JP2007226091A - Speech recognizer and speech recognizing program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognizer which is improved in recognition precision for a predetermined keyword. <P>SOLUTION: The speech recognizer includes a storage means 11 which stores a language model 111 wherein predetermined identification information is imparted to a word corresponding to the specified keyword, and a word string generating means 13 for outputting a word string as a recognition result by searching the language model 111 for a route giving a maximum probability value to the word string of an input voice. The word string generating means 13 includes a bonus imparting means 132a for increasing the connection probability value of the word to which the identification information is given in the language model 111. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a speech recognition apparatus and speech recognition program for recognizing input speech using a language model.

In general, speech recognition is performed by searching for a word as a recognition candidate from a language model that models the appearance frequency and connection probability of a word. As this language model, a language model (N-gram language model) using N-gram is known.
In this N-gram language model, for an input word string (word string) w ₁ w ₂ ... W _n , the word appearance probability P (w ₁ w ₂ ... W _n ) is shown below (1 ) Is a model that is calculated and generated based on a conditional probability, and is a set of entries including a word string w ₁ w ₂ ... W _n and an appearance probability P (w ₁ w ₂ ... W _n ).

That is, in the N-gram language model, the generation probability of the i-th word w _i depends on (N−1) word strings w _{i−N + 1} ... W _i−2 w _i−1 . For example, taking 3-gram (N = 3: trigram) as an example, the probability that the word w ₃ appears following the word string w ₁ w ₂ is expressed as P (w ₃ | w ₁ w ₂ ). .
In such an N-gram language model, unregistered words that have not been learned cannot be recognized.
Therefore, conventionally, a method has been proposed in which the vocabulary of the recognition dictionary is expanded to reduce unregistered words (Non-Patent Document 1, Non-Patent Document 2, etc.).
`` Open Vocabulary ASR for Audiovisual Document Indexation '', ICASSP 2005 I pp.1013-1016 `` Unsupervised Vocabulary Expansion for Automatic Transcription of Broadcast News '', ICASSP 2005 I pp.1021-1024

In general, when a speech signal of a broadcast program, a lecture, or the like is recognized by speech, there is an important word that is known in advance, that is, a keyword that needs to be recognized. For example, terms related to the theme of the lecture. Conventionally, there has been a demand for such keywords to be recognized with high accuracy.
However, in the conventional technique, a word that has not been registered in advance is registered in a recognition dictionary to enable recognition. That is, by simply registering an unregistered word in the recognition dictionary, the unregistered word can be recognized in the same manner as other words, and no contrivance has been made to increase the recognition accuracy of the word as a keyword.

In general, keywords are often proper nouns and technical terms, and are special words. Therefore, it is difficult to estimate the appearance probability based on past appearance frequency information. It was a factor that could not be increased.
The present invention has been made in view of the above problems, and an object of the present invention is to provide a speech recognition apparatus and a speech recognition program with improved recognition accuracy for a predetermined keyword.

  The present invention has been developed to achieve the above object. First, the speech recognition apparatus according to claim 1 is a speech recognition apparatus for recognizing an input speech using a language model. And a word string generating means, wherein the word string generating means has a probability value increasing means.

In such a configuration, the speech recognition apparatus stores a language model obtained by adding predetermined identification information to a word corresponding to a specific keyword in the language model storage unit. Thus, the speech recognition apparatus can identify the keyword when using the language model.
Then, the speech recognition device outputs a word string that is a recognition result by searching for a path having a maximum probability value for the word string of the input speech in the language model by the word string generation unit. At this time, since the word string generation means can identify the keyword by the identification information in the language model, the predetermined value is added to the connection probability value of the word to which the identification information is given by the probability value increasing means. To increase the probability value. As a result, the probability value that the word string including the keyword is output increases, and the recognition rate of the keyword can be increased.

  According to a second aspect of the present invention, the speech recognition apparatus according to the first aspect includes a keyword extracting unit and a language model updating unit.

In such a configuration, the speech recognition apparatus extracts a keyword word from the digitized document by the keyword extraction unit. To extract keywords from this document, for example, the word weight (importance) based on the frequency of the words that appear in the document and the scale that indicates how many documents the word appears in all the documents. The TF / IDF method for calculating the score can be used.
Then, the speech recognition apparatus updates the language model by adding identification information to the keyword extracted by the keyword extracting unit by the language model updating unit.

  Furthermore, the speech recognition apparatus according to claim 3 is a speech recognition apparatus for recognizing input speech using a language model, comprising a language model storage means and a word string generation means, wherein the word string generation means And a probability value increasing means and a probability value decreasing means.

  In such a configuration, the speech recognition apparatus provides the language model storage unit with predetermined first identification information for a word corresponding to a specific keyword and a second predetermined for a word similar to the specific keyword. The language model with identification information is stored. Thus, the speech recognition apparatus can identify a keyword and a word similar to the keyword when using the language model.

  Then, the speech recognition device outputs a word string that is a recognition result by searching for a path having a maximum probability value for the word string of the input speech in the language model by the word string generation unit. At this time, since the word string generation means can identify the keyword by the first identification information in the language model, the probability value increasing means adds the connection probability value of the word to which the first identification information is given, The probability value is increased by adding a predetermined value. Furthermore, since the word string generation means can identify a word similar to the keyword by the second identification information in the language model, the probability of connection of the word to which the second identification information is given by the probability value reduction means The probability value is decreased by subtracting a predetermined value from the value. This directly increases the probability that the word string containing the keyword will be output and also reduces the output probability of the word string containing the word similar to the keyword. The probability value is increased, and the keyword recognition rate can be increased.

  The speech recognition apparatus according to claim 4 is the speech recognition apparatus according to claim 3, wherein the pronunciation dictionary storage means, the keyword extraction means, the phoneme string search means, the similar word extraction means, and the language model update Means.

In such a configuration, the speech recognition apparatus stores a pronunciation dictionary in which pronunciations of words are stored in advance in the pronunciation dictionary storage unit. Then, the speech recognition apparatus extracts a word as a keyword from the digitized document by the keyword extraction unit.
Then, the speech recognition apparatus searches the phoneme string search means for a keyword phoneme string that is a phoneme string indicating the pronunciation of the keyword from the pronunciation dictionary.
Furthermore, the speech recognition apparatus is similar to the keyword based on the keyword phoneme string searched by the phoneme string search means by the similar word extraction means and the registered word phoneme string indicating the pronunciation of the word registered in the pronunciation dictionary. To extract words.
Then, the speech recognition device adds the first identification information to the keyword extracted by the keyword extraction unit by the language model update unit, and adds the second identification information to a word similar to the keyword, Update the model.

  Furthermore, the speech recognition device according to claim 5 is the speech recognition device according to claim 4, wherein the probability value reducing means is based on the degree of similarity between the keyword phoneme sequence and the registered word phoneme sequence. The amount of decrease in the connection probability value is changed.

  In such a configuration, the speech recognition apparatus uses the probability value reduction unit to increase the amount of decrease in the connection probability value as the degree of similarity between the keyword phoneme sequence and the registered word phoneme sequence increases. The probability of being recognized can be lowered.

  According to a sixth aspect of the present invention, there is provided a speech recognition apparatus for recognizing input speech using a language model, a keyword storage means for storing a specific keyword and its part of speech, and a word corresponding to the keyword. A language model storage means for storing a language model learned by replacing with a part of speech, and a word sequence as a recognition result is output by searching the language model for a path having a maximum probability value for the word sequence of the input speech. And a word string generation means that, based on a connection probability value of a part of speech corresponding to the keyword stored in the keyword storage means, in the language model, the word string generation means The probability value increasing means for increasing the connection probability value of the keyword is calculated.

  In such a configuration, the language model stored in the language model storage means is a language model of the part of speech class for the keyword. Therefore, the speech recognition apparatus, when the keyword connection probability value is small by the word string generation means, calculates the keyword connection probability value based on the connection probability value of the part of speech corresponding to the keyword. A connection probability value is obtained.

  Furthermore, the speech recognition program according to claim 7 gives a predetermined first identification information to a word corresponding to a specific keyword and a second identification predetermined for a word similar to the specific keyword. In order to recognize input speech using a language model to which information is added, a computer is caused to function as a feature extraction unit and a word string generation unit, and the word generation unit includes the first identification in the language model. While the connection probability value of the word to which the information is assigned is increased, the connection probability value of the word to which the second identification information is assigned is decreased.

  In such a configuration, the speech recognition program analyzes the input speech by the feature extraction unit and extracts the feature amount. Then, the speech recognition program outputs a word string that is a recognition result by searching for a path having a maximum probability value for the word string of the input speech in the language model by the feature amount by the word string generation unit. At this time, since the word string generation means can identify the keyword by the first identification information in the language model, the probability value increasing means preliminarily sets the connection probability value of the word to which the first identification information is given. The probability value is increased by adding the determined values. Furthermore, since the word string generation means can identify a word similar to the keyword by the second identification information in the language model, the probability of connection of the word to which the second identification information is given by the probability value reduction means The probability value is decreased by subtracting a predetermined value from the value. This directly increases the probability that the word string containing the keyword will be output and also reduces the output probability of the word string containing the word similar to the keyword. The probability value is increased, and the keyword recognition rate can be increased.

The present invention has the following excellent effects.
According to the first aspect of the present invention, it is possible to perform speech recognition with a higher recognition accuracy for keywords registered in advance. As a result, for example, when a unique keyword is used in a broadcast program or lecture, it is possible to increase the recognition accuracy even if the word is usually low in recognition accuracy, which is suitable for individual broadcast programs and lectures. It is possible to perform voice recognition.

According to the second aspect of the present invention, since the keyword can be extracted from the digitized document, the keyword can be registered without human intervention.
Further, according to the present invention, since identification information is only given to a keyword, it is possible to easily return to the voice recognition device with the original accuracy by deleting the information. Thus, for example, even when it is desired to improve the recognition accuracy of different keywords by using a script such as a different broadcast program, it is easy to do so simply by registering the keyword with a new script after deleting the previous identification information. Can do.

According to the invention described in claim 3 or claim 7, since the connection probability value of the word similar to the keyword can be reduced, the probability that the keyword is recognized becomes relatively high. This makes it possible to perform speech recognition with a higher recognition accuracy for keywords registered in advance.
According to the fourth aspect of the present invention, since a word similar to the keyword can be extracted from the digitized document, information for identifying the word similar to the keyword is registered without human intervention. Can do.
Further, according to the present invention, since identification information is only given to a keyword, it is possible to easily return to the voice recognition device with the original accuracy by deleting the information.

According to the fifth aspect of the present invention, as the word resembles the keyword, the connection probability value is further reduced, so that the keyword recognition accuracy can be relatively improved.
According to the sixth aspect of the present invention, since the connection probability value is calculated using the language model of the part-of-speech class for the keyword, the connection probability value can be increased as compared with the case where the connection probability value is obtained only from a normal word. It is possible to improve the keyword recognition accuracy.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
[First embodiment]
(Configuration of voice recognition device)
First, the configuration of the speech recognition apparatus according to the first embodiment will be described with reference to FIG. FIG. 1 is a block diagram showing the configuration of the speech recognition apparatus according to the first embodiment of the present invention. The speech recognition apparatus 1 shown in FIG. 1 recognizes an input speech signal with improved recognition accuracy for a specific keyword. Here, the speech recognition apparatus 1 includes a recognition unit 10 and an update unit 20.

  The recognition unit 10 recognizes an audio signal based on a language model, an acoustic model, and a pronunciation dictionary. Here, the recognition unit 10 includes a storage unit 11, a feature extraction unit 12, and a word string generation unit 13.

The storage means (language model storage means, pronunciation dictionary storage means) 11 stores a language model 111, an acoustic model 112, and a pronunciation dictionary 113, and is a general storage device such as a hard disk.
The language model 111 models the appearance frequency, connection probability, and the like of an output sequence (words, morphemes, phonemes, etc.) learned from a large amount of speech data. As this language model, for example, a general “N-gram language model” can be used.

  The language model 111 is generated by the update unit 20 described later with information for identifying a specific keyword. Here, with reference to FIG. 2, the content of the language model 111 is demonstrated concretely. FIG. 2 is a data structure diagram showing the contents of the language model according to the present invention.

As shown in FIG. 2, the language model 111 sets the probability of connection between a word constituting a sentence and the next word, and the probability of a word that tends to appear at the beginning of a sentence from “beginning” to “end of sentence”. Described data. The language model 111 is expanded based on a general N-gram language model so that a specific keyword can be identified.
Specifically, in the language model 111, identification information (characters “!” In FIG. 2) for specifying the keyword is given to the keyword (“Kosugi” in FIG. 2). An area BA is provided for storing the probability value to be added as a bonus value for the probability value to be connected.
Thereby, when calculating the connection probability value of the word, the accuracy of recognizing the keyword can be increased by adding the bonus value.

  Further, in the language model 111, an area PA for storing a probability value to be subtracted from a probability value to be connected as a penalty value is provided for words similar in pronunciation to the keyword in order to improve the recognition accuracy of the keyword. Yes. In the example of FIG. 2, the area for storing the bonus value and the penalty value is the same, and either one is identified by the symbols “+” and “−”.

Here, the identification information (first identification information) for identifying the keyword by characters such as “!” Is used, but a sign such as “+” of the bonus value may be regarded as the identification information. Here, a sign such as “−” of the penalty value is used as second identification information indicating a word similar to the keyword.
The generation of the language model shown in FIG. 2 is performed in the description of the update unit 20 described later.
Returning to FIG. 1, the description of the configuration of the speech recognition apparatus 1 will be continued.

The acoustic model 112 is obtained by modeling a feature amount for each phoneme learned in advance from a large amount of speech data using a “hidden Markov model”. As the acoustic model 112, a single acoustic model may be used, or a plurality of models may be used for each acoustic type (for example, for each person).
The pronunciation dictionary 113 shows the structure of consonants and vowels that indicate the pronunciation for each word. In this pronunciation dictionary 113, pronunciations of a plurality of words are registered in advance.

The feature extraction unit 12 analyzes a voice (voice signal) input from the outside and extracts a feature amount of the voice.
Note that the feature extraction unit 12 extracts a framed waveform by applying a window function (such as a Hamming window) to the speech waveform of the input speech, and performs frequency analysis on the waveform to obtain various feature quantities. To extract. For example, a cepstrum coefficient or the like, which is a value obtained by inverse Fourier transform of the logarithm of the power spectrum of a framed waveform, is used as the feature amount. In addition to the cepstrum coefficient, a general speech feature quantity such as a mel frequency cepstrum coefficient (MFCC), an LPC (Linear Predictive Cording) coefficient, logarithmic power, or the like can be used as the feature quantity.

  Based on the language model 111, the acoustic model 112, and the pronunciation dictionary 113 stored in the storage unit 11, the word string generation unit 13 generates a word string that is a speech recognition result from the feature amount extracted by the feature extraction unit 12. Is to be generated. Here, the word string generation unit 13 includes an acoustic similarity calculation unit 131 and a search unit 132.

  The acoustic similarity calculation unit 131 extracts the similarity (probability) between the feature quantity extracted by the feature extraction unit 12 and input in time series and the phoneme modeled by the acoustic model 112 stored in the storage unit 11. Value).

  The search means 132 searches the language model 111 for connected output sequence candidates, and outputs the output sequence having the maximum probability value as a recognition result (recognition word string) for the input speech. Here, the search unit 132 includes a bonus grant unit 132a and a penalty grant unit 132b.

  The bonus giving means (probability value increasing means) 132a adds a predetermined probability value as a bonus value to the probability of a word registered as a keyword in the language model 111 when calculating the probability value of the output series. Is. Here, the bonus grant means 132a recognizes a word as a keyword from the identification information (characters “!” In FIG. 2) in the language model 111 described with reference to FIG. 2, and a bonus value assigned to the keyword. Is added to the original probability value. As a result, the probability of keyword connection increases, and the accuracy with which the keyword is recognized can be increased.

  When calculating the connection probability of the output sequence, the penalty granting means (probability value reducing means) 132b penalizes a predetermined probability value from the probabilities of words registered as words similar to the keyword in the language model 111. Subtract as a value. Thereby, the connection probability of words similar to the keyword is lowered, and the accuracy of recognizing the keyword can be relatively increased.

  Note that the search unit 132 adds a connection probability for each word based on the language model 111 and also adds a similarity (probability value) based on the acoustic model 112 calculated by the acoustic similarity calculation unit 131.

  The update unit 20 updates the language model 111 used by the recognition unit 10 so that the recognition accuracy of a specific keyword is increased. Here, the update unit 20 includes a keyword extraction unit 21, a storage unit 22, a phoneme string search unit 23, a similar word extraction unit 24, and a language model update unit 25.

  The keyword extraction means 21 extracts a keyword from a document containing the digitized keyword. This document corresponds to, for example, a broadcast program script or a configuration table. Here, the keyword extraction means 21 weights the words in the document, and specifies the word that becomes the keyword based on the weight (importance score).

Here, referring to FIG. 3 (refer to FIG. 1 as appropriate), a keyword extraction method performed by the keyword extraction means 21 will be described. FIG. 3 is an explanatory diagram for explaining a keyword extraction method.
Here, as shown in FIG. 3, N documents d including a keyword are input, and the keywords are extracted from the document. Note that the symbol “Δ” in the document d indicates a word break. As described above, the input document d may be data divided in advance for each word, or may be classified for each word by the keyword extraction unit 21 performing morphological analysis.

In general, many important words as keywords appear in the document d. However, if many appear in all documents (N), it cannot be said that the word characterizes the document d. Therefore, here, the keyword extracting means 21 uses the “TF / IDF method” as a method for extracting a keyword from the document d, and calculates the weight of the word (importance score) to extract the keyword.
The importance score in the “TF / IDF method” is a scale (IDF :) indicating the frequency (TF: Term Frequency) of a word that appears in a document and how many documents in the entire document the word appears. Inverse Document Frequency).
That is, the keyword extracting means 21 uses tf (t, d) as the frequency of the word t appearing in a certain document d, and idf (t) as a scale indicating how many words the word t appears in all documents. Then, the importance score w _t ^d of the word t is calculated by the following equation (2).

  The scale idf (t) in the equation (2) can be expressed by the following equation (3), where N is the total number of documents and df (t) is the number of documents in which the word t appears.

As described above, the keyword extracting unit 21 calculates the importance score of the word in the document from the document, and extracts the word having the importance score larger than the predetermined threshold as the keyword. Then, the keyword extraction unit 21 stores the extracted keyword and its importance score in the storage unit 22. Here, the importance score is used as the bonus value described above.
In the example shown in FIG. 3, the keywords “Cosgi”, “Karate”, and “Martial art” are extracted from the document d, and “1.000”, “0. 959 ”and“ 0.532 ”are calculated.
Returning to FIG. 1, the description of the configuration of the speech recognition apparatus 1 will be continued.

  The storage means (keyword storage means) 22 stores the keyword extracted by the keyword extraction means 21 and the importance score as a keyword dictionary 221 and is a general storage device such as a hard disk.

  The phoneme string search means 23 searches the pronunciation dictionary 113 for a phoneme string (keyword phoneme string) indicating the pronunciation of a keyword included in the keyword dictionary 221 stored in the storage means 22. The keyword phoneme string retrieved from the pronunciation dictionary 113 is a character string in which consonants and vowels are combined. The phoneme string search means 23 outputs the keyword and the searched keyword phoneme string to the similar word extraction means 24.

  The similar word extracting unit 24 selects a keyword based on the keyword phoneme sequence indicating the pronunciation of the keyword searched by the phoneme sequence searching unit 23 and the registered word phoneme sequence indicating the pronunciation of the word registered in the pronunciation dictionary 113. Words that are similar (including the same) (similar words) are extracted. Here, the similar word extracting unit 24 includes a similarity degree measuring unit 241.

The similarity measurer 241 measures the similarity indicating the degree of similarity between the keyword phoneme string and the registered word phoneme string.
That is, the similar word extracting unit 24 extracts words having a registered word phoneme sequence similar to the keyword phoneme sequence as similar words based on the similarity measured by the similarity measuring unit 241.

Here, with reference to FIG. 4 (refer to FIG. 1 as appropriate), a method of extracting similar words performed by the similar word extracting unit 24 will be described. FIG. 4 is an explanatory diagram for explaining a method of extracting similar words.
Here, as shown in FIG. 4, it is assumed that “martial arts” and “Kosugi” are extracted in advance as keywords.
At this time, the similar word extracting means 24 measures the similarity indicating the degree of similarity between the keyword phoneme string and the registered word phoneme string by the similarity measuring means 241. The similarity between the phoneme strings can be obtained as a distance between the phoneme strings by, for example, DP (Dynamic Programming) matching method.

In the example of FIG. 4, the keyword “martial arts” and the registered word “grape” are both phoneme strings “budo:”, and the distance between phoneme strings is “0”, resulting in the same pronunciation. In addition, the keyword “Kosugi” and the registered word “Kosuge” are different in each phoneme sequence between “kosugi” and “kosugue”, but only one phoneme (“i” and “e”). Is a similar pronunciation of “1”.
Therefore, the similar word extraction unit 24 sets the penalty value larger as the distance between the phoneme strings is shorter. Here, an example is shown in which a penalty value “−1.0” is set for the registered word “grape”, and a penalty value “−0.5” is set for the registered word “Kosuge”.
Returning to FIG. 1, the description of the configuration of the speech recognition apparatus 1 will be continued.

  The language model update unit 25 uses the keywords and importance scores (bonus values) registered in the keyword dictionary 221 stored in the storage unit 22 and the similar words and penalty values extracted by the similar word extraction unit 24. Based on this, the language model 111 is updated. That is, as described in FIG. 2, the language model update unit 25 adds identification information (character “!” In FIG. 2) for specifying the keyword to the word indicating the keyword in the language model 111. A bonus value to be added to the connection probability value is registered in the area BA. Furthermore, the language model update unit 25 searches for similar words in the language model 111 and registers a penalty value to be subtracted from the connection probability value in the area PA.

The language model update unit 25 cancels the registration of the keyword and deletes the bonus value and the penalty value when the keyword registered previously exists in the language model 111. Here, in order to delete the registration of the keyword, the identification information attached to the keyword (the character “!” In FIG. 2) may be deleted. As a result, the language model 111 is in a state in which a bonus value is given only for a keyword for which recognition accuracy is always to be improved.
By configuring the speech recognition device 1 in this way, the speech recognition device 1 can extract a keyword from an electronic document and perform speech recognition with an increased recognition accuracy for the keyword.

The configuration of the speech recognition apparatus 1 according to the first embodiment has been described above, but the present invention is not limited to this configuration. For example, if a keyword is determined in advance, the keyword extraction unit 21 may be omitted from the configuration, and the keyword and bonus value may be registered directly in the keyword dictionary 221.
Here, in the language model 111, a bonus value is set for a keyword and a penalty value is set for a similar word. However, a configuration in which a bonus value is only set for a keyword may be used. That is, the speech recognition apparatus 1 may be configured by omitting the penalty giving means 132b, the phoneme string searching means 23, and the similar word extracting means 24. In this case, the language model update unit 25 updates the language model 111 based on the keywords and bonus values registered in the keyword dictionary 221 stored in the storage unit 22.
The voice recognition apparatus 1 can be operated by a voice recognition program that causes a general computer to function as each of the above-described means.

(Operation of voice recognition device)
Next, the operation of the speech recognition apparatus will be described with reference to FIGS. 5 and 6 (see FIG. 1 for the configuration). FIG. 5 is a flowchart showing the language model update operation of the speech recognition apparatus according to the first embodiment of the present invention. FIG. 6 is a flowchart showing the speech recognition operation of the speech recognition apparatus according to the first embodiment of the present invention.

[Update operation of language model]
As shown in FIG. 5, first, the speech recognition apparatus 1 uses a keyword extraction unit 21 to extract word weights (importance score; bonus value) from a document containing keywords digitized by the “TF / IDF method”. The keyword is extracted by calculating (use) (step S1).
Then, the keyword extraction unit 21 stores the keyword extracted in step S1 and the bonus value of the keyword in the storage unit 22 as the keyword dictionary 221 (step S2).
Thereafter, the speech recognition apparatus 1 searches the phoneme sequence 113 for the phoneme sequence (keyword phoneme sequence) indicating the pronunciation of the keyword included in the keyword dictionary 221 by the phoneme sequence search means 23 (step S3).

Then, the speech recognition apparatus 1 uses the similar word extraction unit 24 as a keyword based on the keyword phoneme string searched in step S3 and the registered word phoneme string indicating the pronunciation of the word registered in the pronunciation dictionary 113. A similar (including the same) word (similar word) is extracted, and a penalty value for the similar word is determined (step S4).
At this time, the similar word extracting unit 24 extracts the similar word by measuring the similarity between the keyword phoneme sequence and the registered word phoneme sequence by the similarity measuring unit 241 based on the distance between the phoneme sequences by the DP matching method. To do. Furthermore, the similar word extracting unit 24 increases the penalty value for similar words having a shorter distance between phoneme strings.

Then, the speech recognition apparatus 1 uses the language model update unit 25 based on the keywords and bonus values registered in the keyword dictionary 221 stored in step S2 and the similar words and penalty values extracted in step S4. The language model 111 is updated.
That is, the speech recognition apparatus 1 uses the language model update unit 25 to add identification information (characters “!” In FIG. 2) for specifying a keyword in the language model 111 to the word indicating the keyword, and the connection probability. A bonus value to be added to the value is registered (step S5).
Further, the speech recognition apparatus 1 searches for similar words in the language model 111 by the language model update unit 25 and registers a penalty value to be subtracted from the connection probability value (step S6).

Through the above operation, the speech recognition apparatus 1 can extract a keyword from an electronic document and newly generate a language model that can recognize the keyword.
When the keyword is updated with another new document, the registered identification information, bonus value, and penalty value are deleted before step S5. As a result, for example, even if a keyword whose recognition accuracy is to be improved differs depending on a script of a broadcast program, a configuration table, or the like, the target keyword can be easily changed.

[Voice recognition operation]
Next, as illustrated in FIG. 6, the speech recognition apparatus 1 uses the search unit 132 to search for a connected output series candidate from the language model 111.
At this time, the search means 132 determines whether the candidate word is a word registered as a keyword (step S11). If it is a keyword (Yes in step S11), the bonus grant means 132a. Thus, the value obtained by adding the bonus value to the connection probability value is added to the probability value of the output series (step S12), and the process proceeds to step S16.

Further, the search unit 132 determines whether the candidate word is a similar word similar to the keyword (step S13). If the word is a similar word (Yes in step S13), the penalty granting unit 132b Then, a value obtained by subtracting the penalty value from the connection probability value is added to the probability value of the output series (step S14), and the process proceeds to step S16.
On the other hand, if the candidate word is neither a keyword nor a similar word, the connection probability value set for the word is added to the probability value of the output series (step S15).

  Further, the speech recognition apparatus 1 extracts the feature amount of the speech by analyzing the input speech (speech signal) by the feature extraction unit 12, and the pronunciation ( The similarity with the phoneme is added to the output series (step S16).

Then, the search means 132 determines whether or not the connected word further continues (step S17), and when it continues (Yes in step S17), returns to step S11 and adds the probability value of the output sequence. I will do it.
Then, the search means 132 outputs the output series having the maximum probability value as a recognized word string at the stage of calculating the probability values of all the output series candidates (step S18).
With the above operation, the speech recognition apparatus 1 can perform speech recognition with improved keyword recognition accuracy.

[Second Embodiment]
(Configuration of voice recognition device)
Next, the configuration of the speech recognition apparatus according to the second embodiment will be described with reference to FIG. FIG. 7 is a block diagram showing the configuration of the speech recognition apparatus according to the second embodiment of the present invention. The speech recognition apparatus 1B shown in FIG. 7 recognizes an input speech signal with improved recognition accuracy for a specific keyword. Here, the speech recognition apparatus 1B includes a recognition unit 10B and an update unit 20B.

The speech recognition apparatus 1B is different from the speech recognition apparatus 1 described in FIG. 1 with respect to the contents of the language model 111B and the keyword dictionary 221B and the functions of the bonus granting means 132Ba, the keyword extracting means 21B, and the language model updating means 25B. Is different. Other configurations are the same as those of the speech recognition apparatus 1 described with reference to FIG.
Note that the speech recognition device 1B is different from the speech recognition device 1 described in FIG. 1 in a method of increasing the connection probability for a keyword (giving a bonus).

The keyword extraction means 21B extracts keywords from a document containing digitized keywords. The keyword extracting unit 21B extracts keywords in the same manner as the keyword extracting unit 21 described with reference to FIG. 1, except that a function for recognizing the keyword part of speech is added. That is, here, the keyword extraction unit 21B analyzes the keyword part of speech by performing morphological analysis.
Then, the keyword extraction unit 21B registers the extracted keyword and its part of speech in the keyword dictionary 221B of the storage unit 22B.

  The language model update unit 25B updates the language model 111B based on keywords and parts of speech registered in the keyword dictionary 221B stored in the storage unit 22B. Here, the language model update unit 25B replaces a word corresponding to a keyword among words registered as learning text with a unique character string indicating the part of speech of the keyword, and then generates a language model of the part of speech class. To do.

Here, with reference to FIG. 8 (refer to FIG. 7 as appropriate), a method in which the language model update unit 25B generates the language model 111B will be described. FIG. 8 is an explanatory diagram for explaining a procedure for generating a language model according to the present invention.
Here, first, as shown in FIG. 8A, it is assumed that there is a learning text for generating a language model. It is assumed that this learning text is normally stored in a storage means (not shown) using about several million sentences.

Further, as shown in FIG. 8B, it is assumed that tens of thousands of words (registered vocabulary) are registered in the language model 111B as the learning result of FIG. 8A.
At this time, if it is assumed that an unregistered vocabulary as shown in FIG. 8 (c) (here, “junior high court”, “child”) is a keyword, the language model updating unit 25B is as shown in FIG. 8 (d). In addition, the word corresponding to the unregistered vocabulary (keyword) of the learning text is replaced with the keyword part of speech with a unique mark (here, “$”).
FIG. 8D shows an example in which “junior court” is replaced with “$ proper noun $” and “child” is replaced with “$ general noun $”.

Thereafter, the language model update unit 25B counts the number of N-grams as shown in FIG. FIG. 8E shows an example of N = 2 among N-grams.
Then, the language model update unit 25B calculates the connection probability value of each word and relearns the language model 111B as shown in FIG. As a result, a language model of the part of speech class is generated for the keyword.
Returning to FIG. 7, the description of the configuration of the speech recognition apparatus 1B will be continued.

The bonus giving means 132Ba calculates the probability value of the output series based on the part of speech of the keyword in the language model 111B. The keyword part of speech is registered in the keyword dictionary 221B.
Here, when the probability value (language score) of the output series is calculated, the bonus grant unit 132Ba uses a language score based on the part of speech as the keyword language score, and inputs a word string from which a larger language score can be obtained. Output as a speech recognition result (a recognition word string).
Specifically, the bonus granting unit 132Ba has a keyword w, a history of (N-1) -gram immediately before the keyword w is h, and a part of speech class is C. A posteriori probability P (w | h) of appearance is estimated.

That is, the bonus granting means 132Ba uses the posterior probability P (w | h) that the keyword w appears after the word string ((N-1) -gram history) h, and the part of speech class C appears after the word string h. And the probability that the keyword w appears in the part-of-speech class C is calculated.
Here, the probability that the keyword w appears in the part-of-speech class C is the reciprocal of the number of keywords belonging to the part-of-speech class C (for example, the proper noun class).
In this case, in the design of the vocabulary used in the speech recognition apparatus 1B in advance, for example, if proper nouns that are generally regarded as special words are excluded, the probability value given to the proper noun word increases. Further, since keywords are generally limited to parts of speech such as proper nouns, a bonus is substantially given to the language score value of the output series including the keywords.

The configuration of the speech recognition apparatus 1B according to the second embodiment has been described above, but the present invention is not limited to this configuration.
For example, if a keyword is determined in advance, the keyword extraction unit 21B may be omitted from the configuration, and the keyword and part of speech may be registered directly in the keyword dictionary 221B. Further, as with the speech recognition apparatus 1 according to the first embodiment, the penalty giving means 132b, the phoneme string searching means 23, and the similar word extracting means 24 may be omitted.
The voice recognition device 1B can be operated by a voice recognition program that causes a general computer to function as each of the above-described means.

(Operation of voice recognition device)
Next, the operation of the speech recognition apparatus will be described with reference to FIGS. 9 and 10 (see FIG. 1 for the configuration). FIG. 9 is a flowchart showing the language model update operation of the speech recognition apparatus according to the second embodiment of the present invention. FIG. 10 is a flowchart showing the speech recognition operation of the speech recognition apparatus according to the second embodiment of the present invention.

[Update operation of language model]
As shown in FIG. 9, first, the speech recognition apparatus 1B extracts a keyword and its part of speech from a document including the digitized keyword by the keyword extraction unit 21B (step S21).
Then, the keyword extraction unit 21B stores the keyword extracted in step S21 and its part of speech in the storage unit 22 as a keyword dictionary 221B (step S22).
Thereafter, the speech recognition device 1B searches the phoneme string search means 23 for a phoneme string (keyword phoneme string) indicating the pronunciation of the keyword included in the keyword dictionary 221B from the pronunciation dictionary 113 (step S23).

  The speech recognition apparatus 1B uses the similar word extraction unit 24 as a keyword based on the keyword phoneme string searched in step S23 and the registered word phoneme string indicating the pronunciation of the word registered in the pronunciation dictionary 113. A similar (including the same) word (similar word) is extracted, and a penalty value is set for the similar word (step S24).

Then, the speech recognition device 1B uses the language model update unit 25B based on the keywords and parts of speech registered in the keyword dictionary 221B stored in step S22 and the similar words and penalty values extracted in step S24. The language model 111B is updated. That is, the speech recognition apparatus 1B replaces the word corresponding to the keyword among the words registered as the learning text by the language model updating unit 25B with the unique character string indicating the part of speech of the keyword, and then the learning text. A language model of the part of speech class is generated by learning (Step S25).
Furthermore, the speech recognition apparatus 1B searches for a similar word in the language model 111B by the language model update unit 25B and registers a penalty value for subtracting the connection probability value (step S26).
Through the above operation, the speech recognition apparatus 1B extracts keywords from the digitized document, and generates a language model modeled with parts of speech for the keywords.

[Voice recognition operation]
Next, as shown in FIG. 10, the speech recognition apparatus 1B searches the output model candidate from the language model 111B by the search unit 132B.
At this time, the search means 132B determines whether or not the candidate word is a word registered as a keyword (step S31). If it is a keyword (Yes in step S31), the bonus grant means 132Ba Thus, the probability value of the part of speech corresponding to the keyword is added to the probability value of the output series (step S32), and the process proceeds to step S36.

Further, the search means 132B determines whether or not the candidate word is a similar word similar to the keyword (step S33), and if it is a similar word (Yes in step S33), the penalty is determined from the connection probability value. The value obtained by subtracting the value is added to the probability value of the output series (step S34), and the process proceeds to step S36.
On the other hand, if the candidate word is neither a keyword nor a similar word, the connection probability value set for the word is added to the probability value of the output sequence (step S35).

  Furthermore, the speech recognition apparatus 1B extracts the feature amount of the speech by analyzing the input speech (speech signal) by the feature extraction unit 12, and the pronunciation ( The similarity with the phoneme is added to the output series (step S36).

Then, the search means 132B determines whether or not the connected word continues (step S37), and if so (Yes in step S37), returns to step S31 and adds the probability value of the output sequence. I will do it.
Then, the search means 132B outputs the output series having the maximum probability value as a recognized word string at the stage of calculating the probability values of all the output series candidates (step S38).
With the above operation, the speech recognition apparatus 1B can perform speech recognition with improved keyword recognition accuracy.

It is a block diagram which shows the structure of the speech recognition apparatus which concerns on 1st embodiment of this invention. It is a data structure figure which shows the content of the language model which concerns on this invention. It is explanatory drawing for demonstrating the extraction method of a keyword. It is explanatory drawing for demonstrating the extraction method of a similar word. It is a flowchart which shows the update operation of the language model of the speech recognition apparatus which concerns on 1st embodiment of this invention. It is a flowchart which shows the speech recognition operation | movement of the speech recognition apparatus which concerns on 1st embodiment of this invention. It is a block diagram which shows the structure of the speech recognition apparatus which concerns on 2nd embodiment of this invention. It is explanatory drawing for demonstrating the procedure which produces | generates the language model which concerns on this invention. It is a flowchart which shows the update operation of the language model of the speech recognition apparatus which concerns on 2nd embodiment of this invention. It is a flowchart which shows the speech recognition operation | movement of the speech recognition apparatus which concerns on 2nd embodiment of this invention.

Explanation of symbols

1 speech recognition device 11 storage means (language model storage means, pronunciation dictionary storage means)
111 Language Model 112 Acoustic Model 113 Pronunciation Dictionary 12 Feature Extraction Unit 13 Word String Generation Unit 131 Acoustic Similarity Calculation Unit 132 Search Unit 132a Bonus Grant Unit (Probability Value Increase Unit)
132b Penalty giving means (probability reduction means)
21 keyword extraction means 22 storage means (keyword storage means)
221 Keyword dictionary 23 Phoneme string search means 24 Similar word extraction means 241 Similarity measurement means 25 Language model update means

Claims (7)

In a speech recognition device that recognizes input speech using a language model,
Language model storage means for storing a language model in which predetermined identification information is given to a word corresponding to a specific keyword;
In the language model, a word string generation unit that outputs a word string that is a recognition result by searching for a path that maximizes a probability value for the word string of the input speech, and
The word string generation means includes a probability value increasing means for increasing the probability value by adding a predetermined value to a connection probability value of a word to which the identification information is given in the language model. A voice recognition device characterized by that. Keyword extracting means for extracting the word as the keyword from the digitized document;
A language model updating unit that updates the language model by adding the identification information to the keyword extracted by the keyword extracting unit;
The speech recognition apparatus according to claim 1, further comprising: In a speech recognition device that recognizes input speech using a language model,
A language model storage means for storing a language model in which predetermined first identification information is given to a word corresponding to a specific keyword and predetermined second identification information is given to a word similar to the specific keyword When,
In the language model, a word string generation unit that outputs a word string that is a recognition result by searching for a path that maximizes a probability value for the word string of the input speech, and
The word string generation means is
In the language model, a probability value increasing means for increasing the probability value by adding a predetermined value to the connection probability value of the word to which the first identification information is given,
In the language model, a probability value reducing means for reducing the probability value by subtracting a predetermined value from a connection probability value of the word to which the second identification information is given,
A speech recognition apparatus characterized by comprising: Pronunciation dictionary storage means for storing a pronunciation dictionary storing the pronunciation of a word;
Keyword extracting means for extracting the word as the keyword from the digitized document;
Phoneme string search means for searching a keyword phoneme string, which is a phoneme string indicating the pronunciation of the keyword, from the pronunciation dictionary;
Similar word extraction means for extracting a word similar to the keyword based on the keyword phoneme string searched by the phoneme string search means and a registered word phoneme string indicating pronunciation of a word registered in the pronunciation dictionary; ,
A language model update that updates the language model by adding the first identification information to the keyword extracted by the keyword extraction means, and adding the second identification information to a word similar to the keyword. Means,
The speech recognition apparatus according to claim 3, further comprising:   5. The speech recognition apparatus according to claim 4, wherein the probability value decreasing unit changes the amount of decrease in the connection probability value based on a degree of similarity between the keyword phoneme string and the registered word phoneme string. . In a speech recognition device that recognizes input speech using a language model,
Keyword storage means for storing a specific keyword and its part of speech;
Language model storage means for storing a language model learned by replacing a word corresponding to the keyword with a part of speech;
In the language model, a word string generation unit that outputs a word string that is a recognition result by searching for a path that maximizes a probability value for the word string of the input speech, and
The word string generation means is
In the language model, the probability value for increasing the connection probability value of the keyword by calculating the connection probability value of the keyword based on the connection probability value of the part of speech corresponding to the keyword stored in the keyword storage means A speech recognition apparatus comprising an increasing means. Using a language model in which predetermined first identification information is given to a word corresponding to a specific keyword and predetermined second identification information is given to a word similar to the specific keyword, To recognize the computer,
A feature extraction means for analyzing the input speech and extracting a feature amount;
A word string that outputs a word string that is a recognition result by searching for a path in which the probability value for the word string of the input speech is maximized in the language model based on the feature amount extracted by the feature extraction unit. Function as a generation means,
The word string generation means increases the probability value by adding a predetermined value to a connection probability value for the word to which the first identification information is given in the language model, and the second A speech recognition program for reducing a probability value by subtracting a predetermined value from a connection probability value for a word assigned with the identification information.

Images