JP2010078877A - Speech recognition device, speech recognition method, and speech recognition program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognition method recognizing speech including a predetermined keyword and accurately extracting the keyword. <P>SOLUTION: A speech recognition device performs speech recognition by using a sound model, a language model, and a dictionary. Once a speech is given, a plurality of candidate patterns each composed of a plurality of words are generated from content of the speech. A candidate pattern having the largest total score is determined as a recognition result and a keyword included in the candidate pattern is extracted in reference to the dictionary. In the language model, an appearance probability and a coefficient of a word string for a keyword similar word are set so that a language score of the candidate pattern including the keyword similar word has a minimum value. Therefore, the language score of the candidate pattern including the keyword similar word has the minimum value and is not regarded as a recognition result. Consequently, the candidate pattern including the keyword is regarded as a recognition result and the keyword is correctly extracted therefrom. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a speech recognition technique using a language model.

  A speech recognition method using a language model is known. The language model is a database that stores appearance probabilities for each combination of adjacent words, and an N-gram model is known as a representative language model by a statistical method. The N-gram language model gives the probability of occurrence for a combination of N consecutive words. A speech recognition method using a language model derives a recognition result using a language model in consideration of the appearance probability of a combination of a plurality of words.

  An example of a speech recognition device using a language model is described in Patent Document 1. In the method of Patent Document 1, a language pattern of a character string specified by a description grammar (network grammar) is mixed in a statistical language model, and a backoff coefficient and 1 for a word of a character string specified by a network grammar are used. -Set the gram probability to 0 (zero). As a result, the connection between a word defined in the network grammar and a word that is not so is prohibited, and a character string in accordance with the network grammar can be correctly detected even when the statistical language model is processed.

Japanese Patent No. 3950957

  However, in the above-described method, there is a possibility that the recognition accuracy of a language pattern that includes a word defined in the network grammar but is not defined in the network grammar is lowered.

  The above-mentioned thing is mentioned as an example as a subject which the present invention tends to solve. An object of the present invention is to provide a speech recognition method capable of recognizing an utterance including a predetermined keyword and accurately extracting the keyword.

  The invention according to claim 1 is a speech recognition device, wherein an acoustic model storage unit that stores an acoustic model, a dictionary storage unit that stores a dictionary having a plurality of words including a keyword, and a plurality of words, A language model storage unit that stores a language model having a coefficient for calculating an appearance probability of a word string including a word and a word string including the word, and a combination of a plurality of words from the utterance content A candidate pattern creation unit for creating a candidate pattern, and for each of the candidate patterns, an acoustic score is calculated based on the acoustic model, a language score is calculated based on the language model, and the acoustic score and the language A score calculation unit that calculates a total score based on the score, and a candidate pattern having the maximum total score among a plurality of candidate patterns A recognition result determination means for determining a recognition result as a recognition result, and a keyword extraction means for extracting a keyword from the recognition result and the determined candidate pattern with reference to the dictionary, and the non-similarity to the keyword in the language model. The appearance probability and the coefficient of the word string for a keyword similar word that is a keyword have values having a language score of a candidate pattern including the keyword similar word as a minimum value.

  The invention according to claim 5 calculates an appearance probability of a word string including the word and an appearance probability of the word string including the word with respect to the acoustic model, a dictionary having a plurality of words including the keyword, and the plurality of words. A speech recognition method using a language model having a coefficient for performing a candidate pattern creating step of creating a candidate pattern composed of a combination of a plurality of words from the utterance content, and each of the candidate patterns, A score calculation step of calculating an acoustic score based on the acoustic model and calculating a language score based on the language model, and calculating a total score based on the acoustic score and the language score, and a plurality of candidate patterns, A recognition result determination step for determining a candidate pattern having the maximum total score as a recognition result; A keyword extraction step of extracting a keyword from the result and the determined candidate pattern, and in the language model, the appearance probability of the word string and the coefficient for the keyword similar word that is a non-keyword similar to the keyword are the keyword It has the value which makes the language score of the candidate pattern containing a similar word the minimum value.

  The invention according to claim 6 is a speech recognition program executed by a computer, an acoustic model storage unit for storing an acoustic model, a dictionary storage unit for storing a dictionary having a plurality of words including keywords, and a plurality of words A language model storage unit for storing a language model having a coefficient for calculating an appearance probability of a word string including the word and an appearance probability of the word string including the word, a combination of a plurality of words from the utterance content For each of the candidate patterns, a candidate pattern creating unit that creates a candidate pattern configured by: calculating an acoustic score based on the acoustic model and calculating a language score based on the language model; and Score calculator that calculates the total score based on the language score, the largest total of multiple candidate patterns A recognition result determining means for determining a candidate pattern having a core as a recognition result, and referring to the dictionary storage unit, causing the computer to function as a keyword extracting means for extracting a keyword from the recognition result and the determined candidate pattern; In the language model, the appearance probability and the coefficient of the word string with respect to a keyword similar word that is a non-keyword similar to a keyword have a value that minimizes the language score of a candidate pattern including the keyword similar word. And

  In a preferred embodiment of the present invention, the speech recognition apparatus includes: an acoustic model storage unit that stores an acoustic model; a dictionary storage unit that stores a dictionary having a plurality of words including a keyword; A language model storage unit for storing a language model having a coefficient for calculating an appearance probability of a word string including the word string and an appearance probability of the word string including the word, and a combination of a plurality of words from the utterance content A candidate pattern creating unit that creates a candidate pattern, and for each of the candidate patterns, calculates an acoustic score based on the acoustic model and calculates a language score based on the language model, and calculates the acoustic score and the language score. A score calculation unit that calculates a total score based on the candidate pattern, and a candidate pattern having a maximum total score among a plurality of candidate patterns A non-keyword similar to a keyword in the language model, comprising: a recognition result determining means for determining a recognition result; and a keyword extracting means for extracting a keyword from the recognition result and the determined candidate pattern with reference to the dictionary. The probability of appearance of the word string and the coefficient with respect to a keyword-similar word are values having a language score of a candidate pattern including the keyword-similar word as a minimum value.

  The speech recognition apparatus performs speech recognition using an acoustic model, a language model, and a dictionary. The language model has, for a plurality of words, coefficients for calculating the appearance probability of a word string including the word and the appearance probability of a word string including the word. The dictionary stores a plurality of words including keywords. When an utterance is made, a candidate pattern composed of a combination of a plurality of words is created from the utterance content. Usually, a plurality of candidate patterns are created for one utterance. Then, an acoustic score and a language score are calculated for each candidate pattern, and a total score is calculated therefrom. A candidate pattern having the maximum total score is taken as a recognition result, and keywords included in the candidate pattern are extracted with reference to the dictionary.

  In the language model, the appearance probability and coefficient of the word string for the keyword similar word have values having the minimum language score of the candidate pattern including the keyword similar word. Therefore, the language score of the candidate pattern including the keyword similar words becomes the minimum value and is not determined as the recognition result. As a result, the candidate pattern including the keyword becomes the recognition result, and the keyword is correctly extracted from the recognition result.

  In one aspect of the speech recognition apparatus, the score calculation unit calculates a product of the acoustic score and the language score as the total score, and the appearance probability of the word string and the coefficient for the keyword similar word are 0. . In this aspect, since the appearance probability and coefficient of the word string for the keyword similar word are 0, the candidate pattern including the keyword similar word has a total score of 0 and is not determined as a recognition result.

  In a preferred example, the language model is an N-gram language model, the word string appearance probability is a 1-gram probability, and the coefficient is a back-off coefficient. For keyword-like words, both the 1 gram probability and the backoff coefficient are set to 0.

  In another preferred example, the keyword-similar word includes a word that forms a part of the keyword and a word that includes a part or all of the keyword.

  In another embodiment of the present invention, for an acoustic model, a dictionary having a plurality of words including keywords, and for a plurality of words, the appearance probability of the word string including the word and the appearance probability of the word string including the word are calculated. A speech recognition method using a language model having a coefficient for performing a candidate pattern creation step of creating a candidate pattern composed of a combination of a plurality of words from speech content, and for each of the candidate patterns, the acoustic model A score calculation step of calculating a language score based on the language model and calculating a total score based on the acoustic score and the language score, and a maximum total score among a plurality of candidate patterns A recognition result determination step for determining a candidate pattern having a recognition result, and recognition with reference to the dictionary And a keyword extraction step of extracting a keyword from the determined candidate pattern, and in the language model, the appearance probability of the word string and the coefficient for the keyword similar word that is a non-keyword similar to the keyword are the keyword It has a value that makes the language score of a candidate pattern including similar words the minimum value.

  Also in the speech recognition method described above, speech recognition is performed using an acoustic model, a language model, and a dictionary. A candidate pattern having the maximum total score is taken as a recognition result, and keywords included in the candidate pattern are extracted with reference to the dictionary. Further, in the language model, the appearance probability and coefficient of the word string for the keyword similar word have values having the minimum language score of the candidate pattern including the keyword similar word. Therefore, the language score of the candidate pattern including the keyword-similar words is the minimum value and does not become a recognition result. As a result, the candidate pattern including the keyword becomes the recognition result, and the keyword is correctly extracted from the recognition result.

  In another embodiment of the present invention, a speech recognition program executed by a computer includes an acoustic model storage unit that stores an acoustic model, a dictionary storage unit that stores a dictionary having a plurality of words including keywords, and a plurality of words, A language model storage unit for storing a language model having a coefficient for calculating the appearance probability of a word string including the word and the appearance probability of the word string including the word, and composed of a combination of a plurality of words from speech content A candidate pattern creating unit that creates a candidate pattern to be calculated, and for each of the candidate patterns, an acoustic score is calculated based on the acoustic model and a language score is calculated based on the language model, and the acoustic score and the language score are calculated A score calculation unit that calculates a total score based on the maximum total score among a plurality of candidate patterns A recognition result determining means for determining a candidate pattern having a recognition result, and referring to the dictionary storage unit, causing the computer to function as a keyword extracting means for extracting a keyword from the recognition pattern and the determined candidate pattern, In the language model, the appearance probability of the word string and the coefficient with respect to a keyword similar word that is a non-keyword similar to the keyword have a value that makes the language score of the candidate pattern including the keyword similar word the minimum value.

  The above speech recognition apparatus can be realized by executing the above speech recognition program on a computer. This voice recognition program can be used in a state where it is recorded on a recording medium.

  Hereinafter, preferred embodiments of the present invention will be described with reference to the drawings.

[Basic explanation]
Prior to description of a specific embodiment, a speech recognition method according to the present invention will be described. FIG. 1 is a schematic configuration diagram of a speech recognition apparatus using a language model. A speech recognition apparatus using a language model recognizes an utterance by a user as a combination of words. The process of recognizing an utterance as a combination of words is called “dictation”. By recognizing an utterance as a combination of words, it is possible to recognize a sentence other than a sentence prepared in advance, that is, a sentence formed by arbitrarily combining a plurality of words.

  As shown in FIG. 1, the speech recognition apparatus includes a dictation unit 20 that performs dictation, a language model database 24 that stores language models (hereinafter, “database” may be abbreviated as “DB”), and an acoustic model. An acoustic model DB 25 for storing, a dictionary DB 26 for storing a dictionary, and a keyword extracting unit 30 are provided.

  The dictionary DB 26 stores a large number of words. By referring to the dictionary, the utterance content is recognized as a combination of a plurality of words. As will be described later, keywords stored in the dictionary include keywords.

  The acoustic model DB 25 is a database that stores the characteristics of sounds in units of syllables and phonemes. The sound feature of each word included in the utterance is determined by comparison with the sound feature stored in the acoustic model, and is calculated as an acoustic score. That is, the “acoustic score” is a similarity between a general sound feature for a combination of words and a sound feature corresponding to an utterance, and the general sound feature is generated from an acoustic model.

  FIG. 2 shows an example of a context-dependent acoustic model as one of acoustic models. A “context-dependent acoustic model” refers to an acoustic model in which the characteristics of sound change depending on the syllables before and after the same syllable. In the example of FIG. 2A, even if the character is the same “shi”, the syllables before and after are different, so that the characteristics of the sound are recognized as different. Therefore, as shown in FIG. 2 (b), the acoustic score varies depending on the combination of words even if the sound is the same “mitakashino”.

  The language model DB 24 is a database that stores the appearance probabilities of adjacent word combinations. In the present invention, an N-gram language model which is one of language models is used. The N-gram language model is a database that stores the probability (= N-gram) that a word W appears under the condition that there is a combination of N−1 words immediately before. A language score is calculated using the language model. “Language score” is a value indicating the appearance probability (appearance frequency) of a combination of adjacent words.

  FIG. 3 shows a configuration example of the 2-gram language model. The 2-gram language model stores the probability that the word W appears under the condition that there is one word X immediately before. A combination of words stored in the language model and an appearance probability thereof are created based on a large number of sentence examples.

  In FIG. 3, “1-gram probability P (w1)” indicates the appearance probability of the word w1. “2-gram probability P (w1 | w2)” indicates the appearance probability of the word w1 under the condition that the word w2 comes immediately before. The “backoff coefficient Bo (w2)” is a coefficient prepared for obtaining the appearance probability of a word combination that does not exist in the 2-gram probability. Specifically, when a combination of the word w1 after the word w2, that is, P (w1 | w2) does not exist in the 2-gram probability, P (w1 | w2) is obtained by the following equation using a back-off coefficient. .

P (w1 | w2) = P (w1) × Bo (w2)
Next, an example of calculating a language score using a language model will be described. FIG. 4 shows an example in which the language score is calculated when it is determined that the utterance “Mitakashi no Shogakuko” is a combination of words “Mitaka City / No / Elementary School” (hereinafter referred to as “candidate pattern”). Show. In the following description, the symbol “/” indicates a break when an utterance is divided into words. In the configuration example of the 2-gram language model, “<s>” is a code indicating a sentence head, and “</ s>” is a code indicating a sentence end.

  The dictation unit 20 divides the utterance “Mitaka City Elementary School” into the words “Mitaka City”, “No”, and “Elementary School”, and refers to the 2-gram language model for the language score corresponding to these combinations. get. The product of these language scores is obtained as the language score for the utterance. The example of FIG. 4 is an example in which 2-gram probabilities exist for all word combinations. Referring to the 2-gram probability column in the configuration example of the 2-gram language model, the language score for each word sequence is as follows.

・ "Mitaka City" at the beginning: 0.0000028
・ "Mitaka City" → "no": 0.0001478
・ "No" → "elementary school": 0.0003822
・ "Elementary school" ends the sentence: 0.1060425
Therefore, as shown in FIG. 4, the language score for the candidate pattern “Mitaka city / no / elementary school” is the product of the individual language scores.

  FIG. 5 shows another example. FIG. 5 is a calculation example of a language score when the same utterance “Mitataka no Shogakuko” is determined as a candidate pattern “Mito / Takashino / elementary school”. Similar to the example of FIG. 4, the dictation unit 20 basically obtains the appearance probability of adjacent word combinations from the 2-gram language model, and uses the product as the language score. However, in the example of FIG. 5, there is no 2-gram probability of a combination in which “Takashino” comes after “see”, so this is expressed as a back-off coefficient of “see” and a 1-gram probability of “Takashino”. Created by multiplication. As described above, the language score for the word combination is calculated using the language model.

  The recognition result by dictation is obtained as a combination of words that maximizes the product of the acoustic score and the language score. Specifically, as shown in FIG. 6A, an acoustic score is calculated for each of the words “Mitaka City”, “No”, and “elementary school”, and a language score for the combination thereof is calculated. The Then, the product of the acoustic score and the language score is calculated as a total score. One utterance is recognized as a plurality of candidate patterns, and a total score is calculated for each candidate pattern. Then, a candidate pattern having the maximum total score is output as a recognition result.

  Next, keyword extraction will be described. As shown in FIG. 1, in the present invention, the speech recognition apparatus extracts keywords included in the speech recognition result. Keywords are determined in advance, and information for identifying keywords and non-keywords is stored in the dictionary. For example, among many words stored in the dictionary, a keyword flag indicating the keyword is added. Preferably, an operation command of a device to which the speech recognition method of the present invention is applied is set as a keyword.

  As shown in FIG. 1, the keyword extraction unit 30 refers to the dictionary and extracts a keyword (“Mitaka city” in this example) included in the recognition result “Mitaka city / no / elementary school” obtained by the dictation unit 20. Output. The output keyword is used to operate the device. For this reason, when the keyword cannot be output, the speech recognition apparatus requests the user to speak again. Therefore, the speech recognition apparatus is required to be able to extract keywords with high accuracy in order to prevent repeated speech by the user.

  In order to increase the keyword extraction accuracy, it is preferable that a candidate pattern including the keyword as one independent word is obtained as a recognition result. FIG. 6B shows an example of keyword extraction. FIG. 6B shows the calculation of the total score when the utterance “Mitakano” is divided into candidate pattern 1 “Mitaka city / no” and divided into candidate pattern 2 “Mitaka / Takashino”. To be done. As described above, the dictation unit 20 outputs a candidate pattern having the highest total score as a recognition result. Therefore, in the example of FIG. 6B, it is desired that the candidate pattern 1 has a higher total score than the candidate pattern 2. Therefore, in the present invention, the language score of “Midori / Takashino” is set to the minimum value, that is, “0 (zero)”. Thereby, the total score of the language pattern 2 becomes 0 which is the minimum value, and the candidate pattern 1 is obtained as a recognition result.

  When the total score is calculated by multiplication using the acoustic score and the language score as described above, the minimum value of the score is “0”. On the other hand, when the values of the acoustic score and the language score are logarithmic values as in the examples described later, the minimum value of the score is a negative value having the maximum absolute value.

  In the example of FIG. 6B, “Mitaka City” is a commonly used word, and the sentence example used when creating the language model also includes the word string “Mitaka City”. As shown in FIG. 4, there is a 2-gram probability corresponding to “Mitaka city / no” in the language model. On the other hand, the word string “Midaka Shino” is a word string that is not normally used. As shown in the example of FIG. 5, the 2-gram probability corresponding to “Min / Takashino” in the language model is not exist. Therefore, the language score corresponding to “Mimi / Takashino” is created by the product of the 1-gram probability and the back-off coefficient as shown in FIG. Therefore, in order to set the language score of “Tomi / Takashino” to 0, both the 1-gram probability and the back-off coefficient of the word “Takashino” may be set to 0. As a result, the language score of “Mimi / Takashino” becomes 0, and the candidate pattern 2 does not become a recognition result. That is, in order to accurately extract the keyword “Mitaka City”, the 1-gram probability and backoff coefficient of the non-keyword “Takashino” similar to the word “Mitaka City” may be set to zero. Hereinafter, such non-keywords similar to the keyword are referred to as “keyword similar words”.

For keyword-like words, both the 1-gram probability and the backoff coefficient need to be 0. This will be described. If the candidate pattern corresponding to the utterance includes the word string “Mimi / Takashino” and there is no 2-gram probability corresponding to it, the language score of “Mimi / Takashino” is
The language score of “Mino / Takashino” = the “off” coefficient of “Mino” × 1−gram probability of “Takashino”. Therefore, if the 1-gram probability of “Takashino”, which is a keyword-similar word, is set to 0, the language score of “Tomi / Takashino” can be set to 0.

On the other hand, if the candidate pattern corresponding to the utterance includes the word string “Takashino / Mimi” and there is no 2-gram probability corresponding to it, the language score of “Takashino / Mimi” is
The language score of “Takashino / Tami” = the back-off coefficient of “Takashino” × the 1-gram probability of “Tami”. Therefore, the language score of the word string “Takashino / Tami” cannot be set to 0 unless the backoff coefficient of “Takashino”, which is a keyword similar word, is set to 0. As described above, in order to set the language score to 0 for both the word string that precedes the keyword-similar word and the word string that follows, both the 1-gram probability and the back-off coefficient of the keyword-similar word are set to 0. It will be necessary.

  FIG. 7 shows a configuration example of the 2-gram language model. FIG. 7A is an example in which the present invention is not applied, and the 1-gram probability and the back-off coefficient of the keyword similar word “Takashino” both have certain values. On the other hand, FIG. 7B is an example to which the present invention is applied. As indicated by reference numeral 90, both the 1-gram probability and the back-off coefficient of the keyword similar word “Takashino” are set to zero.

  Next, keyword similar words will be described. The keyword-similar word is a non-keyword similar to the keyword, and includes a word constituting a part of the keyword, a word including a part or all of the keyword, and the like. The word constituting a part of the keyword is, for example, “Mitaka” for the keyword “Mitaka City”. The word including a part of the keyword is, for example, “Takashino” including “Takashi” which is a part of the keyword “Mitaka City” as in the above example. A word including all keywords is a word in which some word (α) is added to the keyword. When a word is added before the keyword, such as “keyword + α”, “α + This includes the case where a word is added after the keyword such as “keyword”. In this case, the word α may be a number.

  As described above, in the present invention, the speech recognition apparatus first creates a candidate pattern by dividing an utterance into a plurality of words, and a total score that is the product of the acoustic score of each word and the language score of each word combination is The largest candidate pattern is taken as the recognition result. Then, the speech recognition device extracts and outputs keywords included in the recognition result. Here, for a keyword similar word similar to a predetermined keyword, both the 1-gram probability and the back-off coefficient, which are appearance probability values used when calculating the language score, are minimum values, that is, “0”. Set to Therefore, the total score of candidate patterns including keyword-similar words is 0 and is not determined as a recognition result. As a result, a candidate pattern including a keyword can be easily obtained as a recognition result, and a keyword being uttered can be extracted with high accuracy.

[Example]
Next, a specific embodiment of the speech recognition apparatus to which the present invention is applied will be described. In the following description, the appearance probability value in the language model is indicated by a logarithm.

(Language model)
FIG. 8 is an example of a 2-gram language model. The language model format of FIG. 8 is called the ARPA format and is widely used. In this ARPA format language model, the number of types of 1-gram probabilities and the number of types of 2-gram probabilities are described at the top, and 1-gram information and 2-gram information are described in order behind the number of types.

  In the 1-gram information, a logarithmic value of an appearance probability, a corresponding word, and a logarithmic value of a back-off coefficient are sequentially described for each word. As described above, the back-off coefficient is a numerical value for outputting a 2-gram probability for a combination of words that does not exist in an actual sentence example. The 2-gram information is described in the order of the logarithmic value of each 2-gram probability, the preceding word, and the succeeding word.

(Language model creation process)
FIG. 9 is a block diagram of the language model creation unit. The language model creation process will be described below with reference to FIG.

  The input of the language model creation unit is language model learning text data. This is text data that has been transcribed as much as possible for possible sentence examples for speech recognition. FIG. 10A shows an example of learning text data.

  The morphological analyzer 11 performs morphological analysis on the learning text data and outputs an analysis text. “Morphological analysis” refers to breaking a sentence into meaningful units such as parts of speech. FIG. 10B shows an example of the text after the morphological analysis is performed on the language model learning text data and the text data is written in units of words. A sentence head symbol <s> is inserted at the beginning of the sentence, and a sentence end symbol </ s> is inserted at the end of the sentence. This is called “analysis text”. In FIG. 10B, numerical values “+1” and “+0” after each word are keyword flags, “+1” indicates a keyword, and “+0” indicates a non-keyword.

  The used vocabulary extractor 12 extracts vocabulary used in the analysis text and generates a used vocabulary list. FIG. 10C is an example of a used vocabulary list. The language model creation unit 13 creates a language model using the analysis text and the used vocabulary list. Here, a temporary language model is generated.

  The rewrite target vocabulary list DB 15 is a tentative language model in which the vocabulary for which the 1-gram probability and the back-off coefficient are set to 0, that is, the above-described keyword similar words are all described. FIG. 10D is an example of a rewrite target vocabulary list. In this example, a keyword-similar word, that is, a word for which the 1-gram probability and the back-off coefficient are to be 0 is “mi + 0”.

  The backoff & 1-gram probability compulsory rewriter 14 forcibly rewrites the 1-gram probability and backoff coefficient possessed by the temporary language model to 0. Specifically, the 1-gram probability and the back-off coefficient for the vocabulary described in the rewrite target vocabulary list DB 15 are set to zero. The language model output here is called a “countermeasure language model”. This countermeasure language model is used for dictation speech recognition.

  An example of a temporary language model is shown in FIG. FIG. 11B shows a countermeasure language model obtained by rewriting this temporary language model using information in the vocabulary list to be rewritten in FIG. In the countermeasure language model shown in FIG. 11B, it can be seen that the 1-gram probability and the backoff coefficient for the word “mi + 0” described in the vocabulary list to be rewritten are different from the temporary language model (box). 92). In the ARPA format used in this embodiment, since the appearance probability value is given as a logarithmic value, the probability of 1-gram and the back-off coefficient for the word “mi + 0” are not “0” but the minimum value, that is, the maximum value. It has been rewritten with a negative value that has an absolute value.

  The rewrite target vocabulary list DB 15 and the backoff & 1-gram probability compulsory rewrite device 14 are characteristic components of the present invention. That is, keyword similar words in a language model including keywords and non-keywords are prepared as a rewrite target vocabulary list, and the backoff & 1-gram probability compulsory rewrite device 14 refers to the rewrite target vocabulary list to 1 Rewrite gram probabilities and backoff coefficients to minimum values. Thus, the countermeasure language model used in the present invention is created. The prepared countermeasure language model is used as the language model in FIG.

(Dictation process)
Next, the dictation process will be described. FIG. 12 is a block diagram showing the configuration of the dictation unit 20. The dictation unit 20 includes a speech segment detection unit 21, a feature parameter calculation unit 22, and a matching processing unit 23.

  “Speech data” refers to an input signal including voice. For example, in the case of a speech recognition device mounted on a car navigation device, the speech data indicates an input signal recorded from a microphone during a certain time after the user presses the speech button.

  The voice section detection unit 21 detects a voice section from speech data and outputs voice data in the voice section. That is, “voice data” refers to data obtained by cutting out only a section corresponding to voice from speech data.

  The feature parameter calculation unit 22 divides the voice data detected by the voice segment detection unit 21 for each unit time, calculates a feature parameter in each unit, and supplies the feature parameter to the matching processing unit 23.

  The dictionary DB 26 stores words that can be recognized by speech, specifically words that appear in the analysis text. Since the speech recognition method of the present invention assumes that a keyword is extracted from a recognition result by dictation processing, it is necessary to distinguish a keyword from a non-keyword. For this reason, the dictionary DB 26 stores flag data for identifying keywords and non-keywords for each registered word. That is, among the words registered in the dictionary DB 26, the keyword is assigned the keyword flag “+1”, and the non-keyword is assigned the keyword flag “+0”.

  The acoustic model DB 25 stores acoustic features in subword units (syllables, phonemes, etc.) expressed by feature parameters.

  The language model DB 25 is an N-gram language model corresponding to a combination of words registered in the dictionary. In the present invention, the countermeasure language model illustrated in FIG. 11B is used as the language model.

  The matching processing unit 23 predicts the utterance content using the language model DB 24, the acoustic model DB 25, and the dictionary DB 26.

  Next, the dictation operation will be described with reference to FIG.

  First, when utterance data is input, the voice section detection unit 21 detects voice data from the data. Next, the feature parameter calculation unit 22 divides the audio data into frames for each unit time, and calculates the feature parameter in each.

  Next, the matching processing unit 23 performs matching processing for outputting a recognition result by applying the feature parameters obtained for each unit time to the language model DB 24, the acoustic model DB 25, and the dictionary DB 26.

  Specifically, the matching processing unit 23 searches the combination of words registered in the dictionary DB 26 for the best match with the voice data in chronological order from the beginning of the voice data. A plurality of candidate patterns are created by this search. Here, as a result of the matching process from the beginning to the middle of the audio data, a pruning process is performed so as not to perform the matching process thereafter for a combination having a low score.

  The matching processing unit 23 calculates an acoustic score and a language score for a plurality of candidate patterns to obtain a total score, and outputs a candidate pattern having the maximum total score among the plurality of candidate patterns as a recognition result. The recognition result output here is a sentence composed of a combination of words registered in the dictionary DB 26. At this time, in this embodiment, as illustrated in FIG. 11B, the 1-gram probability and the back-off coefficient of the keyword similar word are set to the minimum value in the countermeasure language model. The total score of the candidate patterns to be included becomes the minimum value, and as a result, a recognition result from which keywords can be extracted is obtained.

  In the above configuration, the acoustic model DB 25 corresponds to the acoustic model storage unit of the present invention, the language model DB 24 corresponds to the language model storage unit, and the dictionary DB 26 corresponds to the dictionary storage unit. In particular, the matching processing unit 23 of the dictation unit 20 functions as a candidate pattern creation unit, a score calculation unit, and a recognition result determination unit of the present invention.

(Voice recognition processing)
FIG. 13 shows a speech recognition process according to the present invention. As described above, the recognition result created by the dictation unit 20 is supplied to the keyword extraction unit 30. As a result, the keyword extraction unit 30
"Mitakashi + 1 + 0 ginger + 0"
The recognition result is obtained. The number added after the word is the keyword flag. As described above, “1” is assigned to a keyword and “0” is assigned to a non-keyword.

  The keyword extraction unit 30 extracts a keyword from the words included in the candidate pattern output as the recognition result with reference to the dictionary DB 26, and supplies the keyword to the operation control unit that uses the keyword as a command. For example, when the voice recognition device of this embodiment is applied to a navigation device, the keyword extracted by the keyword extraction unit 30 is sent to the operation control unit of the navigation device as an operation command of the navigation device.

(Language score calculation method)
Next, a method for calculating a language score using the N-gram language model will be described in detail. In the following description, the section in which the word w _i in the candidate pattern matches the utterance is simply referred to as “section of w _i ”. The acoustic score for "w _i section of", referred to as the "acoustic score of w _i". Further, the product of the acoustic score and the language score is called a total score.

  Here, it is assumed that the language score is calculated under the following conditions.

    (1) Use the N-gram language model.

    (2) The number of words included in the candidate pattern is R (N <R).

(3) The candidate pattern is composed of w ₁ ,..., W _R.

(W _i refers to those corresponding to out-th word included in the candidate pattern)
The total score S (X) for the candidate pattern X is expressed by the following equation (1).

P (w _j | w _{j−N + 1} ,..., W _j−1 ) is an N-gram probability for the word string w _{j−N + 1} ,..., W _j included in the candidate pattern, and is obtained from the N-gram language model. It is

Here, the word sequence _w 1 from N-gram language model, ..., the flow of obtaining N-gram probability of the _{w n,} will be described with reference to FIG. 14.

First, a word string _w 1 at step S201, ..., determine whether N-gram probability of _{w n} is present. If it exists, this N-gram probability is output.

If it does not exist if, in step S202 and S203, the word sequence _w 2, ..., in _{w n} (N-1) -gram probability and the word sequence _w 1, _..., in _{w n-1 (N-1} ) -gram Check whether each probability exists. If both present, the word sequence _w 2, ..., in _{w n} (N-1) -gram probability and the word sequence _w 1, _..., output the product of the back-off factor in the _{w n-1} as N-gram probability To do.

If you do not want to present even one of them, a word string w ₂ to _enter, ..., replaced by a w _n, to examine the presence of again (N-1) -gram probability.

(Effects of this embodiment)
Generally, the dictation unit outputs a combination of words registered in the dictionary as a recognition result. Therefore, a plurality of correct answers can be considered depending on the utterance. When a context-dependent acoustic model is used, the acoustic score changes depending on the combination of words.

  Take, for example, “Mitaka City Elementary School” mentioned above. Here, “Mitaka City” is a keyword. In addition, it is assumed that “Takashino” and “Mi” are also registered in the dictionary. In this case, the following two candidate patterns can be considered as correct answers.

Candidate pattern 1: Mitakashi + 1 No + 0 Ginger + 0
Candidate pattern 2: Mi + 0 Takashino + 0 Ginger + 0
When candidate pattern 1 is recognized as a recognition result, the keyword “Mitaka City” can be extracted. Therefore, a device such as a navigation device to which the voice recognition device of the present invention is applied can extract “Mitaka City” as a keyword. However, when the candidate pattern 2 is a recognition result, there is a problem that no keyword is present in the recognition result, and thus the device cannot be operated.

Here, in the analysis text,
・ There is no sentence that comes with “Takashino” after “seeing” ・ As long as there is a sentence that comes with “no” after “Mitaka City”, the language score is basically candidate pattern 1. Get higher. However, the acoustic score may be higher in the candidate pattern 2 due to the utterance. As a result, there is a possibility that the candidate pattern 2 also becomes high in the total score. If this happens, it is possible that “Takashino Elementary School” will be recognized despite the fact that “Mitaka City” should be recognized.

  A specific example will be explained using the language model of FIG. The numerical value of the language model described in FIG. 8 is described in logarithm.

First, logarithmic language scores P _LM1 and P _LM2 of candidate pattern 1 and candidate pattern 2 are calculated.

Here, in candidate pattern 2, there is no 2-gram probability that “Takashino” comes after “Mi”. Therefore, this 2-gram probability is the product of the back-off coefficient for “Mi” and the 1-gram probability of “Takashino”. Therefore, the candidate score 1 is higher in the language score itself. However, there are cases where the candidate pattern 2 is higher in the acoustic score. For example, it is assumed that the logarithmic acoustic scores P _AM1 and P _{AM2 of} candidate pattern 1 and candidate pattern 2 are P _AM1 = −20.0 and P _AM2 = −19.0, respectively. In this case, each log total score _PR1 , _PR2 is

Thus, the candidate pattern 2 that does not exist in the analysis text is surpassed. Therefore, keyword extraction is impossible.

  On the other hand, an example of a 2-gram language model created in the present invention is shown in FIG. As described above, the backoff coefficient and 1-gram probability of the keyword similar word are set to the minimum value (in this example, the negative value having the maximum absolute value). Specifically, the back-off coefficient and 1-gram probability of the non-keyword “mi”, which is a part of the correct keyword, are minimized.

After this countermeasure is taken, a logarithmic language score P _LM2 ′ of candidate pattern 2 is calculated.

As a result, the logarithmic total score _PR2 '

It becomes. Incidentally, since the backoff coefficient is not used in the candidate pattern 1, the logarithmic total score is still P _R1 = −33.77021. Therefore, when the countermeasure language model created in the present invention is used, the candidate pattern 1 is taken as the recognition result, and the keyword “Mitaka city” is correctly extracted.

[Application field]
The present invention can be applied to various devices that perform voice recognition processing. For example, the present invention can be applied to various devices having a voice input function such as a car navigation device, a mobile phone, a personal computer, an AV device, and a home appliance.

It is a schematic explanatory drawing of the speech recognition method using a language model. An example of a context-dependent acoustic model is shown as one of acoustic models. The structural example of a 2-gram language model is shown. An example of language score calculation is shown below. An example of language score calculation is shown below. It is a figure explaining the calculation method of a total score. The structural example of a 2-gram language model is shown. An example language model is shown. It is a block diagram which shows a language model creation process. Examples of learning text, analysis text, vocabulary list used, and vocabulary list to be rewritten are shown. An example language model is shown. It is a block diagram which shows the structure of a dictation part. The flow of voice recognition processing is shown. It is a flowchart which shows the calculation method of a language score. An example language model is shown.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 Morphological analyzer 12 Vocabulary extractor 13 Language model generator 14 Backoff & 1-gram probability compulsory rewriter 15 Rewrite object vocabulary list DB
20 dictation unit 21 speech segment detection unit 22 feature parameter calculation unit 23 matching processing unit 24 language model DB
25 Acoustic model DB
26 Dictionary DB

Claims (7)

An acoustic model storage unit for storing an acoustic model;
A dictionary storage unit for storing a dictionary having a plurality of words including keywords;
For a plurality of words, a language model storage unit that stores a language model having a coefficient for calculating an appearance probability of a word string including the word and an appearance probability of the word string including the word;
A candidate pattern creation unit that creates a candidate pattern composed of a combination of a plurality of words from the utterance content;
For each of the candidate patterns, a score calculator that calculates an acoustic score based on the acoustic model and calculates a language score based on the language model, and calculates a total score based on the acoustic score and the language score;
A recognition result determination means for determining a candidate pattern having the maximum total score among a plurality of candidate patterns as a recognition result;
A keyword extracting means for referring to the dictionary and extracting a keyword from the recognition result and the determined candidate pattern;
In the language model, the appearance probability and the coefficient of the word string with respect to a keyword similar word that is a non-keyword similar to a keyword have a value that minimizes the language score of a candidate pattern including the keyword similar word. Voice recognition device. The score calculation unit calculates a product of the acoustic score and the language score as the total score,
The speech recognition apparatus according to claim 1, wherein the appearance probability of the word string and the coefficient for the keyword-similar word are zero.   The speech recognition apparatus according to claim 2, wherein the language model is an N-gram language model, the appearance probability of the word string is a 1-gram probability, and the coefficient is a back-off coefficient.   The speech recognition apparatus according to any one of claims 1 to 3, wherein the keyword-similar word includes a word that forms part of the keyword and a word that includes part or all of the keyword. An acoustic model, a dictionary having a plurality of words including a keyword, and a language model having a coefficient for calculating an appearance probability of a word string including the word and an appearance probability of the word string including the word for the plurality of words A speech recognition method using
A candidate pattern creating step for creating a candidate pattern composed of a combination of a plurality of words from the utterance content;
For each of the candidate patterns, a score calculation step of calculating an acoustic score based on the acoustic model and calculating a language score based on the language model, and calculating a total score based on the acoustic score and the language score;
A recognition result determination step of determining a candidate pattern having the maximum total score among a plurality of candidate patterns as a recognition result;
A keyword extraction step of referring to the dictionary and extracting a keyword from the recognition result and the determined candidate pattern,
In the language model, the appearance probability and the coefficient of the word string with respect to a keyword similar word that is a non-keyword similar to a keyword have a value that minimizes the language score of a candidate pattern including the keyword similar word. Voice recognition method. A speech recognition program executed by a computer,
An acoustic model storage unit for storing an acoustic model;
A dictionary storage unit for storing a dictionary having a plurality of words including a keyword;
A language model storage unit that stores a language model having a coefficient for calculating an appearance probability of a word string including the word and an appearance probability of the word string including the word for a plurality of words,
A candidate pattern creation unit that creates a candidate pattern composed of a combination of a plurality of words from the utterance content,
For each of the candidate patterns, a score calculation unit that calculates an acoustic score based on the acoustic model and calculates a language score based on the language model, and calculates a total score based on the acoustic score and the language score;
A recognition result determining means for determining a candidate pattern having the maximum total score among a plurality of candidate patterns as a recognition result; and
Referencing the dictionary storage unit, causing the computer to function as keyword extraction means for extracting a keyword from a recognition result and a determined candidate pattern,
In the language model, the appearance probability and the coefficient of the word string with respect to a keyword similar word that is a non-keyword similar to a keyword have a value that minimizes the language score of a candidate pattern including the keyword similar word. Voice recognition program.   A recording medium on which the voice recognition program according to claim 6 is recorded.

Images