JP5276610B2 - Language model generation apparatus, program thereof, and speech recognition system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a language model generating device for generating a language model for precisely recognizing voice. <P>SOLUTION: The language model generating device 1 includes: a learning text storage part 11 for storing a learning text; an equivalent word and chain word selection part 21; a language model generating part 22 for generating the language model indicating at least one appearance probability of a word or a chain word included in the learning text; a language model conversion part 23 for calculating a probability value on the basis of the appearance probability of a language model in synonyms having the same meaning and updating the appearance probability of the synonyms at the probability value; a pronunciation dictionary storage part 17 for storing a pronunciation dictionary; and a pronunciation dictionary conversion part 24 for converting the pronunciation dictionary. <P>COPYRIGHT: (C)2011,JPO&amp;INPIT

Description

  The present invention relates to a technique for generating a language model using a probabilistic language model, and a technique for performing speech recognition using the language model.

  For example, speech recognition using a language model is indispensable in mechanical operation by voice (car navigation device), automatic voice guidance system, or real-time caption generation in a broadcast program. As described above, since the language model is important for determining the accuracy of speech recognition, several conventional techniques related to this language model have been proposed.

  A general language model currently used in speech recognition is called a word N-gram model (see Non-Patent Document 1, for example). Moreover, there is a class N gram model using the concept of word class as a language model that is an extension of this word N gram model. This class N-gram model is an N-gram model handled as a class in which words are classified according to parts of speech and concepts, and is effective when learning cannot be sufficiently performed with the word N-gram model because there are few learning texts.

  Further, as a technique related to a language model other than the N-gram model, the invention described in Patent Document 1 has been proposed. In the invention described in Patent Document 1, a paraphrase pattern (rule) is probabilistically estimated (modeled) using a word list of official names and the paraphrase word list. The invention described in Patent Document 1 adds a paraphrase pattern to the dictionary (language model) in addition to the official name. Furthermore, the invention described in Patent Document 1 calculates how close a short input speech (for example, proper nouns such as facility names and place names) is to the waveform of a word registered in this dictionary. Output the recognition result. That is, the invention described in Patent Document 1 performs isolated word recognition that recognizes an input speech including one word.

Japanese Patent Laid-Open No. 2005-3255

Stochastic language model, University of Tokyo Press, pp. 60-62 and pp. 72-75

However, the N-gram model has the following problems.
There are various notations and readings corresponding to the same expression in Japanese that can be seen mainly in spoken language. For example, “to” often becomes an expression such as “to” or “tchu”. On the other hand, in the written language, all of these are unified as “to”.
In addition, when using a learning text in which a spoken word is transcribed, “to” may be replaced with “to” or “chu”, for example.
As a result, in the language model, for example, the statistic “NO” is dispersed, the amount of learning text becomes insufficient, and a reliable probability value may not be calculated.

  In addition, the learning text may include notation fluctuations so that, for example, “dealing” can be described as “dealing”. In this case, the probability value of “dealing” may be calculated as a small value due to the fluctuation of the expression.

  To summarize the above, the N-gram model is such that when there are words or chain words that have the same meaning or different readings, such as spoken words, written words, and fluctuations in notation, the statistics of these words or chain words are calculated. scatter. For this reason, the learning amount of the N-gram model is relatively short, and there is a problem that the reliability of the probability value becomes low in the generated language model.

  Further, since the invention described in Patent Document 1 performs isolated word recognition, a dictionary is not generated in consideration of the context. For this reason, the invention described in Patent Document 1 cannot cope with large vocabulary continuous speech recognition (speech recognition of input speech including a plurality of words) in which context is very important.

Accordingly, an object of the present invention is to provide a language model generation apparatus and program for generating a language model that solves the above-described problems and enables speech recognition with few recognition errors.
Furthermore, another object of the present invention is to provide a speech recognition system that solves the above-described problems and enables speech recognition with few recognition errors.

In order to solve the above-described problem, the language model generation device according to the first invention of the present application generates a language model by using learning text including synonyms composed of words or chain words having the same meaning or different notation or reading. A language model generation device, comprising a language model generation unit, a chain word extraction unit, an edit distance calculation unit, a minimum edit distance selection unit, a consent word / chain word list generation unit, and a language model conversion unit It is characterized by that.

  According to this configuration, the language model generation device performs learning with the learning text in the language model generation unit, thereby indicating a language model (for example, an appearance probability of at least one of words or chain words included in the learning text) , N-gram model). That is, the language model generation unit generates a probabilistic language model considering the context.

Here, as described above, in the language model generated as a probabilistic language model, when there are synonyms in the learning text, the statistics of these synonyms are dispersed, and the occurrence probability of these synonyms is low. Become. Therefore, the language model generation device extracts word pairs that appear more than a preset frequency in the learning text as the chain words in the order in which the entropy per word of the learning text is most reduced by the chain word extraction unit. To do. Further, the language model generation apparatus calculates the edit distance of the chain word extracted by the chain word extraction means by DP matching by the edit distance calculation means. Furthermore, the language model generation apparatus selects, as a synonym candidate, a chain word that has the minimum edit distance calculated by the edit distance calculation unit by the minimum edit distance selection unit. Further, the language model generation device receives a selection instruction including a chain word selected in advance from the synonym word candidates by the synonym word / chain word list generation unit, and has the same meaning based on the selection instruction A synonym list in which the synonyms are associated in advance is generated. Furthermore, the language model generating device, the language model conversion unit, said with reference to the agreed word list, to calculate a probability value based on the probability of occurrence of synonyms having the same meaning in the language model, the language The appearance probability of the synonym included in the model is updated with the probability value. That is, the language model conversion unit corrects the appearance probability of the synonym calculated with a low value due to the presence of the synonym in the learning text.

Thus, the language model generation device allows the user to select synonyms for the chain words by referring to the synonym candidates in the synonym list.

Moreover, in the language model generation device according to the second invention of the present application, the language model conversion unit converts the synonym synonym type other than the basic type of the synonym with the highest occurrence probability into the language model conversion unit. Language model deleting means for deleting from the updated language model.
According to such a configuration, the language model generation device can reduce the data size of the language model.

A language model generation device according to a third invention of the present application includes a pronunciation dictionary storage unit that stores a pronunciation dictionary in which at least the synonym notation and the pronunciation of the synonym are associated in advance, the pronunciation dictionary, the synonym list The phonetic to be converted into a converted pronunciation dictionary including at least a synonym notation of the synonym, a synonym notation of the synonym corresponding to the basic type, and a synonym pronunciation of the synonym And a dictionary conversion unit.
According to such a configuration, the language model generation device converts the pronunciation dictionary into a converted pronunciation dictionary in which basic type notation, consent type notation and pronunciation are associated with each other.

In order to solve the problems described above, the language model generating program according to the present fourth invention, the computer, characterized in that to function as a language model generating apparatus according to the first aspect of the present invention.

In order to solve the above-described problem, the speech recognition system according to the fifth invention of the present application performs speech recognition using the language model generation device according to the third invention of the present application and the language model generated by the language model generation device. A speech recognition system including a speech recognition device, wherein the speech recognition device includes an acoustic model storage unit that stores an acoustic model generated in advance by learning speech data, a speech analysis unit, and a search unit. It is characterized by.

  According to this configuration, the speech recognition apparatus performs speech analysis on the input speech input by the speech analysis unit, and calculates a feature vector of the input speech. In the speech recognition apparatus, the search unit calculates an acoustic score by matching the feature vector calculated by the speech analysis unit with the acoustic model, and becomes a speech recognition result candidate with reference to the language model. A language score is calculated by adding a second constant to a value obtained by multiplying the appearance probability of the word candidate by the first constant, and a column of word candidates that maximizes the language score and the acoustic score is calculated after the conversion. The phonetic dictionary is referenced and output as a result of the speech recognition. That is, the speech recognition apparatus can output the consent type notation and pronunciation corresponding to the basic type by referring to the converted pronunciation dictionary.

According to the present invention, the following excellent effects can be obtained.
According to the first and fourth inventions of the present application, since the probabilistic language model considering the context is generated, it is possible to cope with large vocabulary continuous speech recognition. According to the first invention of this application, since the synonym is dispersed due to the presence of the synonym in the learning text and the appearance probability of the synonym calculated with a low value is corrected, even if the learning text is small, a recognition error occurs. It is possible to generate a language model that enables voice recognition with less.

According to the first and fourth inventions of the present application, since the user can select synonyms for the chain words by referring to the synonym candidates in the synonym list, the user can compare with the case where the synonym list is not presented. The trouble of selecting synonyms can be greatly reduced.

According to the second aspect of the present invention, since the data size of the language model can be reduced, the memory capacity of the speech recognition apparatus using the language model can be saved.
According to the third invention of the present application, the converted pronunciation dictionary is associated with the basic type notation, the consent type notation, and the pronunciation, so by referring to the converted pronunciation dictionary, the consent type corresponding to the basic type Can be easily output.

According to the fifth aspect of the present invention, since the probabilistic language model considering the context is generated, it is possible to cope with large vocabulary continuous speech recognition. According to the fifth aspect of the present invention, since the synonym is dispersed due to the presence of the synonym in the learning text and the appearance probability of the synonym calculated with a low value is corrected, even if the learning text is small, a recognition error is caused. Enables voice recognition with less. Furthermore, according to the fifth invention of the present application, by referring to the converted pronunciation dictionary, it is possible to easily output the consent type notation and pronunciation corresponding to the basic type, thereby improving the convenience of the speech recognition system. be able to.

It is a block diagram which shows the structure of the speech recognition system which concerns on embodiment of this invention. It is a block diagram which shows the structure of the consent word and chain word selection part of FIG. It is a figure which shows an example of the chain word list which the chain word extraction means of FIG. 2 produced | generated. It is a figure which shows an example of the edit distance minimum chain word relation list | wrist which the minimum edit distance selection means of FIG. 2 produced | generated. 3 is an example of a consent word / chain word list stored in a consent word / chain word list storage unit in FIG. 2. It is a block diagram which shows the structure of the language model conversion part of FIG. It is a figure which shows an example of the language model which the language model memory | storage part of FIG. 1 memorize | stores, (a) is a case of a unigram, (b) is a case of a bigram. It is a figure which shows an example of the speech dictionary which the speech dictionary memory | storage means of FIG. 1 memorize | stores. It is a figure which shows an example of the post-conversion utterance dictionary which converted the utterance dictionary of FIG. It is a flowchart which shows operation | movement of the consent word and chain word selection part of FIG. It is a flowchart which shows operation | movement of the language model conversion part of FIG.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings as appropriate. In each embodiment, means having the same function are denoted by the same reference numerals and description thereof is omitted.

[Outline of speech recognition system]
An outline of a speech recognition system according to an embodiment of the present invention will be described with reference to FIG.
As shown in FIG. 1, the speech recognition system 100 includes a language model generation device 1 and a speech recognition device 3.
The language model generation device 1 generates a language model used for speech recognition using learning text including synonyms consisting of words or chain words that have the same meaning and different notation or reading.
The speech recognition device 3 recognizes the input speech using the language model generated by the language model generation device 1.

  Note that a chain word (word chain) is a frequent expression composed of a plurality of words, and combines words in one word chain and handles them as one word. For example, as a chain word, there is “to_” which is a concatenation of the word “to” and the word “to”. Here, “_” indicates connection between words.

[Configuration of language model generator]
Hereinafter, the configuration of the language model generation device 1 will be described in detail.
As shown in FIG. 1, the language model generation device 1 includes a learning text storage unit 11, a chain word list storage unit 12, an edit distance addition chain word relationship list storage unit 13, and an edit distance minimum chain word relationship list storage unit. 14, consent word / chain word list storage unit 15, language model storage unit 16, pronunciation dictionary storage unit 17, converted pronunciation dictionary storage unit 18, and consent word / chain word selection unit (consent word selection unit) 21, a language model generation unit 22, a language model conversion unit 23, and a pronunciation dictionary conversion unit 24.

[Each storage unit]
The learning text storage unit 11 is a storage unit such as a memory or a hard disk for storing learning text (learning data) necessary for generating a language model. This learning text is stored in advance in the learning text storage unit 11 by a user of the speech recognition system, for example.

The chain word list storage unit 12 is a storage unit such as a memory or a hard disk for storing the chain word list.
The edit distance assignment chain word relation list storage unit 13 is a storage unit such as a memory or a hard disk for storing the edit distance assignment chain word relation list.
The edit distance minimum chain word relationship list storage unit 14 is a storage unit such as a memory or a hard disk for storing the edit distance minimum chain word relationship list.
The consent word / chain word list storage unit 15 is a storage unit such as a memory or a hard disk for storing the consent word / chain word list (synonym list).
The details of the chain word list, the edit distance imparted chain word relation list, the edit distance minimum chain word relation list, and the consent word / chain word list will be described together with the consent word / chain word selection unit 21.

  The language model storage unit 16 is a storage unit such as a memory or a hard disk that stores a language model. This language model indicates the appearance probability of at least one of words or chain words included in the learning text, and is generated by a language model generation unit 22 described later.

The pronunciation dictionary storage unit 17 is a storage unit such as a memory or a hard disk that stores the pronunciation dictionary in advance. This pronunciation dictionary is a dictionary that associates the notation of words or chain words with their pronunciation.
The post-conversion pronunciation dictionary storage unit 18 is a storage unit such as a memory or a hard disk that stores the post-conversion pronunciation dictionary. This post-conversion pronunciation dictionary is obtained by converting the pronunciation dictionary by the pronunciation dictionary conversion unit 24 described later, and is referred to when the speech recognition device 3 performs speech recognition.
Details of the pronunciation dictionary and the converted pronunciation dictionary will be described together with the pronunciation dictionary conversion unit 24.

[Consent word / chain word selection part]
Hereinafter, the consent word / chain word selection unit 21 will be described in detail with reference to FIG.
The consent word / chain word selection unit 21 refers to the learning text, and generates a consent word / chain word list based on the selection instruction input by the user. Here, as shown in FIG. 2, the consent word / chain word selection unit 21 includes a chain word extraction unit 211, an edit distance calculation unit 212, a minimum edit distance selection unit (chain word candidate selection unit) 213, and an agreement A word / chain word list generation unit 214.

The chain word extraction unit 211 performs a bigram for calculating the appearance probability of two consecutive words in order from the beginning of the learning text. For example, if the learning text is “Kyo_ was _Warm_Morning_”, the word pairs would be “Kyo_ha”, “was_warm” and “Morning_”. In this case, since the word pair “Kyo_ha”, “ha_warm”, and “was morning_” appeared once each, the chain word extraction unit 211 sets the appearance probability “1” of each of these word pairs. calculate. The chain word extraction unit 211 extracts, as a chain word, word pairs that are equal to or higher than a preset frequency (threshold value) and appear in the learning text in an order that reduces the entropy per word of the learning text most. To do. Thereafter, the chain word extraction unit 211 generates a chain word list in which the extracted chain words are stored, and stores them in the chain word list storage unit 12.
Details of the method for extracting chain words can be found in, for example, the document “Language model using chain words and classes for conversational speech, Proceedings of the Society of Musical Engineers of Japan, pp. 71-72, March 2006”. Have been described.

Here, an example of a chain word list will be described with reference to FIG.
In the chain word list of FIG. 3, chain words extracted from the learning text by the chain word extraction means 211 (for example, “Yes_Is”, “I_No”, “I_I” ”) are respectively stored. Yes.

Returning to FIG. 2, the explanation of the consent word / chain word selection unit 21 will be continued.
The edit distance calculation means 212 refers to the chain word list, and calculates the edit distance of the chain word included in the chain word list by DP matching. Specifically, the edit distance calculation means 212 performs DP matching for each chain word on the N-1 chain words excluding the chain word and the word unit constituting the chain word, and creates a chain. Calculate the edit distance of the word. Then, the edit distance calculation means 212 generates an edit distance-added chain word relationship list storing each chain word and the edit distance of each chain word, and stores it in the edit distance-added chain word relationship list storage unit 13.
The details of DP matching are described in, for example, the document “Pattern Recognition and Learning Algorithm, Sentence 1 General Publishing, pp. 91-108”.

  The minimum edit distance selection means 213 refers to the edit distance assignment chain word relation list, and selects (lists up) as a synonym candidate the link word having the shortest edit distance included in the edit distance assignment chain word relation list. Then, the minimum edit distance selection means 213 generates an edit distance minimum chain word relationship list storing the selected synonym candidates (chain words), and stores it in the edit distance minimum chain word relationship list storage unit 14.

Here, an example of the edit distance minimum chain word relation list will be described with reference to FIG. 4 (see FIG. 2 as appropriate).
As shown in FIG. 4, the minimum edit distance selection means 213 displays the edit distance minimum chain word relationship list with, for example, a line beginning with an identifier “basic type” indicating a basic type and an identifier “consent type” indicating a consensus type. A format in which the starting line appears alternately. At this time, the minimum edit distance selection means 213 includes one chain word in the line starting with “basic type” and one or more in the line starting with “consent type” in the edit distance minimum chain word relation list. Include chain words. Further, the minimum edit distance selection means 213 selects each chain word of the line starting with “consent type” in the edit distance minimum chain word relation list with respect to the chain word of the line starting with “basic type” on the one line. , The editing distance is minimized. For example, in the edit distance minimum chain word relation list in FIG. 4, two chain words “Ne____________” and “Must___________________________________________________ This indicates that the edit distance is the minimum for the chain word “must_do_no” in the line starting with “basic type”.

  Here, it is possible that the edit distance minimum chain word relation list includes a word that cannot be said to be a synonym for a chain word of a line starting with “basic type” in a chain word of a line starting with “consent type”. There is sex. For this reason, it is preferable that the user checks the edit distance minimum chain word relation list and selects only a chain word having no problem as a synonym from among the chain words of the line starting with “consent type”. Then, the user inputs the chain word selected by the user based on the edit distance minimum chain word relation list to the consent word / chain word list generation unit 214 as a selection instruction.

  This selection instruction indicates at least one of a word selected as a synonym and a chain word. That is, the selection instruction may include a word previously selected by the user in addition to the chain word. Further, when only one of a chain word and a word is used as a synonym, the selection instruction may include only the one.

Hereinafter, returning to FIG. 2, the explanation of the consent word / chain word selection unit 21 will be continued.
The consent word / chain word list generation means 214 receives a selection instruction from the user, and generates a consent word / chain word list based on the selection instruction. That is, the consent word / chain word list generation unit 214 stores the synonyms included in the selection instruction and generates a consent word / chain word list. Then, the consent word / chain word list generation unit 214 stores the generated consent word / chain word list in the consent word / chain word list storage unit 15.

Here, an example of the consent word / chain word list will be described with reference to FIG.
The synonym word / chain word list in FIG. 5 includes a chain word selected as a synonym from the edit distance minimum chain word relation list in FIG. Also, this synonym word / chain word list is similar to the edit distance minimum chain word relation list in FIG. 4 by inserting identifiers of “basic type” and “consent type” at the beginning of the line, and synonyms (same meaning). And a chain word having the same meaning). That is, the synonym word / chain word list indicates that the chain word of the line starting with “basic type” and the chain word of the line starting with “consent type” one line below are synonyms. For example, the consensus word / chain word list of FIG. 5 includes a chain word “must_do_no” in a line starting with “basic type” and a chain word “naki_” in a line starting with “consent type” one line below. "I don't have" is a synonym. 4 is included in the synonym word / chain word list because it is determined by the user that it is not a synonym for the chain word “must_must_must”. I can't.
Although omitted in FIG. 5, it goes without saying that the synonym word / chain word list may be associated with words having the same meaning as synonyms, similarly to the chain words.

In summary, the consent word / chain word selection unit 21 selects a word pair having a high appearance frequency in the learning text, and selects the word pair that greatly reduces the entropy of the learning text among these word pairs. Synonym candidate). The synonym word / chain word selection unit 21 repeats this process as necessary, and obtains M chain words (synonym word candidates) to generate a synonym word / chain word list. Thereby, for example, the user can select synonyms for the chain words by referring to the synonym candidates in the synonym word / chain word list, so that the user's effort to select the synonyms can be greatly reduced. it can. If the user cannot refer to the synonym word / chain word list, the user needs to perform a difficult task of directly selecting the synonym word from the learning text for the chain word, which greatly increases the burden. Here, the value of M is set in advance so as to increase the accuracy of speech recognition.
For words, a list corresponding to the minimum edit distance chain word relation list is not generated because the user has less time to select synonyms than the chain words.

[Language model generator]
Returning to FIG. 1, the description of the language model generation apparatus 1 will be continued.
The language model generation unit 22 generates a language model by learning the learning text using a probabilistic language model (machine learning), and stores the language model in the language model storage unit 16. Here, the language model generation unit 22 uses a word N-gram model as the probabilistic language model. This word N-gram model, a word string _w ¹ n _{= w} ¹ that is included in the learning text,..., Against _{w n,} stochastic to predict the probability of occurrence of word _{w n,} from the N-1 last word It is a language model and can be expressed by the following equation (1).

This word N-gram model is called a unigram when N = 1, called a bigram when N = 2, and when N = 3, It is called a trigram. Also, the immediately preceding N-1 word (w ₁ ^n-1 ) is called a history.

Here, an example of a language model will be described with reference to FIG. 7 (see FIG. 1 as appropriate).
In the language model of FIG. 7, the left column is an N-gram probability value (appearance probability), the center column is a parameter name (word or chain word), and the right column is a back-off coefficient.

The N-gram probability value indicates the appearance probability of the word or chain word described in the parameter name, and in FIG. 7, the value is expressed in logarithm.
The parameter name indicates a word or a chain word. In addition, <s> in the parameter name is a sentence head symbol, and </ s> is a sentence end symbol. That is, in this language model, the beginning symbol and the end symbol are handled as words.
The back-off coefficient is a coefficient that is used when an N-gram probability value that has an appearance probability of zero in the learning text is estimated from a low-order word N-gram model (for example, a unigram).
The details of the back-off coefficient will be described together with back-off coefficient processing means 234 described later.

[Language model converter]
Hereinafter, returning to FIG. 6, the language model conversion unit 23 will be described in detail.
The language model conversion unit 23 performs conversion (correction) of the language model with reference to the consent word / chain word list. Here, as shown in FIG. 6, the language model conversion unit 23 includes a parameter extraction unit 231, a probability value calculation unit 232, a history processing unit 233, a backoff coefficient processing unit 234, and a language model update unit 235. Is provided.

  The parameter extraction unit 231 refers to the synonym word / chain word list and extracts N-gram parameters of the synonym words included in the synonym word / chain word list from the language model. Then, the parameter extracting unit 231 outputs the extracted N-gram parameter to the probability value calculating unit 232. In the following description, the N-gram parameter refers to the N-gram probability value, parameter name, and back-off coefficient of the language model.

  The probability value calculation means 232 receives the N-gram parameter from the parameter extraction means 231. Further, the probability value calculation means 232 refers to the synonym word / chain word list and acquires synonyms having the same meaning associated with the synonym word / chain word list. Then, the probability value calculation unit 232 calculates a probability value for the synonym having the same meaning based on the appearance probability of the input N-gram parameter. Here, the probability value calculation means 232 can calculate a probability value by performing four arithmetic operations for obtaining an added value or the like for the appearance probability of the input N-gram parameter. In addition, the probability value calculation unit 232 can calculate a probability value by performing a statistical calculation for obtaining an average value, a maximum value, and the like for the appearance probability of the input N-gram parameter. Further, the probability value calculation means 232 calculates, as the probability value, any one of the added value (method 1), the average value (method 2), and the maximum value (method 3) for the appearance probability of the input N-gram parameter. Is preferred. Hereinafter, six specific examples of calculating probability values will be described in order.

<First example: Trigram technique 1>
First, as a first example to a third example, a specific example will be described when Method 1 to Method 3 are applied to a trigram in which the chain word w _k appears next to the chain word strings w _i and w _j .
As a result of clustering for each chain word having the same meaning, N chain word classes {C ₁ ,..., C _N } are obtained, and in a certain class C _n (where 1 ≦ n ≦ N), K _n Suppose that there are +1 chain words with the same meaning. In this case, of the K _n +1 chain words, the one with the highest appearance probability is set as the basic type, and the other is set as the consensus type (the word also has the basic type and the consensus type). That is, K _n +1 chain words are represented by the following formula (2).

In this first example, it is interpreted that chain words representing the same meaning appear in a distributed manner in a basic type and a consent type. Therefore, the probability value calculation means 232 calculates the addition value of the appearance probability of a chain word as a probability value using the following formula (3).
In Equation (3), S _n (κ) represents the κ-th chain word in class C _n .

<Second example: Trigram method 2>
In this second example, it is interpreted that the basic type and the consent type appear with an equal probability. Therefore, the probability value calculating means 232 calculates the average value of the appearance probability of the chain word as a probability value using the following formula (4).

<Third example: Trigram technique 3>
In this third example, the probability value calculating means 232 simplifies the methods 1 and 2 and calculates the maximum value of the appearance probability of a chain word as a probability value using the following equation (5). That is, the probability value calculation means 232 replaces the consent type appearance probability with the basic type appearance probability.

<Fourth Example: Method 1 with Unigram>
Subsequently, as a fourth example to a sixth example, specific examples when the methods 1 to 3 are applied to a unigram will be described. In the fourth to sixth examples, it is assumed that the synonyms and their appearance probabilities have the following relationship. Also, in these 4th to 6th examples, the synonym “n is _ but is” is the basic type, and other synonyms are “n is _ but”, “n is _ but” and “n is _”. Is the consent type.

<< Synonyms and their appearance probabilities in the fourth to sixth examples >>
Synonym occurrence probability is _ but 0.4
It is __ but 0.3
But it is 0.2
Is it __

  In this fourth example, the probability value calculating means 232 sets the value obtained by adding the appearance probabilities of synonyms as the probability value, as in the first example. That is, the probability value calculation means 232 performs a calculation “0.4 + 0.3 + 0.2 + 0.1 = 1.0”. Therefore, the probability value of each synonym is as follows.

<< Probability value calculated in the fourth example >>
Synonym is the probability value__ but 1.0
It is _ but 1.0
I'm 1.0
It is __

<Fifth example: Unigram method 2>
In the fifth example, the probability value calculating means 232 sets the probability value to a value obtained by averaging the appearance probabilities of synonyms as in the second example. That is, the probability value calculation means 232 performs a calculation of “(0.4 + 0.3 + 0.2 + 0.1) /4=0.25”. Therefore, the probability value of each synonym is as follows.

<< Probability value calculated in the fifth example >>
Synonym Probability Value _ but 0.25
_ But 0.25
_ But 0.25
It is _ 0.25

<Sixth example: Method 3 with unigram>
In the sixth example, the probability value calculating means 232 calculates the maximum value “0.4” of the appearance probability in the synonym as in the third example. Therefore, the probability value of each synonym is as follows.

<< Probability value calculated in the sixth example >>
Synonym Probability value _ but 0.4
But it is 0.4
But it is 0.4
It is 0.4

  After that, the probability value calculating unit 232 updates the appearance probability included in the N-gram parameter input from the parameter extracting unit 231 with the calculated probability value. Then, the probability value calculation unit 232 outputs the N-gram parameter updated with the probability value to the history processing unit 233.

  Note that the probability value calculation means 232 may use any method to set the probability value, and for example, may set in advance which method is used to calculate the probability value. Further, the probability value calculation means 232 can calculate the probability value for the word as well as the chain word.

  The history processing means 233 performs history processing when an N-gram parameter is input from the probability value calculation means 232 and a word having a consent type exists in the history. Here, in order to simplify the description of the history processing, it is assumed that the number of consent type patterns K = 1, that is, one consent type exists for the basic type. At this time, the language model is assumed to be a bigram.

The probability that the word w _n appears in the next word w _n-1 in the training text, can be represented by the following formula (6).
In Equation (6), C (•) indicates the appearance probability in the learning text.

Similarly, the probability that the word w _n-1 of the next word w _n of synonyms _w'n-1 appears, can be represented by the following formula (7).

From these, the appearance probability obtained by integrating the basic type and the consent type in the history can be expressed by the following equation (8).
In Expression (8), N represents the sum of appearance probabilities in the unigram for all words in the learning text.

  Then, the history processing unit 233 updates the appearance probability of the N-gram parameter input from the probability value calculation unit 232 using the following equation (9). Thereafter, the history processing unit 233 outputs the N-gram parameter whose appearance probability is updated to the back-off coefficient processing unit 234.

  That is, according to the above equations (8) and (9), in the learning text, when one of the occurrence probabilities is zero for the basic type and the consent type, the appearance probability is zero. N-gram parameters are newly generated. In this new N-gram parameter, the appearance probability is the other appearance probability whose appearance probability is not zero.

By the way, when the degree of the word N-gram model is increased, the above-described equation (8) becomes complicated, and thus it is preferable to approximate it practically. As this approximation method, for example, the following method A or method B can be considered.
The history processing means 233 can perform history processing on chain words as well as words.

Method A: Only an N-gram parameter with an appearance probability of zero is newly generated, and calculations for other N-gram parameters are omitted.
Method B: The value of the N-gram parameter whose history is the basic type is substituted for the N-gram parameter whose history is the consent type.

  The back-off coefficient processing means 234 receives the N-gram parameter from the history processing means 233 and performs back-off coefficient processing for updating the back-off coefficient. Here, in order to simplify the description of the back-off coefficient processing, as in the history processing, the number of synonymous patterns K = 1 (the synonym w ′ exists for the basic type w), and the language model is bigram. Suppose there is.

Backoff smoothing is a method of estimating the appearance probability P (w _n | w _n−1 ) from the appearance probability P (w _n ) when the appearance probability C (w _n−1 w _n ) = 0 of the learning text. is there. Here, in the Katz method, which is one of backoff smoothing, the following equations (10) and (11) are used. At this time, it is preferable that the appearance probability of low-frequency words (including words with an appearance probability of zero) in the learning text is corrected in advance using a good Turing estimation method (for example, “probabilistic” Language model, University of Tokyo Press, pp. 67-68 ”).
Note that the back-off coefficient is α in the equations (10) and (11).

Here, when integrating the basic type w _n-1 and the consent form w _'n-1, the back-off factor alpha, Substituting in can be expressed by the following equation (12) (which in the formula (8) Can be further expanded).

  Then, the back-off coefficient processing unit 234 updates the back-off coefficient of the N-gram parameter input from the history processing unit 233 using the following equation (13). Thereafter, the back-off coefficient processing unit 234 outputs the N-gram parameter with the updated back-off coefficient to the language model update unit 235.

By the way, when the degree of the word N-gram model is increased, the above-described equation (12) becomes complicated, and therefore, it is preferable to approximate in practice. As this approximation method, for example, similarly to the history processing, calculation omission (method A) or substitution of a basic N-gram parameter (method B) can be considered.
Note that the back-off coefficient processing means 234 can perform back-off coefficient processing for chain words as well as words.

  The language model update unit 235 receives the N-gram parameter from the back-off coefficient processing unit 234 and updates the language model stored in the language model storage unit 16 using the N-gram parameter. That is, the language model update unit 235 updates the appearance probability included in the language model of the language model storage unit 16 with the appearance probability included in the N-gram parameter, and the backoff coefficient included in the language model of the language model storage unit 16 Is updated with the back-off coefficient included in the N-gram parameter.

  Here, the language model update unit 235 includes a language model deletion unit 236 as shown in FIG. The language model deletion unit 236 deletes the consent type N-gram parameter from the language model after the language model update unit 235 updates the language model. Thus, since the data size of the language model is reduced, the speech recognition apparatus 3 that refers to the language model can save the memory capacity during speech recognition.

[Pronunciation dictionary converter]
The details of the pronunciation dictionary conversion unit 24 will be described below with reference to FIGS. 8 and 9 (see FIG. 1 as appropriate).
The pronunciation dictionary conversion unit 24 converts the pronunciation dictionary format with reference to the consent word / chain word list. As shown in FIG. 8, in the pronunciation dictionary, the left column is a notation of a chain word or word, and the right column is the pronunciation of the chain word or word. In this pronunciation dictionary, pronunciation is expressed in Roman letters, and “:” indicates that the vowel immediately before that is extended and pronounced. For example, in this pronunciation dictionary, two pronunciations “toiu sp” and “toyou: sp” are registered for the chain word “to_say”.

  Here, from the synonym word / chain word list, it is possible to discriminate between the synonym word and chain word, and the basic type and the synonym type among the synonyms. For example, it is assumed that the basic type “to_say” and the consent type “te_say” are set in the consent word / chain word list. In this case, since the pronunciation dictionary of FIG. 8 has two pronunciations of the basic type “to_say”, the pronunciation dictionary conversion unit 24 adds the basic type notation “to” to the converted utterance dictionary of FIG. “_” Means that two basic pronunciations “toiu sp” and “toyou: sp” are registered. That is, as shown in FIG. 9, the pronunciation dictionary conversion unit 24 registers the basic notation “to_say” in the left column and the center column and the first pronunciation “toiu sp” in the right column. . Further, the pronunciation dictionary conversion unit 24 registers the basic type expression “to_say” in the left column and the center column and the first pronunciation “toyou: sp” in the right column (FIGS. 8 and 9). (See symbol α).

  Further, since the pronunciation dictionary of FIG. 8 has three pronunciations of the consensus type “te_say” registered, the pronunciation dictionary conversion unit 24 adds the basic expression “to” to the post-conversion utterance dictionary of FIG. “Constant”, “Consent” notation “Tell”, and consensus pronunciations “Qteiu sp”, “Qteyu: sp” and “Qtu: sp” are registered respectively. That is, as shown in FIG. 9, the pronunciation dictionary conversion unit 24 uses the basic type notation “to_say” in the left column, the consent type notation “te_say” in the center column, and the agreement in the right column. The first pronunciation of the mold “Qteiu sp” is registered. The pronunciation dictionary conversion unit 24 also displays the basic type notation “to_say” in the left column, the consent type notation “te_say” in the center column, and the second pronunciation of the consent type in the right column. “Qteyu: sp” is registered. Further, the pronunciation dictionary conversion unit 24 displays the basic type notation “to_say” in the left column, the consent type notation “te_say” in the center column, and the second pronunciation of the consent type in the right column. “Qtu: sp” is registered (see symbol β in FIGS. 8 and 9).

  That is, the pronunciation dictionary conversion unit 24 converts the pronunciation dictionary of FIG. 8 into a converted pronunciation dictionary having basic notation, consent type notation, and consent type pronunciation. Therefore, the converted pronunciation dictionary of FIG. 9 is a consensus type in which the center column is the left column when the notation of the left column and the center column is different.

[Operation of language model generator]
<Consent word / chain word selection part>
Hereinafter, the operation of the consent word / chain word selection unit 21 of FIG. 2 will be described with reference to FIG. 10 (see FIG. 2 as appropriate).
First, the language model generation device 1 extracts chain words from the learning text by the chain word extraction unit 211 (step S1). In addition, the language model generation device 1 calculates the edit distance of the extracted chain word by DP matching by the edit distance calculation means 212 (step S2).

  In addition, the language model generation device 1 selects a chain word having the minimum edit distance as a synonym candidate by the minimum edit distance selection unit 213, and generates a minimum edit distance chain word relation list (step S3). Then, the language model generation device 1 generates the consent word / chain word list based on the input selection instruction by the consent word / chain word list generation unit 214 (step S4).

<Language model conversion unit>
The operation of the language model conversion unit 23 in FIG. 6 will be described below with reference to FIG. 11 (see FIG. 6 as appropriate).
First, the language model generation device 1 generates a language model by using the language model generation unit 22 (step S11). In addition, the language model generation device 1 extracts the N-gram parameter of the synonym from the language model by the parameter extraction unit 231 (step S12).

  Moreover, the language model generation apparatus 1 calculates a probability value based on the appearance probability by the probability value calculation unit 232 (step S13). Then, the language model generation device 1 performs history processing by the history processing unit 233 (step S14).

  Further, the language model generation device 1 performs back-off coefficient processing by the back-off coefficient processing means 234 (step S15). Then, the language model generation device 1 updates the language model with the N-gram parameter for which the probability value and the back-off coefficient are calculated by the language model update unit 235 (step S16).

  As described above, the language model generation device 1 according to the embodiment of the present invention generates a word N-gram model in consideration of the context by the language model generation unit 22, and therefore can support large vocabulary continuous speech recognition. . Then, the language model generation device 1 corrects the appearance probability of the synonym calculated with a low value by the language model conversion unit 23 by dispersing the presence of the synonym in the learning text. Even in this case, it is possible to generate a language model that enables speech recognition with few recognition errors.

  Note that the language model generation device 1 updates only the N-gram parameter including the recommended notation determined in advance after updating the language model in order to solve the problem of notation fluctuation (eg, “trade” and “deal”). And the other N-gram parameters are preferably deleted.

  The language model generation device 1 has been described with an example using the identifier “basic type” indicating the basic type and the identifier “consent type” indicating the consent type, but is not limited thereto. For example, the language model generation device 1 may use “ref” as the identifier indicating the basic type and “hyp” as the identifier indicating the consent type.

  In the embodiment, the language model generation apparatus according to the present invention has been described as an independent apparatus. However, in the present invention, a general computer can be realized by a program that functions as each of the above-described units. This program may be distributed via a communication line, or may be distributed by writing in a recording medium such as a CD-ROM or a flash memory.

[Configuration of voice recognition device]
Returning to FIG. 1, the configuration of the speech recognition apparatus 3 will be described.
As shown in FIG. 1, the speech recognition device 3 includes an acoustic model storage unit 31, a speech analysis unit 33, and a search unit 35.

  The acoustic model storage unit 31 is a storage unit such as a memory or a hard disk that stores an acoustic model in advance. This acoustic model is a probability model generated in advance by learning (machine learning) a large amount of speech data.

  The voice analysis unit 33 receives the input voice (voice signal), analyzes the input voice, calculates a feature vector of the input voice, and outputs it to the search unit 35. Specifically, the speech analysis unit 33 cuts out the input speech with a Hamming window, performs linear prediction analysis (LPG) or mel cepstrum analysis, and obtains a feature vector (MFCC feature amount) of the input speech.

  The search unit 35 receives the feature vector of the input speech from the speech analysis unit 33, and outputs the result of speech recognition using the language model, the acoustic model, and the converted pronunciation dictionary from the feature vector of the input speech. To do. Specifically, the search unit 35 obtains a probability value (likelihood) by matching the feature vector of the input speech with the acoustic model, and calculates a value obtained by taking the logarithm (log) of the probability value as an acoustic score. To do. Further, the search unit 35 obtains an appearance probability (N-gram probability) from the language model for the word candidate that is a candidate for the speech recognition result during the speech recognition. At this time, the search unit 35 preferably performs a correct word search using basic parameters (appearance probability and backoff coefficient). Then, the search unit 35 takes the logarithm of the appearance probability, multiplies it by a first constant called a language weight, and sets a value obtained by adding a second constant called an insertion penalty as a language score. Thereafter, the search unit 35 refers to the post-conversion pronunciation dictionary in FIG. 9 and outputs a word candidate string having the maximum language score and acoustic score as a speech recognition result (recognition result in FIG. 1).

  As described above, the speech recognition apparatus 3 according to the embodiment of the present invention can refer to the pronunciation sequence corresponding to the word candidate (basic type) selected during the correct word search by the search unit 35. By using the pronunciation dictionary after conversion, it is possible to output the consent type notation described in the center column. That is, the voice recognition device 3 can output the consent type notation and pronunciation corresponding to the basic type, and can improve the convenience of the voice recognition system 100.

The effects of the present invention will be described below as examples.
Here, the language model is generated using the method 1 to the method 3 by the language model generation device 1 of FIG. Then, using each language model, the speech recognition apparatus 3 in FIG. 1 performed speech recognition using the news report program (large vocabulary continuous speech recognition) as input speech. In addition, as a comparison target, the same news report program was voice-recognized using a language model generated by a conventional method, and a word error rate was obtained. Table 1 below shows the results of word error rate.

  As shown in Table 1, it can be seen that the language models of Method 1 to Method 3 have a lower word error rate (recognition error) than the language models of Comparative Examples 1 and 2. That is, the language model generation device 1 can generate a language model that enables speech recognition with fewer recognition errors than in the related art.

Further, in order to use the post-conversion pronunciation dictionary of FIG. 9, the language model deleting unit 236 deletes the consensus type N-gram parameter (Example 4) and the language model generated by the conventional method. The data size was compared.
The method of Comparative Example 3 adds the non-existing N-gram parameter when the basic N-gram parameter exists in the language model and the consent-type N-gram parameter does not exist, or vice versa. Is.

  As shown in Table 2, the language model of Example 4 has a smaller data size than the language models of Comparative Examples 1 and 3. That is, the language model generation device 1 can reduce the data size of the language model as compared with the prior art.

DESCRIPTION OF SYMBOLS 1 Language model production | generation apparatus 11 Learning text memory | storage part 12 Chain word list memory | storage part 13 Edit distance addition chain word relation list memory | storage part 14 Edit distance minimum chain word relation list memory | storage part 15 Consent word and chain word list memory | storage part 16 Language model memory | storage part 17 Pronunciation dictionary storage unit 18 Converted pronunciation dictionary storage unit 21 Consent word / concatenated word selection unit (synonymous word selection unit)
211 Chain word extraction means 212 Edit distance calculation means 213 Minimum edit distance selection means (chain word candidate selection means)
214 Consent word / linked word list generation means 22 Language model generation section 23 Language model conversion section 231 Parameter extraction means 232 Probability value calculation means 233 History processing means 234 Backoff coefficient processing means 235 Language model update means 236 Language model deletion means 24 Pronunciation Dictionary conversion unit 3 Speech recognition device 31 Acoustic model storage unit 33 Speech analysis unit 35 Search unit 100 Speech recognition system

Claims (5)

A language model generation device that generates a language model using learning text including synonyms consisting of words or chain words that have the same meaning or different notation or reading,
A language model generation unit that generates a language model indicating the appearance probability of at least one of words or chain words included in the learning text by learning the learning text using a probabilistic language model;
Chain word extraction means for extracting, as the chain word, word pairs that appear more frequently than a preset frequency in the learning text in an order that most reduces entropy per word of the learning text;
Editing distance calculation means for calculating the editing distance of the chain words extracted by the chain word extraction means by DP matching;
Minimum edit distance selection means for selecting as a synonym candidate a chain word that minimizes the edit distance calculated by the edit distance calculation means;
A selection instruction including a chain word preselected from the synonym candidates is input, and based on the selection instruction, a synonym word for generating a synonym list in which the synonyms having the same meaning are associated in advance A chain word list generating means;
With reference to the synonym list, to calculate a probability value based on the probability of occurrence of synonyms having the same meaning in the language model, the probability of occurrence of the synonyms included in the language model with the probability value A language model conversion unit to be updated;
A language model generation apparatus comprising: The language model conversion unit includes:
Language model deletion means for deleting the synonym synonym type other than the basic type of the synonym having the maximum appearance probability from the language model after the language model conversion unit is updated,
The language model generation apparatus according to claim 1, further comprising: A pronunciation dictionary storage unit that stores a pronunciation dictionary that associates at least the notation of the synonym and the pronunciation of the synonym in advance;
With reference to the synonym list, the phonetic dictionary includes at least a synonym representation of the synonym, a synonym representation of the synonym corresponding to the basic type, and a synonym pronunciation of the synonym A pronunciation dictionary converter for converting to a converted pronunciation dictionary including:
Language model generating apparatus according to claim 1 or claim 2, further comprising a. Language model generation program for the computer to function as a language model generating apparatus according to claim 1. A speech recognition system comprising: the language model generation device according to claim 3; and a speech recognition device that performs speech recognition using a language model generated by the language model generation device,
The speech recognition device
An acoustic model storage unit that stores an acoustic model that is a probability model generated in advance by learning speech data;
A voice analysis unit that performs voice analysis on the input voice and calculates a feature vector of the input voice;
An acoustic score is calculated by matching the feature vector calculated by the speech analysis unit and the acoustic model, and a first constant is set for the appearance probability of a word candidate as a speech recognition result candidate with reference to the language model. A language score obtained by adding a second constant to the multiplied value is calculated, and a sequence of word candidates that maximizes the language score and the acoustic score is obtained as a result of the speech recognition with reference to the converted pronunciation dictionary. A search unit that outputs as
A speech recognition system comprising:

Images