthermore, the speech recognition apparatus according to claim 6 is the speech recognition apparatus according to any one of claims 1 to 3, further comprising phoneme recognition means and second unknown word substitution means. It was.

In such a configuration, the speech recognition device generates phonetic sound data based on the acoustic model by the phoneme recognition means. The speech recognition apparatus acquires, from the phoneme recognition means, the pronunciation of the unknown word that is not registered in the language model for each word of the corrected character string input from the correction apparatus by the second unknown word substitution means. A probability value at which an unknown word appears is set in place of a probability value registered in advance as an unknown word connection probability value in the language model. As a result, the unknown word probability value registered in advance in the language model is assigned and assigned to the unknown word input from the correction device.
Then, the speech recognition apparatus searches for a word string having the maximum connection probability value based on the connection probability value in the language model and the connection probability value of the substituted unknown word by the search means.

  A speech recognition device according to claim 7 is the speech recognition device according to any one of claims 1 to 3, wherein an external dictionary storage unit, a phoneme recognition unit, a known word conversion unit, An unknown word substitution unit and a second unknown word substitution unit are further provided.

In such a configuration, the speech recognition apparatus stores an external dictionary, which is a second pronunciation dictionary having a larger number of registered words than the pronunciation dictionary, in the external dictionary storage unit. Further, the speech recognition apparatus generates sound generation data of the speech signal based on the acoustic model by the recognition unit.
Then, the speech recognition apparatus decomposes the unknown words not registered in the language model into known words registered in the language model for each word of the corrected character string input from the correction apparatus by the known word conversion unit. . As a result, unknown words that are not registered in the language model are converted into known words for which probability values are set in the language model and stored in the cache storage means.

  In addition, the speech recognition device acquires from the external dictionary the pronunciation of the unknown word that could not be decomposed into the known word by the known word conversion means by the unknown word substitution means, and the probability value that the unknown word appears in the language model in advance. A probability value registered as an unknown word connection probability value is used instead. As a result, the unknown word probability value registered in advance in the language model is assigned and assigned to the unknown word input from the correction device.

Then, the speech recognition apparatus acquires, from the phoneme recognition unit, the pronunciation of the unknown word that could not be acquired from the external dictionary by the unknown word substitution unit by the second unknown word substitution unit, and the probability value that the unknown word appears Is substituted with a probability value registered in advance as an unknown word connection probability value in the language model. As a result, the unknown word probability value registered in advance in the language model is assigned and assigned to the unknown word input from the correction device.
Then, the speech recognition apparatus searches for a word string having the maximum connection probability value based on the connection probability value in the language model and the connection probability value of the substituted unknown word by the search means.

  Furthermore, the speech recognition program according to claim 8 corrects a recognition error of a character string obtained by speech recognition of a speech signal by a speech recognition device using a pronunciation dictionary, an acoustic model, and a language model. In a speech recognition system that is corrected by a device and outputs as a speech recognition result, in order to reduce the recognition error by a character string corrected by the correction device, a computer of the speech recognition device is connected to a cache score calculation means, a language score The configuration is made to function as correction means and search means.

  In such a configuration, the speech recognition program uses the cache score calculation unit to store the words stored in the cache storage unit that stores the words of the corrected character string corrected by the correction device by a predetermined number of words. A probability value that the word appears in the cache storage means is calculated as a cache score.

  Then, the speech recognition program adds the cache score, which is the appearance probability value of the word stored in the cache storage means, to the language score, which is the word appearance probability value obtained from the language model, by the language score correction means. Thus, a corrected language score obtained by correcting the language score of the word obtained from the language model is generated. As a result, the weight for the most recent word input from the correction device is added to the language score, and the language score is set higher for that word.

  Then, the speech recognition program searches the language model as a speech recognition result for the word string having the maximum connection probability value based on the corrected language score generated by the language score correcting unit. As a result, the language score is increased for the most recent word input from the correction device, and the probability that the most recent word is selected in the word string having the maximum connection probability value is increased.

The present invention has the following excellent effects.
According to the first and eighth aspects of the present invention, the language score of the word input from the correction device is increased in speech recognition by setting a high language score for the most recent word input from the correction device. Can do. As a result, a recognition error occurs in speech recognition, and the appearance probability of the word corrected by the correction device can be increased, so that the recurrence of the same recognition error can be prevented.

  According to the second aspect of the present invention, since the IDF value is added as a weighting value to the cache score, the appearance probability can be increased for important words such as keywords. As a result, it is possible to prevent the adverse effect that the cache score becomes large for a function word (attachment word) such as a particle, and it is possible to suppress errors in the function word.

  According to the invention described in claim 3, since the language score is set higher for the word corrected by the operator in the correction device, the appearance probability of the word corrected by the operator, that is, the correct word can be increased. And recognition errors in voice recognition can be suppressed.

  According to the fourth aspect of the present invention, an accurate appearance probability value for an unknown word is obtained by decomposing an unknown word composed of a chain of known words input from a correction device into individual known words. Therefore, the accuracy of speech recognition can be improved.

  According to the fifth and sixth aspects of the invention, the unknown word appearance probability value registered in advance in the language model can be substituted for the unknown word appearance probability value input from the correction device. An unknown word input from the apparatus can be a word to be recognized by speech recognition.

  According to the seventh aspect of the present invention, an accurate appearance probability value for an unknown word is obtained by decomposing an unknown word composed of a chain of known words input from the correction device into individual known words. Therefore, the accuracy of speech recognition can be improved. Furthermore, according to the invention described in claim 7, the appearance probability value of the unknown word registered in advance in the language model can be substituted for the appearance probability value of the unknown word input from the correction device. An unknown word input from the apparatus can be a word to be recognized by speech recognition.

1 is a block diagram showing an overall configuration of a voice recognition system according to an embodiment of the present invention. It is an example of a screen for demonstrating the operation screen of the speech recognition system which concerns on embodiment of this invention. It is a block diagram which shows the structure of the speech recognition apparatus which concerns on 1st Embodiment of this invention. It is a flowchart which shows operation | movement of the speech recognition apparatus which concerns on 1st Embodiment of this invention. It is a block diagram which shows the structure of the speech recognition apparatus which concerns on 2nd Embodiment of this invention. It is a block diagram which shows the structure of the speech recognition apparatus which concerns on 3rd Embodiment of this invention. It is a block diagram which shows the structure of the speech recognition apparatus which concerns on 4th Embodiment of this invention. It is a block diagram which shows the structure of the speech recognition apparatus which concerns on 5th Embodiment of this invention. It is a block diagram which shows the structure of the speech recognition apparatus which concerns on 6th Embodiment of this invention.

Embodiments of the present invention will be described below with reference to the drawings.
<Voice recognition system>
Initially, with reference to FIG. 1, the structure of the speech recognition system which concerns on embodiment of this invention is demonstrated. This speech recognition system S is a system that recognizes an input speech, and manually corrects an erroneously recognized character string and outputs it as a speech recognition result in real time (including a correction time). For example, the voice recognition system S converts a voice of a news manuscript or the like into character data in real time and outputs it as subtitle data in order to perform broadcasting with subtitles added in a live broadcast program such as news. Here, the speech recognition system S includes a speech recognition device 1 and a correction device 3.

The speech recognition apparatus 1 recognizes an input speech signal (input speech) and outputs a recognition result as a character string. For example, the speech recognition apparatus 1 outputs a program sound as a character string when a speaker (re-speaker) who has heard the sound of a program such as news inputs the sound again through a microphone (not shown). To do. The character string that is the recognition result of the voice recognition device 1 is output to the correction device 3.
When the correction device 3 corrects the recognition result, the voice recognition device 1 feeds back the correction result (corrected character string) to reflect the subsequent voice recognition. The voice recognition device 1 will be described in detail later.

The correction device 3 is for the operator to correct the character string recognized by the speech recognition device 1 when there is an error. For example, the correction device 3 displays a screen G as shown in FIG. 2 on a display device not shown.
In the example of the screen G in FIG. 2, the correction device 3 displays a recognition result display region 3 a, a pre-correction character string display region 3 b, a corrected character string input region 3 c, and a send button 3 d on the screen G of the display device. Is displayed.

The recognition result display area 3 a is an area for displaying a character string of a voice recognition result input from the voice recognition device 1.
The pre-correction character string display area 3b is an area for displaying a character string in which a recognition error has occurred in the character string displayed in the recognition result display area 3a. In the pre-correction character string display area 3b, the operator compares the voice heard by the operator with the recognition result displayed in the recognition result display area 3a, and a recognition error occurs in the recognition result display area 3a. A character string selected by an operation with a mouse or a touch panel (not shown) is displayed as a character string.

The corrected character string input area 3c is an area in which the operator inputs a correct character string in which the character string displayed in the pre-correction character string display area 3b is corrected.
The example of FIG. 2 shows an example in which “Ministry of Education, Culture, Sports, Science and Technology” is mistakenly recognized as “MEXT or movement” in the recognition result input from the speech recognition apparatus 1.
In this case, the operator selects “sentence or movement” in the recognition result display area 3a. At this time, “text or movement” is displayed in the pre-correction character string display area 3b. Then, the operator inputs “Ministry of Education, Culture, Sports, Science and Technology” in the corrected character string input area 3c as a correct character string.
In this way, the correction device 3 corrects the erroneously recognized character string by the operator.

The send button 3d on the screen G is pressed by the operator when the correction is completed or when correction is not necessary, so that the corrected character string (the original character string if there is no correction). ) Is a button for outputting. By pressing the send button 3d, the correction device 3 outputs a character string (or a corrected character string) that is voice recognition of the voice recognition device 1 as subtitle data.
When the speech recognition result is corrected, the correction device 3 converts the corrected character string (in the example of FIG. 2, “Ministry of Education, Culture, Sports, Science and Technology”) at the timing of pressing the send button 3d. As “Text / Science / Ministry”: “/” indicates a word segmentation), it is fed back to the speech recognition apparatus 1.

By configuring the speech recognition system S in this way, the speech recognition system S feeds back the character string (corrected character string) in which the error of the speech recognition result is corrected in the correction device 3 to the speech recognition device 1. As a result, the recognition accuracy of voice recognition is improved.
Hereinafter, the speech recognition apparatus 1 (1B to 1F) according to the present invention will be described in detail.

[First Embodiment]
<Configuration of voice recognition device>
First, the configuration of the speech recognition apparatus according to the first embodiment of the present invention will be described with reference to FIG. The speech recognition device 1 shown in FIG. 3 performs speech recognition of input speech with reference to a character string (corrected character string) fed back from the correction device 3. Here, the speech recognition apparatus 1 includes a pronunciation dictionary storage unit 10, an acoustic model storage unit 11, a language model storage unit 12, a speech analysis unit 20, a cache processing unit 21, a score combination unit 22, and a word string. Generating means 23.

  The pronunciation dictionary storage means 10 stores the pronunciation dictionary 10a and is a general storage device such as a hard disk. The pronunciation dictionary 10a shows the structure of consonants and vowels indicating the pronunciation for each word, and the pronunciation of a plurality of words is registered in advance.

  The acoustic model storage unit 11 stores the acoustic model 11a and is a general storage device such as a hard disk. The acoustic model 11a is obtained by modeling a feature amount for each phoneme learned in advance from a large amount of speech data using a “hidden Markov model”. As the acoustic model 11a, a single acoustic model may be used, or a plurality of models may be used for each acoustic type (for example, for each person).

The language model storage unit 12 stores a language model 12a and is a general storage device such as a hard disk. The language model 12a models the appearance probability of an output sequence (words, morphemes, etc.) learned from a large amount of text. As this language model, for example, a general “N-gram language model” can be used.
Here, the pronunciation dictionary storage means 10, the acoustic model storage means 11, and the language model storage means 12 are configured as separate storage devices, but the pronunciation dictionary 10a and the acoustic model are stored in one storage device. 11a and the language model 12a may be stored. Further, here, an example is shown in which the pronunciation dictionary 10a, the acoustic model 11a, and the language model 12a are stored in the hard disk. However, during operation, the word string generation means 23 can refer to them for speeding up. It is assumed that it is expanded in a memory not shown.

The voice analysis means 20 analyzes a voice signal (input voice) input from the outside and extracts a feature amount of the voice signal as a feature vector. The feature vector extracted by the voice analysis unit 20 is output to the word string generation unit 23.
Note that the voice analysis means 20 extracts a framed waveform by applying a window function (such as a Hamming window) to the voice waveform of the voice signal, and extracts various feature amounts by performing frequency analysis on the waveform. To do. For example, a cepstrum coefficient that is a value obtained by inverse Fourier transform of the logarithm of the power spectrum of a framed waveform is used as the feature amount. In addition to the cepstrum coefficients, general audio feature quantities such as a mel frequency cepstrum coefficient (MFCC), an LPC (Linear Predictive Cording) coefficient, logarithmic power, and the like can be used as the feature quantity.

  The cache processing means 21 inputs the corrected character string fed back from the correction device 3 for each word, stores (caches) it, and calculates a cache score that is the appearance probability of the word in the cache. Here, the cache processing unit 21 includes a cache storage unit 211 and a cache score calculation unit 212.

  The cache storage unit 211 stores a predetermined number (M) of the latest words for the corrected character string fed back from the correction device 3, and is configured by a general memory or the like. . The cache storage unit 211 sequentially stores the latest M (for example, 3000) words.

The cache score calculation unit 212 calculates a cache score that is an appearance probability value of a word in the cache (cache storage unit 211). The cache score calculated by the cache score calculating unit 212 is output to the language score correcting unit 221 of the score combining unit 22.
Here, the latest word that has been cached notation _w n, the word of M one before and _{w n-M,} a word string _{w n-M, w n-} M + 1, ..., a _{w n-1 w n-M} when expressed as ^n-1, the cache score calculating means 212, a word _{w n} cache score _{P C (w n | w n} -M n-1) the following (1) is calculated by the equation.

In this equation (1), δ (·) is the Kronecker δ function, and when the arguments are equal, that is, when _wn and wn _−m are equal in equation (1), “1”, otherwise In this case, the function is “0”. Hereinafter, the probability model given by the equation (1) is referred to as a cache model.

  The score combining means 22 corrects the language score by combining a cache score, which is a word appearance probability value stored as a cache, with a language score, which is a word appearance probability value obtained from the language model 12a. It is. Here, the score combining unit 22 includes a language score correcting unit 221.

  The language score correcting unit 221 corrects the language score by combining the cache score calculated by the cache score calculating unit 212 with the language score obtained from the language model 12a. The language score (corrected language score) corrected by the language score correcting unit 221 is output to the searching unit 232 of the word string generating unit 23.

Here, the language model 12a and N-gram language model, a score (language scores) of a word _{w n} obtained from the language model _12a, P _LM = _| when the ^{_{(w n w n-N +}} 1 n-1), The language score correction means 221 combines the score of the language model 12a (language score) and the score of the cache model (cache score) by the following equation (2) to correct the corrected language score P (W _n | w ₁ ⁿ⁻¹ ) is generated.

  In the equation (2), λ is a constant satisfying 0 ≦ λ ≦ 1. The constant λ may be set to a value that reduces speech recognition errors by experiments in advance, or may be set to a value that minimizes a predetermined text perplexity (the number of predicted output sequences of text). .

  The word string generation means 23 generates a word string (recognized character string) as a voice recognition result from the feature vector extracted by the voice analysis means 20 based on the pronunciation dictionary 10a, the acoustic model 11a, and the language model 12a. It is. Here, the word string generation unit 23 includes an acoustic score calculation unit 231 and a search unit 232.

  The acoustic score calculation means 231 calculates the similarity (probability value) between the feature vector extracted by the speech analysis means 20 and input in time series and the phoneme modeled by the acoustic model 11a as an acoustic score. It is. The acoustic score calculation unit 231 calculates an acoustic score for each output sequence search candidate sequentially output from the search unit 232 described later. The calculated acoustic score is output to the search means 232.

Based on the acoustic score calculated by the acoustic score calculating unit 231, the searching unit 232 searches the language model 12 a for a candidate for the output series to be connected, and the search candidate that is the search result is stored in the acoustic score calculating unit 231. In addition to outputting, an output sequence having the maximum acoustic score and connection probability (language score) is output to the outside as a recognition result for the input speech.
The search means 232 uses the language score (corrected language score) corrected by the language score correcting means 221 for the language model 12a. Then, the search unit 232 searches the language model 12a (corrected language model) for an output series that maximizes the product of the acoustic score and the language score (corrected language score).

  By configuring the speech recognition apparatus 1 in this manner, the score of the corrected word (cache score) stored in the cache (cache storage unit 211) is used as the language score for speech recognition using the conventional language model. Therefore, a weight is given to the word input from the correction device 3 most recently, so that the accuracy of speech recognition can be improved and the recurrence of recognition errors can be prevented.

  The voice recognition device 1 can be realized by a general computer having a CPU and a memory (not shown). At this time, the speech recognition apparatus 1 operates by a speech recognition program that causes a computer to function as each of the above-described means.

<Operation of voice recognition device>
Next, the speech recognition operation of the speech recognition apparatus according to the first embodiment of the present invention will be described with reference to FIG. Note that the corrected character string output from the correction device 3 is input to the speech recognition device 1 and stored in the cache storage unit 211. In the following description, description of the cache operation is omitted.

First, the speech recognition apparatus 1 analyzes a speech signal (input speech) input from the outside by the speech analysis unit 20, and extracts a feature amount of the speech signal as a feature vector (step S1).
Then, the speech recognition apparatus 1 uses the search means 232 to sequentially list output sequence candidates to be connected from the language model 12a (step S2). Then, the speech recognition apparatus 1 uses the acoustic score calculation unit 231 to calculate the phoneme feature amount in the pronunciation acoustic model 11a indicated by the pronunciation dictionary 10a for each search candidate of the output series listed in step S2, and step S1. The similarity (probability value) with the feature amount of the input speech extracted in step S3 is calculated as an acoustic score (step S3).

  Furthermore, the speech recognition apparatus 1 uses the search unit 232 to calculate a connection probability (language score) in the language model 12a for each output series candidate listed in step S2. That is, the speech recognition apparatus 1 uses the cache score calculation unit 212 of the cache processing unit 21 to determine the cache score that is the appearance probability value of the word stored in the cache storage unit 211 among the words in the output series (1 ) (Step S4).

Then, the speech recognition device 1 uses the language score correcting unit 221 of the score combining unit 22 to determine the language score that is the appearance probability value of the word in the output series candidate and the cache score calculated in step S4 in the language model 12a. Are combined to generate a corrected language score in which the language score is corrected (step S5).
Then, the speech recognition apparatus 1 outputs, as the speech recognition result, an output sequence in which the product of the acoustic score calculated in step S3 and the corrected language score generated in step S5 is maximized by the search unit 232 (step S6).

  With the above operation, the speech recognition device 1 can correct the language score by using the word input from the correction device 3 most recently and weight the word, thereby improving the accuracy of speech recognition and recognizing The recurrence of errors can be prevented. Furthermore, it is possible to reduce the trouble of correction by the operator.

[Second Embodiment]
Next, the configuration of the speech recognition apparatus according to the second embodiment of the present invention will be described with reference to FIG. The speech recognition device 1B shown in FIG. 5 performs speech recognition of input speech with reference to a character string (corrected character string) fed back from the correction device 3 as in the speech recognition device 1 described in FIG. It is. This speech recognition apparatus 1B is different from the speech recognition apparatus 1 in that the words in the cache are further weighted.

  Here, the speech recognition apparatus 1B includes a pronunciation dictionary storage unit 10, an acoustic model storage unit 11, a language model storage unit 12, a weight value storage unit 13, a speech analysis unit 20, a cache processing unit 21B, a score Combining means 22 and word string generating means 23 are provided. Since the configuration other than the weight value storage unit 13 and the cache processing unit 21B is the same as the configuration of the speech recognition apparatus 1 described with reference to FIG. The internal configuration of the cache processing unit 21, the score combining unit 22, and the word string generating unit 23 is the same as that of the speech recognition apparatus 1 described with reference to FIG.

  The weight value storage means (IDF value storage means) 13 stores in advance a weight value set for a word stored in the cache storage means 211 in association with the word. This weighting value has a general meaning other than that for a function word (attachment) that is a word having a grammatical role such as a particle (“GA”, “HA”, “NO” ...). Set so that the content word that is a word has a greater weight.

  Here, an IDF (Inverse Document Frequency) value, which is a scale indicating how many specific documents appear in the entire document, is used as the weighting value, and an IDF table 13a in which an IDF value is associated with a word. Is stored in the weight value storage means 13. The IDF value is learned in advance by a set of documents (IDF learning document set 30a) prepared in advance for IDF learning stored in the external storage unit 30. The IDF learning document set 30a is, for example, a news manuscript used for news or the like during a certain period.

Here, when the number of documents contained in the IDF learning document set 30a _{N d,} a number of documents that contain word _{w n} in a document and df _n, IDF value of word _{w n} (IDF _(w n)) is given by the following expression (3), in association with the IDF value to word _{w n,} the IDF table 13a.

  This IDF value is a small value for words such as function words that appear evenly in each document, and a large value for words such as keywords that appear only in a specific document.

  The cache processing unit 21B inputs the corrected character string fed back from the correction device 3, stores (caches) the latest number of words (assumed to be M), and weights the cached words. Is to do. Here, the cache processing unit 21B includes a cache storage unit 211 and a cache score calculation unit 212B. The cache storage unit 211 has the same configuration as the voice recognition device 1 described in FIG.

The cache score calculation unit 212B calculates a cache score by adding a weight to the appearance probability value of the word in the cache (cache storage unit 211). The cache score calculated by the cache score calculating unit 212B is output to the language score correcting unit 221 of the score combining unit 22.
Here, the cache score calculation means 212B is an appearance probability value obtained by weighting the words in the cache storage means 211 with the weight values (IDF values) of the IDF table 13a stored in the weight value storage means 13. Calculate the cash score.
Specifically, the latest cached word is _denoted by wn, the Mth previous word is denoted by _wn-M, and the word string wn _-M , wn _{-M + 1} , ..., wn _-1 is _denoted by wn. when expressed as _-M ^n-1, the cache score calculating unit 212B, the word _{w n} cache score _{P C (w n | w n} -M n-1) the following (4) is calculated by the equation.

In this equation (4), [delta] (·), like the equation (1), a [delta] function Kronecker, when the argument is equal, i.e., is equal to _{w n} and _{w n-m} in (4) It is a function that is “1” at times, and “0” at other times. Z is a normalization coefficient for satisfying the probability axiom, and is a value calculated by the following equation (5).

As described above, the speech recognition apparatus 1B weights the appearance probability value of the word in the cache (cache storage unit 211) by the IDF value, and therefore, the relative score of the function word such as the particle included in the word in the cache is relative. Therefore, it is possible to suppress the occurrence of error in function words in the speech recognition result.
In the present embodiment, the cache score calculation unit 212B may use the part of speech information of a word instead of the IDF value to reduce the weight of the function word. However, the part-of-speech of a word may not be uniquely determined, and it is difficult to estimate the correct part-of-speech of each word from the context of a character string including an error such as a speech recognition result. Therefore, it is desirable to use an IDF value. It can be said.

Here, the cache score calculation unit 212B weights only the IDF value to the appearance probability value of the word in the cache (cache storage unit 211). May be performed.
Specifically, when the weight w indicating whether or not the word w fed back from the correction device 3 is a corrected word is expressed as weight (w), the cache score calculation means 212B uses the equation (4) and (5) instead of equation by the following equation (6) and (7), a word _{w n} cache score _P C _| calculates the ^{_{(w n w n-M n}} -1).

The value of weight (w) is “0” if the word w is the recognition result itself, and a positive constant if the word w is a corrected word.
Here, whether or not the word w is a corrected word can be determined by the cache processing unit 21B by adding information in the correction device 3. For example, in the example of FIG. 2, the correction device 3 adds “0” for the word “sentence” of the recognition result itself that has not been corrected, and adds a positive constant for “science” that has been corrected. And feed back to the speech recognition apparatus 1B.

  Whether or not the word w is a corrected word is determined by feedback from the correction device 3 of the character string before correction (recognition result) and the character string after correction (corrected character string). By performing DP matching using a correction determination unit that omits the above, both character strings may be compared, and a word with a difference may be determined to be a corrected word. For example, in the example of FIG. 2, the correction device 3 uses the speech recognition device 1 </ b> B as “text section or movement” as the character string before correction (recognition result) and “Ministry of Education, Culture, Sports, Science and Technology” as the character string after correction (corrected character string). To give feedback. In this way, in addition to the IDF value, the speech recognition apparatus 1B assigns a large weight to the language score for the word corrected by the operator by performing weighting depending on whether or not the correction is performed. Generation of errors can be prevented.

  The voice recognition operation of the voice recognition device 1B is basically the same as the operation of the voice recognition device 1 described with reference to FIG. Differences in operation between the speech recognition device 1B and the speech recognition device 1 are stored in the cache storage unit 211 among the words in the output series by the cache score calculation unit 212B of the cache processing unit 21 in step S4 of FIG. In other words, the cache score, which is the appearance probability value of the word that is being used, is calculated by the above equation (4) or (5).

  The voice recognition device 1B described above can be realized by a general computer having a CPU and a memory (not shown). At this time, the speech recognition apparatus 1B operates by a speech recognition program that causes the computer to function as each of the above-described means.

[Third Embodiment]
Next, with reference to FIG. 6, the structure of the speech recognition apparatus according to the third embodiment of the present invention will be described. The speech recognition apparatus 1C shown in FIG. 6 can recognize unknown words that are not registered in the language model by using the corrected character string fed back from the correction apparatus 3 in addition to the functions of the speech recognition apparatus 1 described in FIG. To do.

  Here, the speech recognition apparatus 1C includes a pronunciation dictionary storage unit 10, an acoustic model storage unit 11, a language model storage unit 12, a speech analysis unit 20, a cache processing unit 21, a score combining unit 22, a word string. A generation unit 23 and an unknown word processing unit 24 are provided. Since the configuration other than the unknown word processing means 24 is the same as the configuration of the speech recognition apparatus 1 described with reference to FIG. 3, the same reference numerals are given and description thereof is omitted. The internal configuration of the cache processing unit 21, the score combining unit 22, and the word string generating unit 23 is the same as that of the speech recognition apparatus 1 described with reference to FIG.

The unknown word processing means 24 converts an unknown word that is an unknown word into a word string of a known word that is a known word in the corrected character string that is corrected by the operator in the correction device 3.
Usually, in the correction device 3, there may be an unknown word that is not registered in the language model 12a in the word corrected by the operator. Therefore, the unknown word processing unit 24 includes a known word conversion unit 241 that converts an unknown word into a known word, and the unknown word conversion unit 241 converts the unknown word into a known word. The corrected character string converted into the word string of the known word is output to the cache processing means 21.

  That is, the known word conversion means 241 converts each word of the input corrected character string into a known word by referring to the language model 12a and decomposing it into words registered in the language model 12a. For example, when “Hello Work” is input as an unknown word, the known word conversion unit 241 decomposes the word into “Hello” and “Work” which are known words registered in the language model 12a. The individually decomposed known words are output to the cache processing unit 21 and stored in the cache storage unit 211.

By configuring the speech recognition device 1C in this manner, the speech recognition device 1C is known even when an unknown word is input as a corrected character string from the correction device 3 in addition to the effects of the first embodiment. By converting to a word, it can be reflected in an accurate language score, and the accuracy of speech recognition can be improved.
Further, the speech recognition apparatus 1C may be configured to include the weight value storage unit 13 described with reference to FIG. 5 in place of the cache processing unit 21B described with reference to FIG.

  The voice recognition operation of the voice recognition device 1C is the same as that of the voice recognition device 1 described with reference to FIG. In the cache operation of the speech recognition apparatus 1C, the modified character string fed back from the modification apparatus 3 is input to the speech recognition apparatus 1C and stored in the cache storage unit 211 by the known word conversion unit 241 by the unknown word conversion unit 241. Is different from the operation of the speech recognition apparatus 1 in that it is converted into a known word.

  The speech recognition apparatus 1C described above can be realized by a general computer equipped with a CPU and a memory (not shown). At this time, the speech recognition apparatus 1C operates by a speech recognition program that causes the computer to function as each of the above-described means.

[Fourth Embodiment]
Next, with reference to FIG. 7, the structure of the speech recognition apparatus which concerns on 4th Embodiment of this invention is demonstrated. The voice recognition device 1D shown in FIG. 7 can recognize unknown words that are not registered in the pronunciation dictionary by using the corrected character string fed back from the correction device 3 in addition to the function of the voice recognition device 1 described in FIG. To do.

  Here, the speech recognition apparatus 1D includes a pronunciation dictionary storage unit 10, an acoustic model storage unit 11, a language model storage unit 12, an external dictionary storage unit 14, a speech analysis unit 20, a cache processing unit 21, a score A combining unit 22, a word string generating unit 23, an unknown word processing unit 24B, and an unknown word registering unit 25 are provided. Since the configuration other than the external dictionary storage unit 14, the unknown word processing unit 24B, and the unknown word registration unit 25 is the same as the configuration of the speech recognition apparatus 1 described with reference to FIG. To do. The internal configuration of the cache processing unit 21, the score combining unit 22, and the word string generating unit 23 is the same as that of the speech recognition apparatus 1 described with reference to FIG.

  The external dictionary storage means 14 stores an external dictionary 14a, which is a pronunciation dictionary (second pronunciation dictionary), and is a general storage device such as a hard disk. Normally, in speech recognition, only a predetermined number of pronunciations of words are registered in the pronunciation dictionary 10a in order to increase recognition accuracy and recognition speed. The external dictionary 14a stored in the external dictionary storage means 14 is a huge dictionary in which pronunciations of more words than the pronunciation dictionary 10a are registered.

The unknown word processing means 24B gives an appearance probability value to an unknown word that is an unknown word in the corrected character string corrected by the operator in the correction device 3.
This unknown word processing means 24B gives the appearance probability value to the unknown word by the unknown word substitution means 242 that gives a predetermined substitute appearance probability value to the unknown word. The appearance probability value for this unknown word is output to the unknown word registration means 25.

  The unknown word substitution means 242 obtains pronunciations from the external dictionary storage means 14 for words of unknown words that are not registered in the language model 12a for the individual words of the input corrected character string, and stores them in the language model 12a in advance. This is output to the unknown word registration means 25 together with the appearance probability value of the registered unknown word. The unknown word processing unit 24B outputs the known words to the cache processing unit 21 as they are.

The unknown word registration unit 25 registers the pronunciation and appearance probability value of the unknown word output from the unknown word processing unit 24B in the pronunciation dictionary 10a and the language model 12a.
That is, the unknown word registration unit 25 registers the unknown word text data output from the unknown word processing unit 24B and the pronunciation of the unknown word in the pronunciation dictionary 10a. Further, the unknown word registration unit 25 registers the text data of the unknown word output from the unknown word processing unit 24B in the language model 12a, and registers the appearance probability value in association with the unknown word.
In the present embodiment, it is assumed that the pronunciation dictionary 10a and the language model 12a are expanded and operated in advance in a memory (not shown) that can be referred to by the word string generation unit 23. An output from the word registration unit 25 is used as a word string generation unit 23.

  By configuring the speech recognition device 1D in this way, the speech recognition device 1D can be used even when an unknown word is input as a corrected character string from the correction device 3 in addition to the effects of the first embodiment. An unknown word appearance probability value of a language model determined in advance can be substituted for the unknown word, and voice recognition can be performed for the unknown word.

  The voice recognition operation of the voice recognition device 1D is the same as that of the voice recognition device 1 described with reference to FIG. The speech recognition device 1D refers to the external dictionary 14a by the unknown word substitute unit 242B when the corrected character string fed back from the correction device 3 is input to the speech recognition device 1D. The point that the pronunciation and the appearance probability value are assigned is different from the operation of the speech recognition apparatus 1.

  The voice recognition device 1D described above can be realized by a general computer having a CPU and a memory (not shown). At this time, the speech recognition apparatus 1D operates by a speech recognition program that causes the computer to function as each of the above-described means.

[Fifth Embodiment]
Next, with reference to FIG. 8, the structure of the speech recognition apparatus according to the fifth embodiment of the present invention will be described. The voice recognition device 1E shown in FIG. 8 can recognize unknown words that are not registered in the pronunciation dictionary by using the corrected character string fed back from the correction device 3 in addition to the functions of the voice recognition device 1 described in FIG. To do.

  Here, the speech recognition apparatus 1E includes a pronunciation dictionary storage unit 10, an acoustic model storage unit 11, a language model storage unit 12, a speech analysis unit 20, a cache processing unit 21, a score combination unit 22, a word string. A generation unit 23B, an unknown word processing unit 24C, an unknown word registration unit 25, and a phoneme recognition unit 26 are provided. Since the configuration other than the word string generation unit 23B, the unknown word processing unit 24C, and the phoneme recognition unit 26 is the same as the configuration of the speech recognition apparatus 1D described with reference to FIG. . The internal configuration of the cache processing unit 21 and the score combining unit 22 is the same as that of the speech recognition apparatus 1 described with reference to FIG.

The word string generator 23B generates a word string (recognized character string) that is a voice recognition result from the feature vector extracted by the voice analyzer 20 based on the pronunciation dictionary 10a, the acoustic model 11a, and the language model 12a. It is. The internal configuration of the word string generation unit 23B is the same as that of the word string generation unit 23 of the speech recognition apparatus 1 described with reference to FIG.
Here, in addition to the function of the word string generation unit 23 described with reference to FIG. 3, the word string generation unit 23B adds time information (time stamp) measured by the time measurement unit (not shown) to the word. To do.

The unknown word processing means 24C gives an appearance probability value to an unknown word that is an unknown word in the corrected character string corrected by the operator in the correction device 3.
Here, the unknown word processing means 24C gives the appearance probability value to the unknown word by the unknown word substitution means 242B that gives a predetermined substitute appearance probability value to the unknown word. The appearance probability value for this unknown word is output to the unknown word registration means 25.

  The unknown word substitution means (second unknown word substitution means) 242B obtains pronunciation from the phoneme recognition means 26 for words of unknown words that are not registered in the language model 12a for individual words of the input corrected character string. Then, together with the appearance probability value of the unknown word registered in advance in the language model 12a, it is output to the unknown word registration means 25. The unknown word processing unit 24C outputs the known words to the cache processing unit 21 as they are.

  The unknown word substitution unit 242B compares the character string with the time stamp generated by the word string generation unit 23B with the corrected character string output from the correction device 3, thereby correcting the time of the corrected word. You can get a stamp. When the word is an unknown word, the unknown word substitution unit 242B associates the pronunciation output from the phoneme recognition unit 26 corresponding to the time stamp with the unknown word.

  The phoneme recognizing unit 26 decomposes a speech signal (input speech) input from the outside into phonemes with reference to the acoustic model 11a to generate pronunciation data. The phoneme recognition means 26 may be a general phoneme recognition device, and may be provided inside the speech recognition device 1E or may be configured to be connected to the outside. The phoneme recognizing means 26 gives time information (time stamp) to each phoneme by a time measuring means (not shown) as a recognition result, and outputs it to the unknown word processing means 24C.

  By configuring the speech recognition device 1E as described above, the speech recognition device 1E can be used in addition to the effects of the first embodiment, even when an unknown word is input as a corrected character string from the correction device 3. An unknown word appearance probability value of a language model determined in advance can be substituted for the unknown word, and voice recognition can be performed for the unknown word.

  The voice recognition operation of the voice recognition device 1E is similar to the operation of the voice recognition device 1 described with reference to FIG. Note that the speech recognition device 1E assigns pronunciation and appearance probability values to unknown words by the unknown word substitution means 242B when the corrected character string fed back from the correction device 3 is input to the speech recognition device 1E. Is different from the operation of the speech recognition apparatus 1.

  The voice recognition device 1E described above can be realized by a general computer equipped with a CPU and a memory (not shown). At this time, the speech recognition apparatus 1E operates according to a speech recognition program that causes the computer to function as each of the above-described means.

  As mentioned above, although 1st-5th embodiment was described as embodiment of this invention, this invention is not limited to these embodiment. For example, the third to fifth embodiments may be appropriately combined. For example, it can be set as 6th Embodiment which combined 3rd-5th embodiment.

[Sixth Embodiment]
Here, with reference to FIG. 9, the speech recognition apparatus of 6th Embodiment which combined 3rd-5th embodiment is demonstrated.
The speech recognition apparatus 1F shown in FIG. 9 is configured by combining the third to fifth embodiments with the unknown word processing means 24D, the known word conversion means 241, and the unknown word substitution means 242, 242B. Since each configuration has been described in the first to fifth embodiments, the same reference numerals are used and description thereof is omitted. However, since the function of the unknown word processing unit 24D is different, its operation will be described.

When a corrected character string is input from the correction device 3 and an unknown word is present in the corrected character string, the speech recognition device 1F uses the known word conversion unit 241 to decompose the unknown word into a plurality of known words. Convert to word.
Here, when the unknown word cannot be converted into the known word, the speech recognition apparatus 1F acquires the pronunciation from the external dictionary 14a of the external dictionary storage unit 14 by the unknown word substitution unit 242, and is registered in advance in the language model 12a. The unknown word appearance probability value is output to the unknown word registration means 25 for registration.

  Further, when the pronunciation of the unknown word cannot be acquired even in the external dictionary 14a, the speech recognition apparatus 1F acquires the pronunciation from the phoneme recognition means 26 by the unknown word substitution means 242B, and the unknown words registered in the language model 12a in advance. Is output to the unknown word registration means 25 and registered.

As a result, the speech recognition apparatus 1F can correct the language score by using the word of the corrected character string that has been input most recently, and can weight the word, thereby improving the accuracy of speech recognition and reducing the recognition error. Relapse can be prevented. Furthermore, even if the word is an unknown word, the speech recognition apparatus 1F can substitute the unknown word appearance probability value of the language model in advance for the unknown word. Recognition can be performed.
Further, the speech recognition apparatus 1F may be configured to include the weight value storage unit 13 described with reference to FIG. 5 in place of the cache processing unit 21B described with reference to FIG.

  The voice recognition device 1F described above can be realized by a general computer equipped with a CPU and a memory (not shown). At this time, the speech recognition apparatus 1E operates according to a speech recognition program that causes the computer to function as each of the above-described means.

S Speech recognition system 1 (1B to 1F) Speech recognition device 10 Pronunciation dictionary storage means 10a Pronunciation dictionary 11 Acoustic model storage means 11a Acoustic model 12 Language model storage means 12a Language model 13 Weight value storage means (IDF value storage means)
13a IDF table 14 External dictionary storage means 14a External dictionary (second pronunciation dictionary)
20 speech analysis means 21 cache processing means 211 cache storage means 212 cache score calculation means 22 score combination means 221 language score correction means 23 word string generation means 231 acoustic score calculation means 232 search means 24 unknown word processing means 241 known word conversion means 242 Unknown word substitution means 242B Unknown word substitution means (second unknown word substitution means)
25 Unknown word registration means 26 Phoneme recognition means 3 Correction device

Claims (8)

Speech recognition that the operator corrects the recognition error of the character string obtained by speech recognition by the speech recognition device using the pronunciation dictionary, acoustic model and language model, and outputs it as a speech recognition result In the speech recognition device in the system,
A cache storage means for inputting a corrected character string corrected in the correction device, and storing words constituting the corrected character string by a predetermined number of words;
For a word stored in the cache storage means, a cache score calculation means for calculating a probability value that the word appears in the cache storage means as a cache score;
By adding a cache score that is an appearance probability value of the word stored in the cache storage unit to a language score that is an appearance probability value of the word obtained from the language model, the word obtained from the language model Language score correction means for generating a corrected language score obtained by correcting the language score;
Based on the corrected language score generated by the language score correcting means, search means for searching the language model as a speech recognition result for a word string having a maximum connection probability value;
A speech recognition apparatus comprising: IDF value storage means for storing an IDF value, which is a scale indicating how many specific documents appear in a plurality of documents, in association with each specific word;
The cache score calculation means adds an IDF value stored in the IDF value storage means associated with a word corresponding to the cache score as a weighting value to the cache score. The speech recognition apparatus according to claim 1. The corrected character string is a character string in which a word corrected by the operator and a word not corrected are mixed,
2. The cache score calculation unit adds a predetermined weight value to the cache score for a word corrected by the operator in a corrected character string input from the correction device. Or the speech recognition apparatus of Claim 2.   For each word of the corrected character string input from the correction device, the device further comprises known word conversion means for decomposing an unknown word not registered in the language model into a known word registered in the language model, The speech recognition apparatus according to claim 1, wherein a known word is stored in the cache storage unit. An external dictionary storage means for storing an external dictionary which is a second pronunciation dictionary having a larger number of registered words than the pronunciation dictionary;
For each word of the corrected character string input from the correction device, the pronunciation of an unknown word that is not registered in the language model is obtained from the external dictionary, and the probability value of the unknown word appearing in advance in the language model An unknown word substitution means that substitutes and sets a probability value registered as an unknown word connection probability value in
The search means searches for a word string having a maximum connection probability value based on a connection probability value in the language model and a connection probability value of the substituted unknown word. The speech recognition apparatus according to claim 3. Phoneme recognition means for generating pronunciation data of the speech signal based on the acoustic model;
For each word of the corrected character string input from the correction device, the pronunciation of an unknown word that is not registered in the language model is acquired from the phoneme recognition means, and the probability value that the unknown word appears is preliminarily set in the language. A second unknown word substitution means that substitutes and sets a probability value registered as a connection probability value of an unknown word in the model,
The search means searches for a word string having a maximum connection probability value based on a connection probability value in the language model and a connection probability value of the substituted unknown word. The speech recognition apparatus according to claim 3. An external dictionary storage means for storing an external dictionary which is a second pronunciation dictionary having a larger number of registered words than the pronunciation dictionary;
Phoneme recognition means for generating pronunciation data of the speech signal based on the acoustic model;
For each word of the corrected character string input from the correction device, an unknown word that is not registered in the language model is decomposed into known words that are registered in the language model, and stored in the cache storage unit. A known word conversion means;
The pronunciation of an unknown word that could not be decomposed into known words by this known word conversion means is acquired from the external dictionary, and the probability value that the unknown word appears is registered in the language model in advance as a connection probability value of the unknown word. An unknown word substitution means that substitutes with a certain probability value, and
The unknown word substitution means obtains the pronunciation of the unknown word that could not be obtained from the external dictionary from the phoneme recognition means, and the probability value of the unknown word appearing in advance in the language model is connected to the unknown word connection probability value. A second unknown word substitution means that substitutes and sets the probability value registered as:
The search means searches for a word string having a maximum connection probability value based on a connection probability value in the language model and a connection probability value of the substituted unknown word. The speech recognition apparatus according to claim 3. Speech recognition that the operator corrects the recognition error of the character string obtained by speech recognition by the speech recognition device using the pronunciation dictionary, acoustic model and language model, and outputs it as a speech recognition result In the system, in order to reduce the recognition error by the character string corrected by the correction device, the computer of the voice recognition device,
For a word stored in the cache storage means in which each word of the corrected character string corrected by the correction device is stored for a predetermined number of words, a probability value that the word appears in the cache storage means is determined as a cache score. Cash score calculation means for calculating as
By adding a cache score that is an appearance probability value of the word stored in the cache storage unit to a language score that is an appearance probability value of the word obtained from the language model, the word obtained from the language model A language score correcting means for generating a corrected language score obtained by correcting the language score;
Search means for searching a word string having a maximum connection probability value from the language model as the speech recognition result based on the corrected language score generated by the language score correcting means,
A voice recognition program characterized by functioning as

Images