JP2010197411A - Language model update device for voice recognition device, and voice recognition device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a language model update device for voice recognition devices, which can improve a recognition rate of a specified word in which improvement of the recognition rate is desired, and can suppress the degradation of the recognition rate of a similar word. <P>SOLUTION: The language model update device for voice recognition devices 100 comprises: a keyword storage means 101 for storing a keyword; an influence degree calculation means 105 for calculating an influence degree to a similar word by improving a recognition rate of the keyword; a language score adjusting means 104 for adjusting at least either a bonus given to the language score of the keyword, or a penalty given to the language score of the similar word based on the influence degree; and a language score update means 106 for giving at least either the bonus which is adjusted by the language score adjusting means 104 to the language score of the keyword, or the penalty which is adjusted by the language score adjusting means 104 to the language score of the similar word. Thus, the language score of a language model 107 is updated. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a language model update device for a speech recognition device, a speech recognition device, a language model update method for a speech recognition device, a speech recognition method, a computer program, and a recording medium.

  A speech recognition device for recognizing a speaker's speech has been developed and partly put into practical use. For example, Patent Documents 1 and 2 are prior art documents related to such a speech recognition apparatus.

  The speech recognition device described in Patent Document 1 is a speech recognition device that recognizes input speech using a language model. In this speech recognition apparatus, the language model is searched for a path having the maximum probability value for the word sequence of the input speech, and a word sequence as a recognition result is output. The speech recognition apparatus described in Patent Literature 1 has a word registration function that allows a user to register unregistered words that are not registered in the recognition dictionary.

  In the speech recognition apparatus described in Patent Literature 2, the word score is added to the acoustic score indicating how close the speech signal and the word dictionary are acoustically at a specified rate, the distance value is calculated, and the distance value is calculated. The result with the smaller value is output.

  Among these conventional speech recognition devices, for example, a speech recognition device having a word registration function for registering unregistered words that are not registered in the recognition dictionary, such as the speech recognition device described in Patent Document 1, There are the following problems. That is, the word registration function in such a conventional speech recognition apparatus is a function that simply gives notation and reading, and the language score after registration cannot be set appropriately. For this reason, another word other than the registered word added by the user may be mistakenly recognized as a registered word. The cause of such a problem is that a conventional speech recognition apparatus can improve the recognition rate of a specific word, but cannot prevent an excessive decrease in the recognition rate of similar words.

JP 2007-226091 A Japanese Patent Laid-Open No. 10-097285

  Therefore, the present invention provides a speech recognition apparatus having a word registration function that improves both the recognition rate of a specific word (also referred to as “keyword” in the present invention) for which the recognition rate is to be increased and suppresses the decrease in the recognition rate of a similar word. An object of the present invention is to provide a speech model recognition device language model update device, a speech recognition device, and the like.

The language model update device for a speech recognition apparatus of the present invention includes a keyword storage unit, an influence degree calculation unit, a language score adjustment unit, and a language score update unit,
The keyword storage means stores a keyword,
The influence calculating means calculates an influence on a similar word by increasing a recognition rate of the keyword;
The language score adjusting means adjusts at least one of a bonus given to the language score of the keyword and a penalty given to the language score of the similar word based on the degree of influence,
The language score update unit performs at least one of the provision of the bonus adjusted by the language score adjustment unit with respect to the language score of the keyword and the provision of a penalty adjusted by the language score adjustment unit with respect to the language score of the similar word. Thus, the language score of the language model is updated.

The speech recognition device of the present invention includes the language model update device for the speech recognition device of the present invention,
Voice input means for inputting voice;
Feature quantity extraction means for extracting the feature quantity of the input speech;
Language model,
Voice recognition means,
Updating the language model in the language model updating device;
The speech recognition means recognizes the speech using the updated language model from the feature quantity extracted by the feature quantity extraction means.

The method for updating a language model for a speech recognition apparatus according to the present invention uses a keyword storage unit, an influence calculation unit, a language score adjustment unit, and a language score update unit.
A keyword storage step in which the keyword storage means stores a keyword;
An influence degree calculating step in which the influence degree calculating means calculates an influence degree on a similar word by increasing a recognition rate of the keyword;
A language score adjusting step in which the language score adjusting means adjusts at least one of a bonus given to the language score of the keyword and a penalty given to the language score of the similar word based on the degree of influence;
The language score update means performs at least one of the provision of the bonus adjusted by the language score adjustment means for the language score of the keyword and the provision of the penalty adjusted by the language score adjustment means for the language score of the similar word. Thus, a language score update step of updating the language score of the language model is included.

The speech recognition method of the present invention uses the speech recognition device of the present invention,
Updating the language model by the language model updating method for a speech recognition apparatus of the present invention;
A feature amount extraction step in which the feature amount extraction means extracts the feature amount of the voice input by the voice input means;
The speech recognition unit includes a speech recognition step of recognizing the speech using the language model from the feature amount extracted by the feature amount extraction unit.

  The computer program of the present invention is a computer program capable of executing the language model updating method for a speech recognition apparatus of the present invention or the speech recognition method of the present invention on a computer.

  The storage medium of the present invention is a recording medium storing the computer program of the present invention.

  According to the speech model recognition device language model update device and the speech recognition device of the present invention, the speech recognition device achieves both the improvement of the recognition rate of a specific word for which the recognition rate is to be increased and the suppression of the decrease in the recognition rate of a similar word. it can.

FIG. 1 is a block diagram showing a configuration of Embodiment 1 of the present invention. FIG. 2 is a language model update processing flowchart according to the first embodiment of the present invention. FIG. 3 is a block diagram showing the configuration of the second embodiment of the present invention. FIG. 4 is a flowchart of the speech recognition processing according to the second embodiment of the present invention. FIG. 5 is a block diagram showing a configuration of the third embodiment of the present invention. FIG. 6 is a language model update processing flowchart according to the third embodiment of the present invention. FIG. 7A is a block diagram showing a configuration of the fourth embodiment of the present invention. FIG. 7B is a diagram illustrating the relationship between the similarity of similar words and the penalty value. FIG. 8 is a block diagram showing a configuration of the sixth embodiment of the present invention. FIG. 9A is a block diagram showing a configuration of the seventh embodiment of the present invention. FIG. 9B is a diagram illustrating the relationship between the influence level input by the user and the penalty reduction level. FIG. 10 is a flowchart of a language model update process according to the seventh embodiment of the present invention. FIG. 11 is a block diagram showing the configuration of the eighth embodiment of the present invention. FIG. 12 is a diagram illustrating a configuration of a language model used in the eighth embodiment of the present invention.

  In the present invention, the “keyword” means a word whose recognition rate is to be increased in the speech recognition apparatus as described above, and includes, for example, a word newly added to a recognition dictionary used for a language model. The term “similar word” means a word that is acoustically similar to the word. “Language score” means information obtained by quantifying the linguistic accuracy of each word in a word string, such as the probability of connection between words and the position of a word string. Such a language score includes, for example, information such as a connection probability value recorded for each word in the language model. In the present invention, “bonus” means an additional amount of the language score to be added to the language score of the keyword in order to increase the recognition rate of the keyword by a speech recognition device. “Penalty” means a discount amount of the language score to be given to the language score of the similar word in order to reduce the substitution to the similar word and the error rate of the similar word in the speech recognition apparatus.

As described above, the language model update device for a speech recognition apparatus according to the present invention includes a keyword storage unit, an influence degree calculation unit, a language score adjustment unit, and a language score update unit.
The keyword storage means stores a keyword,
The influence calculating means calculates an influence on a similar word by increasing a recognition rate of the keyword;
The language score adjusting means adjusts at least one of a bonus given to the language score of the keyword and a penalty given to the language score of the similar word based on the degree of influence,
The language score update means performs at least one of the provision of the bonus adjusted by the language score adjustment means for the language score of the keyword and the provision of the penalty adjusted by the language score adjustment means for the language score of the similar word. Thus, the language score of the language model is updated.

  According to the language model update device for a speech recognition apparatus of the present invention, the keyword is preferentially recognized in the speech recognition device, and the recognition rate of similar words of the keyword can be prevented from being excessively reduced. In other words, in the language model update device for a speech recognition device according to the present invention, the language score update means gives the bonus or the penalty to the language score of at least one of the keyword and the similar word in the language model for the speech recognition device. Can be granted. Thereby, the language score update means can make a difference in the language score of the word so that, for example, the keyword can be preferentially recognized by the speech recognition means between the keyword and the similar word. Here, for example, when a penalty is uniformly given to the language score of the similar word simply considering only the similarity of the similar word to the keyword, the recognition rate of the keyword is improved. However, on the other hand, depending on similar words, for example, the recognition rate in the speech recognition apparatus may be excessively lowered. To give a specific example, if you give a penalty to a similar word to add the word “Japanese confectionery” to the dictionary as the keyword, “I”, which is acoustically different from one phoneme, will incur a large penalty. . For this reason, instead of improving the recognition rate of “Japanese confectionery”, there is a problem that the recognition rate of “I”, which is a frequent word, deteriorates. As described above, the similar words include languages that need to be frequently used in conversation. If the recognition rate of similar words with high importance decreases excessively, the recognition rate as a whole deteriorates. However, in the language model update device for a speech recognition device according to the present invention, at least one of the bonus given to the language score of the keyword and the penalty given to the language score of the similar word can be adjusted based on the degree of influence. Therefore, it is possible to achieve both improvement in the recognition rate of the keyword and suppression of excessive reduction in the recognition rate in the similar word. That is, according to the present invention, the recognition rate of the keyword is increased, and the similar word can be recognized appropriately. The specification of the keyword and the similar word in the present invention will be described later.

  In the language model update device for a speech recognition apparatus according to the present invention, the keyword storage means stores, for example, a keyword specified by a user. The keyword storage means can extract and store the keyword from a recognition dictionary used in a speech recognition device, for example, but is not limited thereto. For example, it is preferable that the keyword storage unit can newly store the word in the recognition dictionary, for example, when a word corresponding to the keyword is not registered in the recognition dictionary. The keyword storage means is preferably capable of storing a word corresponding to the keyword. According to such a keyword storage means, a word that is not in the recognition dictionary can be recognized as a keyword in the speech recognition apparatus. In the language model update device for a speech recognition apparatus according to the present invention, by using the keyword storage unit, for example, the keyword is specified from a word input to the keyword storage unit by a user or a document including the word, and the language Model update processing can be performed.

  In the language model update device for a speech recognition apparatus according to the present invention, the similar words can be specified using, for example, a similarity measurer. The similarity measuring unit is a unit that can measure the degree of similarity between the keyword and the similar word. As such a similarity measurement means, for example, a means for measuring the acoustic similarity between the keyword and each word included in the recognition dictionary with reference to the recognition dictionary can be used. As a scale for measuring the acoustic similarity, for example, a phoneme editing distance, a distance between phonemes calculated from a mixed Gaussian distribution model (GMM) of phonemes, and the like can be used. In the language model update device for a speech recognition apparatus according to the present invention, based on the similarity measured by the similarity measurement unit, for example, a word having a high similarity can be extracted as a similar word from the recognition dictionary. The extraction of the similar words from the recognition dictionary can be performed using, for example, similar word extraction means.

  In the language model update device for a speech recognition apparatus according to the present invention, the influence degree calculating means calculates the influence degree to which the similar word is affected by increasing the recognition rate of the keyword in the speech recognition apparatus. The influence degree calculating means calculates, for example, the degree of influence that the similar word receives due to new assignment of the language score of the keyword in the language model. The degree of influence means, for example, the degree of influence that a similar word is affected by newly adding a language score of the keyword in the language model. For example, when the similar word is a word that is frequently used in a conversation, it can be said that the degree of the influence by increasing the recognition rate of the keyword in the speech recognition device is larger than the similar word that is less frequently used in the conversation. . The influence degree calculating means can calculate the influence degree based on various information about the similar words. In the language model update device for a speech recognition apparatus according to the present invention, for example, information on the similar words may be stored in a separate storage unit, and the influence degree calculating unit may use the information to make the influence degree. Can be calculated. The information is, for example, information such as the importance of a similar word, ease of recognition, and more specifically, for example, the appearance frequency of the similar word, the recognition rate of the similar word, and judgment information by the user. . For example, when the information is the appearance frequency of the similar word, the influence degree calculating means can increase the influence degree as the appearance frequency of the similar word increases. For example, when the information is the recognition rate of the similar word, the influence degree calculating unit can increase the influence degree as the recognition rate of the similar word is lower.

  In the language model update device for a speech recognition device according to the present invention, the language score adjustment unit may be configured to provide both the bonus given to the language score of the keyword and the penalty given to the language score of the similar word based on the degree of influence. Adjust either one. In the language model update device for a speech recognition apparatus according to the present invention, either the bonus or the penalty may be adjusted. However, in order to increase the overall recognition rate, it is preferable to adjust both. For example, the language score adjusting means can decrease the size of the bonus or the penalty as the influence increases. Thereby, the characteristic of each similar word can be reflected in the degree of suppression of the recognition rate of the similar word for the keyword, the recognition rate can be appropriately suppressed for each similar word, and the overall recognition rate can be improved. .

  In the language model update device for a speech recognition device according to the present invention, the language score update means includes both the adjustment bonus for the language score of the keyword and the adjustment penalty for the language score of the similar word, or Do one. In the language model update device for a speech recognition apparatus according to the present invention, the language score update unit may perform either the bonus or the penalty, but in order to increase the overall recognition rate. It is preferable to do both. Thereby, the recognition rate as a whole can be improved as mentioned above. In addition, the language score update means performs only one by both giving the adjusted bonus to the language score of the keyword and giving the adjusted penalty to the language score of the similar word. The change in the language score is smaller than in the case, and the deterioration of the recognition rate as a whole can be suppressed. The method in which the language score updating unit gives the bonus and / or the penalty to the language score in the language model is not particularly limited. That is, such a method is a method in which the speech recognition means using the language model can recognize the language score of the keyword to which the bonus is given and the language score of the similar word to which the penalty is given. Any method is acceptable. For example, the language model update device for a speech recognition device of the present invention may change the language model by giving the bonus or the penalty to the language score stored in the language model, or change the language model. Alternatively, the language score may be given by a method in which the voice recognition means can recognize the language score given the bonus or the penalty.

  In the language model update device for a speech recognition apparatus according to the present invention, the influence degree calculation means is, for example, at least one of the appearance frequency of the similar words, the recognition rate of the similar words, and judgment information by the user as described above. Based on this, the degree of influence can be calculated. The appearance frequency of the similar word can be calculated from the past recognition history of the speech recognition apparatus, for example. As the recognition rate, for example, the actual recognition rate of each similar word in the speech recognition apparatus can be used. Further, it is possible to use a recognition rate calculated by similarity prediction in which pseudo speech data is created by speech synthesis or feature amount synthesis and speech recognition is performed. The judgment information by the user is information indicating the judgment value of the influence degree by the user, for example. For example, the similarity calculation means can increase the influence as the appearance frequency of the similar word increases. The similarity calculation means can increase the influence as the recognition rate of the similar word is lower.

  The language model update device for a speech recognition device according to the present invention can be used for a language model used in the speech recognition device. The language model to which the language model update device for a speech recognition apparatus of the present invention can be applied is not particularly limited as long as it is used in the speech recognition device. The language model is data indicating statistical criteria for calculating the linguistic accuracy of a word string, and includes, for example, various pieces of information regarding words. Examples of the various information include data describing information such as the probability of easy connection between a word constituting a sentence and the next word and the probability of a word likely to appear at the beginning of a sentence from the beginning to the end of the sentence. Such a language model may be, for example, an N-gram language model using N-grams, or a model modified or extended based on the N-grams. In the present invention, the language model is preferably one whose contents can be modified as necessary. For example, one which can newly add the language score and can update the language score of a keyword is preferable.

  The language model update device for a speech recognition device of the present invention is used in a speech recognition device. The language model update device for a speech recognition device of the present invention may be included in the speech recognition device as a part of the speech recognition device of the present invention, or may be configured as a separate device, for example. Good.

  In the voice recognition apparatus of the present invention, the voice input means is means capable of acquiring a voice of a speaker. In the voice input means, for example, the user can input voice. Such voice input means is, for example, a microphone.

  In the speech recognition apparatus of the present invention, the feature amount extraction unit extracts, for example, information (referred to as a feature amount in the present invention) indicating the feature of the speech input by the speech input unit. The feature amount is information indicating the feature of the voice. Such a feature amount is not particularly limited as long as it is information that can detect a word string as a candidate in a speech recognition unit to be described later. Examples of the feature amount include information extracted based on the speech waveform of the speech, such as a cepstrum, a spectrum pitch, and power. The feature amount can be extracted using acoustic analysis such as spectrum analysis.

  In the speech recognition apparatus according to the present invention, the speech recognition means recognizes the speech as a word string using the language model from the feature quantity extracted by the feature quantity extraction means. The speech recognition means, for example, calculates the probability of a word string from the language score in the language model, for example, detects the word string with the highest probability as a candidate word string corresponding to the speech, and outputs it it can. The certainty can be calculated, for example, by adding the language scores. The speech recognition means can also perform speech recognition using an acoustic model in combination with the language model. As the acoustic model, an acoustic model normally used in a speech recognition apparatus, for example, a hidden Markov model, can be used. The speech recognition means can use another model for improving the accuracy of speech recognition as required. Examples of such other models include a grammar model.

  According to the voice recognition apparatus of the present invention, not only can the keyword be recognized by the voice recognition means with priority, but also the following is possible. That is, in the speech recognition device of the present invention, the language score adjusting means adjusts either or both of the bonus given to the language score of the keyword and the penalty given to the language score of the similar word based on the degree of influence. it can. Therefore, both improvement of the recognition rate of the keyword and suppression of excessive decrease in the recognition rate of the similar word can be realized.

Next, a method for updating a language model for a speech recognition apparatus according to the present invention will be described.
As described above, the language model update method for a speech recognition apparatus of the present invention uses a keyword storage unit, an influence degree calculation unit, a language score adjustment unit, and a language score update unit.
A keyword storage step (a) in which the keyword storage means stores a keyword;
An influence degree calculating step (b) in which the influence degree calculating means calculates an influence degree on a similar word by increasing the recognition rate of the keyword;
A language score adjusting step (c) in which the language score adjusting means adjusts at least one of a bonus given to the language score of the keyword and a penalty given to the language score of the similar word based on the degree of influence;
The language score update means performs at least one of the provision of the bonus adjusted by the language score adjustment means for the language score of the keyword and the provision of the penalty adjusted by the language score adjustment means for the language score of the similar word. Thus, a language score update step (d) for updating the language score of the language model is included.

  According to the language model update method for a speech recognition device of the present invention, the keyword is recognized by the speech recognition device with priority in the language model, and the recognition rate of similar words of the keyword is excessively reduced. The language model can be updated so that there is no problem. That is, in the language model update method for a speech recognition apparatus of the present invention, both or either of the bonus given to the language score of the keyword and the penalty given to the language score of the similar word by the language score adjustment step of (c). This can be adjusted according to the influence degree calculated by the influence degree calculating means. Therefore, according to the language model update method for a speech recognition apparatus of the present invention, it is possible to realize both the improvement of the keyword recognition rate and the suppression of the excessive decrease in the recognition rate of the similar words.

  The language model update method for a speech recognition apparatus according to the present invention can be implemented using, for example, the language model update apparatus for a speech recognition apparatus according to the present invention. That is, the language model update method for a speech recognition apparatus according to the present invention includes, for example, the language score update means, the influence degree calculation means, and the language score adjustment means in the language model update apparatus for the speech recognition apparatus according to the present invention. And using the language score update means. The configurations and functions of these means are as described in the description of the language model update device for a speech recognition device of the present invention.

  In the language model update method for a speech recognition apparatus according to the present invention, in the step (c), the language score adjusting means preferably decreases the bonus or the penalty as the influence increases. Thereby, the characteristic of each similar word can be reflected in the degree of suppression of the recognition rate of the similar word for the keyword, the recognition rate can be appropriately suppressed for each similar word, and the overall recognition rate can be improved. .

  In the language model update method for a speech recognition apparatus according to the present invention, in the language score update step of (d), the language score update means assigns the adjusted bonus to the language score of the keyword and the language of the similar word Both and / or any of the adjustment of the adjusted penalty for the score is performed. Thereby, the recognition rate as a whole can be improved as mentioned above. The language score updating means performs only one of the processes by both giving the adjusted bonus to the language score of the keyword and giving the adjusted penalty to the language score of the similar word. The change in the language score is smaller than in the case, and the deterioration of the recognition rate as a whole can be suppressed. In the language score update means, a method of updating the language model by giving either or either of the bonus and the penalty to the language score is not particularly limited. That is, any method can be used as long as the speech recognition apparatus using the language model can recognize the language score after giving the bonus or the penalty to the language score before keyword storage. It may be a method. For example, the language model update device for a speech recognition device according to the present invention may give the bonus or the penalty to the language score recorded in the language model, or the speech recognition device may change the language model without changing the language model. You may give by the other method which can recognize the language score after giving a bonus or the said penalty.

  In the language model update method for a speech recognition apparatus according to the present invention, in the step (b), for example, in the influence calculation means, the appearance frequency of the similar words, the recognition rate of the similar words, and the judgment information by the user The influence degree can be calculated based on at least one. The appearance frequency of the similar word, the recognition rate of the similar word, and the determination information by the user are as described for the language model update device for a speech recognition device of the present invention.

  In the language model updating method for a speech recognition apparatus according to the present invention, in the step (b), the influence degree can be increased in the influence degree calculation means, for example, according to the appearance frequency of the similar word.

  In the language model updating method for a speech recognition apparatus according to the present invention, in the step (b), the influence calculation means can increase the influence according to the recognition rate decrease of the similar words.

The speech recognition method of the present invention can be implemented using the language model update method for speech recognition apparatus of the present invention.
As described above, the speech recognition method of the present invention uses the speech recognition device of the present invention,
Updating the language model by the language model updating method for a speech recognition apparatus of the present invention;
A feature quantity extraction step (e) in which the feature quantity extraction means extracts a feature quantity of the voice input by the voice input means;
The voice recognition means includes a voice recognition step (f) for recognizing the voice using the language model from the feature quantity extracted by the feature quantity extraction means.

  The speech recognition method of the present invention can not only recognize the keyword preferentially, but also can: That is, in the speech recognition method of the present invention, in the step (c), either or both of the bonus given to the language score of the keyword and the penalty given to the language score of the similar word are set according to the degree of influence. Can be adjusted. Therefore, it is possible to achieve both improvement in the recognition rate of the keyword and suppression of excessive decrease in the recognition rate of the similar word.

  The speech input means, feature amount extraction means, speech recognition means, and language model used in the speech recognition method of the present invention include the speech input means, feature amount extraction means, speech recognition means in the speech recognition apparatus of the present invention. And language model can be used respectively. The configurations and functions of the voice input unit, the feature amount extraction unit, the voice recognition unit, and the language model are as described in the description of the voice recognition device of the present invention. The voice recognition method of the present invention may further use other means for improving the accuracy of voice recognition. The configuration and function of such other means are as described in the description of the speech recognition apparatus of the present invention.

  The language model updating method for a speech recognition apparatus according to the present invention may further include steps other than the steps (a) to (d), or may not include the steps. In the language model updating method for a speech recognition apparatus of the present invention, the order in which the steps are performed is not particularly limited as long as the effects of the present invention are obtained. Moreover, the speech recognition method of the present invention may or may not include steps other than the steps (a) to (f). In the speech recognition method of the present invention, the order in which the steps are performed is not particularly limited as long as the effects of the present invention are obtained.

The computer program of the present invention is a computer program capable of executing on a computer at least one of the language model update method for a speech recognition apparatus of the present invention and the speech recognition method of the present invention.
The recording medium of the present invention is a recording medium storing the computer program of the present invention.

Hereinafter, more specific embodiments of the present invention will be described with reference to the drawings. However, the present invention is not limited to the following embodiments.
[Embodiment 1]
The block diagram of FIG. 1 shows the configuration of a language model update device for speech recognition apparatus (hereinafter simply referred to as “language model update device”) of the present embodiment. As shown in the figure, the language model update device 100 of the present embodiment includes a keyword storage unit 101, an influence degree calculation unit 105, a similar word extraction unit 102, a language score adjustment unit 104, and a language score update unit 106. Including. The language model update device 100 according to the present embodiment communicates with a language model 107 and a recognition dictionary 108 used in the speech recognition device. The language model 107 records at least the language score of each word stored in the recognition dictionary 108. The recognition dictionary 108 includes notation and reading information of words to be recognized. The language model updating apparatus 100 according to the present embodiment can give a bonus to the language score of the keyword in the language model 107 so that the keyword can be recognized with priority over similar words. Penalize the language score of similar words. That is, the language model update device 100 according to the present embodiment can update the language model 107. The language model update device 100 according to the present embodiment includes, for example, at least a CPU (Central Processing Unit), a ROM (Read Only Memory), and a RAM (Random Access Memory). This can be realized by the CPU. That is, for example, by incorporating a computer program that provides each function of each means into the CPU, the means in the language model update device 100 can be constructed, and the language model update device 100 of the present invention can be realized. In addition, the language model updating apparatus 100 according to the present embodiment is implemented by mounting a circuit component composed of a hardware component such as an LSI (Large Scale Integration) in which a computer program for realizing the function of each unit is incorporated. It can also be realized. Such a computer program can also be used in the form of a recording medium storing the computer program, for example, HDD, FD, CD-ROM (CD-R, CD-RW), MO, DVD, memory card, etc. . With reference to the drawings, the operation of the language model update apparatus 100 of the present embodiment will be described.

  When the user wants the speech recognition apparatus using the language model 107 to recognize a specific word as a keyword, in the present embodiment, the language model 107 is processed by the following processes (1-1) to (1-5). Can be updated.

(1-1) Keyword Storage Processing First, a user inputs a word (hereinafter referred to as a keyword) to be recognized or a sentence including the keyword to the keyword storage unit 101. Based on this input, the keyword storage means 101 extracts the keyword from the recognition dictionary 108 and stores it. In the present embodiment, for example, when the keyword is not registered in the recognition dictionary 108, the keyword storage unit 101 can store the keyword as it is. Next, the similar word extraction means 102 performs the similar word extraction process (1-2) below.

(1-2) Similar Word Extraction Processing The similar word extraction unit 102 extracts a word having a certain degree of similarity with the keyword from the recognition dictionary 108 as a similar word of the keyword. Next, the influence degree calculation means 105 performs the influence degree calculation process (1-3) below.

(1-3) Influence Level Calculation Processing The degree of influence calculation unit 105 calculates the degree of influence on the recognition rate of the similar words as the language model 107 increases the language score of the keyword. Subsequently, the language score adjustment means 104 performs the language score adjustment process of the following (1-4).

(1-4) Language Score Adjustment Processing The bonus given to the language score of the keyword in the language model 107 and the language of the similar word according to the influence degree calculated by the influence degree calculation means 105 by the language score adjustment means 104 Adjust both and / or penalties for the score. Further, the language score adjusting unit 104 outputs information about the adjusted bonus and the penalty to the language score updating unit 106. The language score adjusting unit 104 can adjust the language score according to the similar word, for example, the similar word having a larger influence degree reduces the bonus and the penalty.

(1-5) Language Score Update Processing The language score update means 106 that has received information on both or either of the bonus and the penalty uses the keyword and the similar word in the language model 107 based on the information. The bonus or the penalty adjusted by the language score adjusting means 104 is given to both or any of the language scores. That is, the language score update unit 106 updates the language model 107.

  According to this embodiment, the language model 107 used by the speech recognition apparatus can be updated so that keywords can be recognized with priority over similar words. However, in the present embodiment, either or both of the bonus and the penalty are adjusted according to the influence calculated by the influence calculation means 105. For this reason, for example, when a similar word is an important word or a word with a low recognition rate, the recognition rate of the similar word does not decrease too much. That is, according to the present embodiment, such similar words can also be recognized appropriately, and the recognition rate as a whole of the speech recognition apparatus can be improved.

In the present embodiment, the language score adjusting means 104 can be configured so that the bonus or penalty can be reduced as the influence increases. Thereby, it is possible to reliably prevent an excessive decrease in the recognition rate of important similar words.
In the present embodiment, in order to increase the recognition rate of the keyword, both the adjustment of the adjusted bonus and the adjustment of the penalty may be performed on the keyword. Thereby, the change of a language score becomes small compared with the case where only any one process is performed, and the deterioration of the recognition rate as a whole can be suppressed.

  In the present embodiment, the language score updating method by the language score updating unit 106 in the language score updating process (1-5) is not particularly limited. For example, the language score updating unit 106 may give the bonus or the penalty by changing the language score recorded in the language model 107, or the language model 107 without changing, for example, The language model 107 can be updated in other ways in which the speech recognizer can recognize the language score with the bonus or penalty.

FIG. 2 shows a flow chart of the language model update method for the speech recognition apparatus of this embodiment. As shown in the figure, the language model update method for a speech recognition apparatus of the present embodiment (hereinafter simply referred to as “language model update method”) includes the following steps (a) to (d).
(A) inputting and storing the keyword in the keyword storage means 101 (step S1);
(B) In the influence degree calculation means 105, a step of calculating an influence degree on a similar word by increasing the recognition rate of the keyword in the speech recognition device (step S2);
(C) The language score adjusting unit 104 adjusts either or both of the bonus given to the language score of the keyword and the penalty given to the language score of the similar word based on the degree of influence (step S3). ,
(D) In the language score update unit 106, the bonus adjusted in the language score adjusting unit 104 for the language score of the keyword and the penalty adjusted in the language score adjusting unit 104 for the language score of the similar word A language score update step (step S4) for updating the language score of the language model 107 by performing both or any of the above.

[Embodiment 2]
The present embodiment is an example of a speech recognition device using the language model update device of the first embodiment. The block diagram of FIG. 3 shows the configuration of the speech recognition apparatus of this embodiment. As shown in the figure, the speech recognition apparatus of the present embodiment has the same configuration as the speech input unit 211, the feature amount extraction unit 210, the speech recognition unit 209, the recognition dictionary 208, the language model 207, and the first embodiment. And a language model update device 200. The language model 207 and the recognition dictionary 208 are each stored in general storage means such as a hard disk. The recognition dictionary 208 includes notation and reading information of words to be recognized. The language model 207 includes data that models the appearance frequency, connection probability, and the like of words, morphemes, and phonemes learned from a large amount of speech data. The speech recognition apparatus according to the present embodiment can be realized, for example, by including at least a CPU, a ROM, and a RAM, and controlling the processing by each means by the CPU. That is, for example, each means in the speech recognition apparatus can be constructed by incorporating a computer program that provides each function of each means in the CPU, and the speech recognition apparatus of the present invention can be realized. In addition, the voice recognition device of the present embodiment can also realize its operation by mounting a circuit component made up of hardware components such as an LSI, which incorporates a computer program for realizing the functions of the respective means. Such a computer program can also be used in the form of a recording medium storing the computer program, for example, HDD, FD, CD-ROM (CD-R, CD-RW), MO, DVD, memory card, etc. . With reference to FIG. 3, the operation of the speech recognition apparatus of this embodiment will be described.

In the present embodiment, when performing speech recognition without specifying a keyword, the speech recognition apparatus of the present embodiment performs processing as follows. First, when a user inputs a voice to the voice input unit 211, the feature quantity extraction unit 210 performs a feature quantity extraction process (2-1) below.
(2-1) Feature Extraction Processing The feature extraction unit 210 analyzes the input speech and extracts a feature amount, and outputs the extracted feature amount to the speech recognition unit 209. Next, the voice recognition unit 209 performs the following voice recognition process (2-2).

(2-2) Speech Recognition Processing The speech recognition means 209 detects a word string that can be determined to be most likely based on the feature amount and uses the language model 207, and outputs it as a recognition result.

(2-3) Keyword Recognition Processing When the user wants the speech recognition device to recognize a specific word as a keyword, the speech recognition device of the present embodiment uses the keyword storage processing (1-1) in the first embodiment. Thus, the keyword storage unit 201 stores the keyword.

(2-4) Language Model Update Process After the keyword storage process, the language model update apparatus 200 updates the language model 207 by the processes (1-2) to (1-5) of the first embodiment. Update.

(2-5) Keyword Recognition Process After the language model update process (2-4), the user inputs voice to the voice input unit 211. Based on this input, after the processing by the feature amount extraction unit 210, the speech recognition unit 209 performs speech recognition processing using the updated language model 207 and outputs a recognition result.

  In the present embodiment, in the language model update device 200, the bonus or the penalty is added to the language score of both the keyword and the similar word or any language score with respect to the language model 207 with reference to before the keyword is recorded. Has been granted. Therefore, the voice recognition unit 209 can recognize the keyword preferentially over the similar word. However, in the present embodiment, both or either of the bonus given to the language score of the keyword and the penalty given to the language score of the similar word are adjusted according to the degree of influence calculated by the influence degree calculating unit 205. It is. Thereby, in this embodiment, when the said similar word is an important word, or when it is a word with a low recognition rate, the recognition rate of a similar word is not reduced too much, for example. Therefore, in the present embodiment, such similar words can also be recognized appropriately, and the recognition rate as a whole of the speech recognition apparatus can be improved.

FIG. 4 shows a flowchart of the speech recognition method of this embodiment. As shown in the figure, the speech recognition method of the present embodiment includes the following steps (e) to (f) in addition to the steps (a) to (d) in the first embodiment.
(E) a step of the feature extraction unit 210 analyzing the voice input from the voice input unit 211 and extracting a feature amount (step S5);
(F) The voice recognition means recognizes and outputs the voice as a word string using the language model 207 based on the feature amount (steps S6 and S7).

[Embodiment 3]
FIG. 5 shows the configuration of the language model update device 300 of this embodiment. As shown in the figure, the language model update device 300 of the present embodiment has the same configuration as that of the first embodiment except that it further includes a similarity measurer 303. The similarity measurer 303 can measure the acoustic similarity between words, and can measure the acoustic similarity between the keyword and each word registered in the recognition dictionary 308. With reference to FIG. 5, the operation of the language model update device 300 of this embodiment will be described.

When the user wants the speech recognition apparatus using the language model 307 to recognize a specific word as a keyword, in this embodiment, the language model 307 can be updated as follows.
(3-1) Keyword Storage Processing First, the user inputs an unregistered word (keyword) to be recognized or a sentence including the word to the keyword storage unit 301. Based on this input, the keyword storage unit 301 stores the keyword by the same keyword storage processing as in the first embodiment. Next, the similarity measurer 303 performs the similarity measure process (3-2) below.

(3-2) Similarity Measurement Processing The similarity measurement unit 303 refers to the recognition dictionary 308 and measures the acoustic similarity between the keyword and each word included in the recognition dictionary 308. After the similarity measurement process, the similar word extraction unit 302 performs the following similar word extraction process (3-3).

(3-3) Similar Word Extraction Processing The similar word extraction unit 302 extracts a word having a certain degree of similarity with the keyword as a similar word that is acoustically similar to the keyword. Next, the influence degree calculation unit 305 performs the influence degree calculation process (3-4) below.

(3-4) Influence Level Calculation Processing The influence level calculation unit 305 calculates the degree of influence on the recognition rate of the similar word by increasing the recognition rate of the keyword in the speech recognition apparatus. Subsequently, the language score adjustment means 304 performs the language score adjustment process of the following (3-5).

(3-5) Language Score Adjustment Processing In this embodiment, the language score adjustment unit 304 adjusts either or both of a bonus given to the language score of the keyword and a penalty given to the language score of the similar word. The language score adjusting unit 304 adjusts the bonus or the penalty so as to decrease as the influence degree increases. Next, information indicating the bonus or the penalty is output to the language score updating means 306.

(3-6) Language model update processing Upon receiving the output of the information, the language model update device 306 gives the adjusted bonus to the keyword and applies it to the similar word in order to increase the recognition rate of the keyword. Both the adjusted penalty may be given. Either one of them may be performed, but if both are performed, the change in the language score becomes smaller, and the deterioration of the recognition rate as a whole can be suppressed.

  According to this embodiment, the language model used by the speech recognition apparatus can be updated so that keywords can be recognized with priority over similar words, and the recognition rate of important similar words is excessively reduced. Can be properly recognized.

FIG. 6 shows a flowchart of the speech recognition method of the present embodiment. As shown in the figure, the speech recognition method of the present embodiment is a language model update method that further includes the following step (g) in the steps (a) to (d) in the first embodiment.
(G) The similarity measuring unit 303 refers to the recognition dictionary 308 and measures the acoustic similarity between the keyword and each word included in the recognition dictionary 308 (step S2 ′).

[Embodiment 4]
FIG. 7A shows the configuration of the language model update device 400 of this embodiment. The language model update device 400 according to the present embodiment has the same configuration as that of the third embodiment except that the appearance number storage unit 410 is included. That is, as shown in the figure, the language model update device 400 of this embodiment includes an appearance count storage unit 410 in the influence degree calculation unit 405. The appearance number storage means 410 stores a recognition result at the time of operation of the speech recognition apparatus before the keyword storage processing as a recognition history. With reference to FIG. 7A, operation | movement of the language model update apparatus 400 of this embodiment is demonstrated.

(4-1) Keyword Storage Processing First, the user inputs an unregistered word (keyword) to be recognized or a sentence including the word to the keyword storage unit 401. Based on this input, the keyword storage unit 401 stores the word corresponding to the keyword by the same keyword storage processing as in the first to third embodiments. Next, the similarity measurer 403 performs the similarity measure process (4-2) below.

(4-2) Similarity Measurement Processing The similarity measurement unit 403 refers to the recognition dictionary 408 and measures the acoustic similarity between the keyword and each word included in the recognition dictionary 408. In the present embodiment, the similarity measuring unit 403 measures the similarity using the phoneme editing distance. Specifically, for example, it is assumed that the user inputs “Japanese confectionery” in the keyword storage unit 401. The phoneme notation of the “Japanese confectionery” is “wagasi”, the phoneme notation of “I”, which is one of the words in the recognition dictionary 408, is “watasi”, and there is a difference in replacement of one phoneme in the phoneme notation. Yes, the similarity is “1”. For example, the phoneme notation of the word “candy” is “okasi”, and compared to “wagasi”, there is a difference in dropping one phoneme and replacing two phonemes, and the similarity is “3” in total. To do. When the phoneme editing distance is used as the similarity, the similarity is smaller as the two words are acoustically similar. After the similarity measurement process, the similar word extraction unit 402 performs the following similar word extraction process (4-3).

(4-3) Similar Word Extraction Processing The similar word extraction unit 402 extracts a word having a certain degree of similarity with the keyword as a similar word that is acoustically similar to the keyword. Next, the influence degree calculation means 405 performs the influence degree calculation process (4-4) below.

(4-4) Influence degree calculation process The influence degree calculation means 405 uses the appearance frequency storage means 410, for example, in a speech recognition device, to increase the recognition rate of the keyword to the similar word recognition rate. Calculate the impact. For example, it is assumed that “I” appears 30 times and “sweets” appears twice as the recognition history in the appearance count storage unit 410. The influence degree calculating means 405 calculates the influence degree based on the number of appearances. Further, the influence degree calculating unit 405 outputs the result to the language score adjusting unit 404.

(4-5) Language Score Adjustment Processing The language score adjustment unit 404, for example, gives a penalty to similar words when “Japanese confectionery” is added as a registered word by the user. Adjust to give. That is, in the present embodiment, the language score adjusting unit 404 calculates a penalty value (Pi) given to the language score of the similar word according to the following formula (I) using the degree of influence.

Pi = S (sim (w, wi)) − αlogC (wi) (I)

In the formula (I), w is a keyword, wi is a similar word, sim (a, b) is a similarity between the words a and b, and S (sim) is a function for determining a penalty value Pi from the similarity. is there. C (w) is the number of appearances of the word w in the recognition history, and is the influence degree calculated in the influence degree calculation process. α is a constant coefficient. The similarity sim uses the phoneme editing distance as described above. FIG. 7B illustrates the relationship between S (sim) calculated by the language score adjusting unit 404 and the similarity of the similar words. As shown in the figure, S is a function in which the penalty increases as the two words are similar, that is, as the similarity value is smaller. For example, S (sim) when the similarity is 1 is When the similarity is 1.0, it can be 0.25. In the present embodiment, by multiplying the logarithm of the number of appearances by a coefficient from the S (sim), the penalty can be reduced for similar words having a larger number of appearances. For example, when α = 0.1, the penalty value of “I” is P = 1.0−0.1 × log 30 = 1.0−0.1 × 3.401 = 0.6599. The penalty value of “confectionery” is P = 0.25−0.1 × log 2 = 0.25−0.1 × 0.6931 = 0.8069. After calculating the penalty value, the language score adjustment unit 404 outputs information indicating the penalty value to the language model update device 406.

(4-6) Language model update process After the language score adjustment process, the language model update device 406 uses the information before the keyword storage process as a reference for the language score of each similar word in the language model 407 based on the information. Give the penalty. That is, the language model update device 406 updates the language model 407.

  In the present embodiment, the language score adjusting unit 404 gives a penalty to the language score of each similar word in the language model 407. For this reason, the speech recognition apparatus using the language model 407 can recognize the keyword preferentially over similar words. However, in the present embodiment, the influence degree calculation means 405 calculates the influence degree based on the information stored in the number of times detection means 410, and adjusts the magnitude of the penalty according to the influence degree. . Therefore, for example, when similar words are important words or words with low recognition rate, the recognition rate of similar words does not decrease too much, and such similar words are also recognized appropriately. And the recognition rate as a whole can be improved.

  In this embodiment, the similarity measurer 403 can use a phoneme distance calculated from a phoneme GMM instead of the phoneme editing distance. In this case, it is possible to reduce the similarity value, assuming that the words having acoustically similar phonemes are more similar.

The language model update method of the present embodiment can be described with the same flowchart as that of the first embodiment. As shown in FIG. 2, the language model update method of this embodiment is a language model update method including the following steps (a) to (d).
(A) the keyword storage unit 401 stores a word corresponding to the keyword (step S1);
(B) In the influence degree calculating means 405, using the appearance number storage means 410, calculating the influence degree with respect to similar words by increasing the recognition rate of the keyword in the speech recognition device (step S2);
(C) The language score adjusting means 404 calculates a penalty value (Pi) to be given to the language score of the similar word according to the formula (I) (step S3),
(D) In the language score update unit 406, in the language model 407, the penalty value is assigned to the language score of the similar word, and the language score is updated (step S4).

[Embodiment 5]
The configuration of the present embodiment is the same as that of the fourth embodiment except that a bonus can be given to the language score of the keyword according to the number of occurrences of similar words in the recognition history. The language score adjusting unit 404 can calculate not only the penalty value (Pi) based on the influence level but also a bonus value (B) to be given to the language score of the keyword. Further, the language score updating means 406 can not only give a penalty value (Pi) to the language score of the similar word but also give the bonus value to the language score of the keyword. With reference to FIG. 7A, operation | movement of the speech recognition apparatus of this embodiment is demonstrated. First, the process from when the user inputs an unregistered word (keyword) to be recognized or a sentence including the word to the keyword storage unit 401 until calculating the influence degree is the same as that in the fourth embodiment (( 4-1) to (4-4) are the same. In the present embodiment, the following (5-1) language score adjustment process is performed after the influence degree calculation process (4-4).

(5-1) Language Score Adjustment Processing The language score adjustment unit 404 follows the above formula (I) using the degree of influence in the same manner as the language score adjustment processing of (4-5) in the fourth embodiment. Then, a penalty value (Pi) to be given to the language score of the similar word is calculated. In the present embodiment, the language score adjusting unit 404 calculates a bonus value (B) to be given to the language score of the keyword according to the following formula (IV) using the penalty value. Next, the language score adjustment unit 404 outputs information indicating the penalty value and the bonus value to the language model update device 406.

B = b−βmaxPi (IV)

In the formula (IV), b is a constant bonus, β is a coefficient, and maxPi is the maximum value of the penalty Pi of the similar word. Instead of maxPi, an average value of the penalty of the similar word may be used.

(5-2) Language Model Update Processing After the language score adjustment processing, the language model update device 406 determines the penalty value and the bonus for the language scores of similar words and keywords in the language model 407 based on the information, respectively. Assign a value.

  In the present embodiment, the adjustment of the keyword is also performed by the bonus value, so that the change in the adjusted language score is smaller than the adjustment using only the penalty value, and the adverse effect on the recognition rate as a whole is reduced. Can be suppressed.

The language model update method according to the present embodiment includes the steps (a) to (d) in the fourth embodiment, and further includes the following (c ′) and the following (d ′).
(C ′) The language score adjusting unit 404 calculates the bonus value using the formula (IV) and outputs information indicating the bonus value to the language model update device 406.
(D ′) The language model update device 406 gives the bonus value to the language score of the keyword in the language model 407 based on the information.

[Embodiment 6]
In the present embodiment, an example is shown in which both or one of the penalty given to the language score of the similar word and the bonus given to the language score of the keyword can be adjusted according to the recognition rate of the similar word. The configuration of the language model update device of the present embodiment is the same as that of the fifth embodiment except that it includes a similar word recognition rate storage unit 610 instead of the appearance number storage unit 410. FIG. 8 is a block diagram showing the configuration of the language model update apparatus 600 of this embodiment. As illustrated, in the present embodiment, the language model update device 600 includes a similar word recognition rate storage unit 610 in the influence degree calculation unit 605. In the present embodiment, the similar word recognition rate storage unit 610 stores the actual recognition rate of each similar word in the speech recognition apparatus before storing keywords. With reference to FIG. 8, the operation of the language model update device 600 of this embodiment will be described. First, a process from when a user inputs an unregistered word (keyword) to be recognized to the keyword storage unit 601 or a sentence including the word until a similar word is extracted is described in the fourth and fifth embodiments. The same as (4-1) to (4-3) above. In this embodiment, after the similar word extraction process of (4-3), the influence degree calculation process of the following (6-1) is performed.

(6-1) Influence Level Calculation Processing Influence on the similar word recognition rate by the influence degree calculating unit 605 using the similar word recognition rate storage unit 610 to increase the keyword recognition rate in the speech recognition apparatus. The degree is calculated and the result is output to the language score adjusting means 604. Subsequently, the language score adjustment means 604 performs the language score adjustment process of the following (6-2).

(6-2) Language Score Adjustment Processing The language score adjustment unit 604 calculates a penalty value (Pi) to be given to a language score of a similar word according to the following formula (II) using the degree of influence, and the penalty Information indicating the value is output to the language model update device 606.

Pi = S (sim (w, wi)) × P (wi) (II)

In the formula (II), P (w) is a recognition rate of the word w. That is, P (wi) is the recognition rate of the similar words, and the degree of influence increases when P (wi) decreases. The penalty value Pi decreases as the degree of influence increases. That is, P (wi) and Pi have a positive correlation. For example, if the recognition rate of “I” is 90%, P = 1.0 × 0.9 = 0.9, and if the recognition rate of “sweets” is 70%, P = 0.25 × 0.7 = 0.175, and the penalty is reduced as the word has a lower recognition rate.

(6-3) Language model update processing After the language score adjustment processing, the language model update device 606 uses the penalty value in the language score of the similar word before the keyword storage processing in the language model 607 based on the information. Is granted. That is, the language model update device 606 updates the language score.

  In the present embodiment, the following effects can be obtained. That is, similar words with a low recognition rate are greatly reduced by giving a penalty. For example, by increasing the degree of influence and reducing the penalty, the recognition rate of the similar word is excessively lowered. Can be prevented.

  In the present embodiment, as the recognition rate, a recognition rate obtained by recognition rate prediction may be used instead of the actual recognition rate. The recognition rate obtained by the recognition rate prediction can be calculated, for example, by creating pseudo speech data and performing speech recognition by speech synthesis or feature amount synthesis.

  In the present embodiment, as in the fifth embodiment, the language score adjusting unit 604 may further calculate a bonus value (B) to be given to the language score of the keyword. Thereby, the language score update means 606 can give the bonus value to the language score of the keyword. Thereby, the change of the adjusted language score can be made smaller than the case where the similar word is adjusted only by the penalty value, and the adverse effect on the recognition rate as a whole can be suppressed.

The language model update method of the present embodiment includes the steps (a) to (d) in the fourth embodiment, and the step (c) includes the following step (c ″).
(C ″) The language score adjusting unit 604 calculates a penalty value (Pi) given to the language score of the similar word according to the formula (II) using the degree of influence, and shows information indicating the calculated penalty value. Outputting to the language model update device 606;

[Embodiment 7]
In the present embodiment, an example is shown in which the user can determine the degree of influence on a similar word and adjust the penalty given to the language score of the similar word. The configuration of the language model update device of this embodiment is the same as that of Embodiment 4 except that it includes similar word presentation means 710 and influence degree input means 711. FIG. 9A is a block diagram showing the configuration of the language model update apparatus 700 of this embodiment. As shown in the figure, in the present embodiment, the language model update device 700 includes a similar word presentation unit 710 in the similar word extraction unit 702 and an influence degree input unit 711 in the influence degree calculation unit 705. In the present embodiment, the user can input the influence degree determined by the user using the influence degree input means 711. With reference to FIG. 9A, the operation of the language model update apparatus 700 of this embodiment will be described. First, the process from when the user inputs an unregistered word (keyword) to be recognized or a sentence including the word to the keyword storage unit 701 until the similar word is extracted is the same as that in the fourth embodiment ( 4-1) to (4-3) are the same. In the present embodiment, after the similar word extraction process (4-3) by the similar word extraction means 702, the extracted similar words are presented to the user by the similar word presentation means 710. The user can perform the influence degree judgment value input process (7-1) as desired.

(7-1) Influence Level Input Process With respect to the similar words presented by the similar word presentation unit 710, the user can input the degree of influence information determined to be appropriate in the influence level input unit 711. For example, the influence degree information can be evaluated in five stages, and 5 can be input to reduce the penalty, and 1 can be input to increase the penalty. Upon receiving the input, the influence degree calculating means 705 calculates the influence degree based on the influence degree information, and outputs the calculated influence degree to the language score adjusting means 704.

(7-2) Language Score Adjustment Processing The language score adjustment unit 704 that receives the influence degree calculates a penalty value (Pi) given to the language score of the similar word according to the following formula (III) using the influence degree. To do.

Pi = S (sim (w, wi)) × T (ui) (III)

In the formula (III), ui is a degree of influence input by the user, T represents a penalty reduction degree, and is a function that decreases as the degree of influence increases as shown in FIG. 9B. For example, when the user inputs “5” as the influence degree of “I”, the penalty is P = 1.0 × 0.1 = 0.1, and the influence degree of “sweets” is “3”. The penalty of time can be P = 0.25 × 0.5 = 0.125. After calculating the penalty value, the language score adjusting unit 704 outputs information indicating the penalty value to the language model update device.

(7-3) Language Model Update Processing After the language score adjustment processing, the language model update device 706 gives the penalty value to the language score of similar words in the language model 707 based on the information. That is, the language model update device 706 updates the language score.

  In the present embodiment, in addition to the effect achieved by the fourth embodiment, the penalty value can be adjusted by the user's judgment, and the effect that voice recognition that best meets the user's wish can be realized.

  In this embodiment, the user can input a bonus value to the language score of the keyword in the influence input device 711. In this case, the language score adjusting unit 704 can calculate a bonus value to be given to the language score of the keyword in the language model 707 based on the influence level input by the user in the influence level input unit 711. Then, the language score updating means 706 can add and update the bonus value to the language score of the keyword. Thereby, the change of a language score can be made smaller than a user adjusting only a penalty freely, and the bad influence on the recognition rate as a whole can be suppressed.

FIG. 10 shows a flowchart of the language model update method of the present embodiment. As shown in the figure, the language model update method of the present embodiment includes the following steps (h) in steps (a) to (d) in the fourth embodiment, and the steps (c) are described below. This is a method performed by the step (c ″ ′).
(H) A step of inputting influence degree information to be judged appropriate by the influence degree input means 711 in the influence degree input means 711 for the similar words presented by the similar word presentation means 710 at the user (step S3 ′).
(C ″ ′) The language score adjusting unit 704 that has received the output from the influence degree calculating unit 705 uses the influence degree to assign a penalty value (Pi) to the language score of the similar word according to the formula (III). ).

[Eighth embodiment]
In the present embodiment, yet another example of the speech recognition apparatus and speech recognition method of the present invention is shown. FIG. 11 shows the configuration of the speech recognition apparatus of this embodiment. As shown in the figure, the speech recognition apparatus of this embodiment includes a speech input means 811, a feature quantity extraction means 810, a speech recognition means 809, an acoustic model 813, a recognition dictionary 808, a language model 807, and the implementation described above. Language model updating apparatus 800 according to the third aspect. The acoustic model 813, the language model 807, and the recognition dictionary 808 are each stored in a general storage unit such as a hard disk. In the present embodiment, a hidden Markov model (HMM) is used as the acoustic model 813. The recognition dictionary 808 includes notation and reading information of a word that is a recognition target of the voice recognition unit 809. The language model 807 is data obtained by modeling the appearance frequency, connection probability, and the like of words, morphemes, and phonemes learned from a large amount of speech data. As the language model 807, for example, an N-gram language model can be used. In the present embodiment, the language model 807 is expanded based on a general N-gram so that the language model update device 800 can identify and / or identify keywords and / or similar words. FIG. 12 illustrates the contents of the language model used in the present embodiment. As shown in the figure, the language model 807 describes from the beginning of the sentence to the end of the sentence, the probability of easy connection between the word constituting the document and the word following the word, and the probability of the word that is likely to be located at the beginning of the sentence. . In the present embodiment, in the language model 807, for example, identification information can be newly given to each of the keyword and the similar word. In this embodiment, the language model 807 can record a language score for each word stored in the recognition dictionary 808. In the present embodiment, an example in which a connection probability value is newly given as a language score will be described. In other words, in the present embodiment, the language model 807 is provided with an area (A) for storing a bonus or penalty to be given to the probability value to which the word is connected, as shown. Specifically, when the language model updating unit 800 gives a bonus to the language score, the language model 807 records a positive numerical value (B) in the area (A), and adds the language score to the language score. When a penalty is given, a numerical value (Pi) of − can be recorded in the area (A). With reference to FIG. 11 and FIG. 12, the operation of the speech recognition apparatus of this embodiment will be described.

(8-1) Normal Speech Recognition Processing When performing speech recognition without specifying a keyword, the speech recognition apparatus according to the present embodiment performs processing as follows. First, when the user inputs a voice to the microphone 811 which is a voice input unit, the feature extraction unit 810 performs a feature amount extraction process (8-2) below.
(8-2) Feature Amount Extraction Processing The feature extraction unit 810 analyzes the input speech and extracts a feature amount, and outputs the extracted feature amount to the speech recognition unit 809. Next, the voice recognition unit 809 performs voice recognition processing (8-3) below.
(8-3) Speech Recognition Processing The speech recognition means 809 calculates a probability value indicating the similarity with the phoneme modeled in the acoustic model 813 based on the feature amount. In addition, the speech recognition unit 809 searches for word string candidates from the language model 807, adds a connection probability for each word, and adds a probability value calculated using the acoustic model 813, The word string having the maximum connection probability value is output as the most likely word string.

(8-4) Keyword Recognition Processing When the user wants the speech recognition apparatus to recognize a specific language as a keyword, the keyword storage unit 801 does not want to recognize an unregistered word (keyword) or a sentence containing the word Enter. Subsequently, the keyword storage unit 801 stores the word corresponding to the keyword by the same keyword storage process as in the third embodiment. Next, the similarity measurer 803 performs the similarity measure process (8-5) below.

(8-5) Similarity Measurement Processing The similarity measurement means 803 refers to the recognition dictionary 808 and measures the acoustic similarity between the keyword and each word included in the recognition dictionary 808. After the similarity measurement process, the similar word extraction unit 802 performs the similar word extraction process (8-6) below.

(8-6) Similar Word Extraction Processing The similar word extraction unit 802 extracts a word having a certain degree of similarity with the keyword as a similar word that is acoustically similar to the keyword. Next, the influence degree calculation means 805 performs the influence degree calculation process (8-7) below.

(8-7) Influence degree calculation process The influence degree calculating means 805 calculates the influence degree of the language model 807 with respect to the recognition rate of similar words that accompanies increasing the recognition rate of the keyword in the speech recognition apparatus. Next, the language score adjustment unit 804 performs the language score adjustment process of (8-8) below.

(8-8) Language Score Adjustment Processing In this embodiment, the language score adjustment unit 804 has a smaller bonus that is given to the language score of the keyword and a penalty that is given to the language score of the similar word as the degree of influence is larger. Adjust to Specifically, the penalty degree (Pi) given to the language score of the similar word and / or the bonus value (B) given to the language score of the keyword are calculated using the degree of influence. Next, information indicating the penalty value (Pi) and the bonus value (B) is output to the language score updating means 806.

(8-9) Language model update processing In response to the information, the language model update device 806 specifies the word for the keyword and / or the similar word in the language model 807. The identification information is assigned. Further, for the keyword, the bonus value (B) is recorded in the area (A). For the similar words, the penalty value (Pi) is recorded in the area (A). Furthermore, the language model update device 806 performs either or both of the following processes (1) and (2) when the speech recognition unit 809 searches for the language model 807. Thereby, the connection probability of the keyword is increased, and the accuracy with which the keyword is recognized can be increased.
(1) The penalty value recorded in the area (A) is assigned to the language score of the similar word.
(2) The bonus value recorded in the area (A) is assigned to the language score 807 of the keyword.

  According to this embodiment, the language score adjusted according to the influence degree with respect to a similar word can be provided with respect to the language model 807 by the said language score adjustment means 804. FIG. Therefore, the recognition rate of keywords can be increased, and for example, when the similar words are important words or the words with low recognition rates, the recognition rate of the similar words may decrease too much. Absent. According to the present embodiment, such similar words can be recognized appropriately, and the overall recognition rate can be improved.

  In this embodiment, the voice recognition means 809 may include means for performing (1) and (2). In this case, when the speech recognition unit 809 performs speech recognition processing, the identification information identifies the keywords and similar words, and performs either (1) or (2) or any one of the processing. be able to. Therefore, in this case, the language model update device 806 may not have a function for the processes (1) and (2).

The speech recognition method of the present embodiment is a speech recognition method further including the following steps (e) to (f) in the steps (a) to (d) in the fourth embodiment.
(E) The feature quantity extraction unit 810 analyzes the voice input by the voice input unit 811 and extracts an acoustic feature quantity.
(F) The speech recognition unit 809 refers to the acoustic model 812 and the language model 807, comprehensively evaluates the acoustic score and the language score, and outputs a word string that can be determined to be most likely as a recognition result. Step.

100, 200, 300, 400, 600, 700, 800 Language model update apparatus 101, 201, 301, 401, 601, 701, 801 Keyword storage means 102, 202, 302, 402, 602, 702, 802 Similar word extraction means 104, 204, 304, 404, 604, 704, 804 Language score adjustment means 105, 205, 305, 405, 605, 705, 805 Similarity calculation means 106, 206, 306, 406, 606, 706, 806 Language score update Means 107, 207, 307, 407, 607, 707, 807 Language model 108, 208, 308, 408, 608, 708, 808 Recognition dictionary 109, 209, 309, 409, 609, 709, 809 Speech recognition means 210, 810 Feature amount extraction means 211, 811 Voice input means 303 (or 803) the similarity measurement unit 410 Occurrences storage unit 610 recognition rate storage unit 710 similar word presentation means 711 impact input means 813 acoustic model

Claims (22)

Including keyword storage means, influence calculation means, language score adjustment means, and language score update means,
The keyword storage means stores a keyword,
The influence calculating means calculates an influence on a similar word by increasing a recognition rate of the keyword;
The language score adjusting means adjusts at least one of a bonus given to the language score of the keyword and a penalty given to the language score of the similar word based on the degree of influence,
The language score update means performs at least one of the provision of the bonus adjusted by the language score adjustment means for the language score of the keyword and the provision of the penalty adjusted by the language score adjustment means for the language score of the similar word. Thus, a language model update device for a speech recognition device, wherein the language score of the language model is updated.   2. The language model update device for a speech recognition device according to claim 1, wherein the language score adjusting means decreases the size of the bonus or the penalty in accordance with the increase in the degree of influence.   The influence degree calculating means calculates the influence degree based on at least one selected from the group consisting of an appearance frequency of the similar words, a recognition rate of the similar words, and judgment information by a user. 3. The language model update device for a speech recognition device according to 1 or 2.   The language model update device for a speech recognition apparatus according to any one of claims 1 to 3, wherein the influence degree calculating unit increases the influence degree according to an increase in appearance frequency of the similar words. The language model update device for a speech recognition device according to any one of claims 1 to 4, wherein the language score adjustment unit calculates the penalty (Pi) according to the following formula (I).

Pi = S (sim (w, wi)) − αlogC (wi) (I)

In the formula (I), w is a keyword, wi is a similar word, sim (a, b) is a similarity between the words a and b, and S (sim) is a function for determining a penalty Pi from the similarity. C (w) is the number of appearances of the word w in the recognition history, and α is a constant coefficient.   The language model update device for a speech recognition apparatus according to any one of claims 1 to 5, wherein the influence degree calculation means increases the influence degree in accordance with a decrease in the recognition rate of the similar words. The language model updating apparatus for a speech recognition apparatus according to any one of claims 1 to 6, wherein the language score adjusting unit calculates a penalty (Pi) according to the following formula (II).

Pi = S (sim (w, wi)) × P (wi) (II)

In the formula (II), P (w) is a recognition rate of the word w. The language model update device for a speech recognition device according to any one of claims 1 to 7, wherein the language score adjusting unit calculates the penalty (Pi) according to the following formula (III).

Pi = S (sim (w, wi)) × T (ui) (III)

In the above formula (III), ui is the influence degree input by the user, and T is the penalty reduction degree. The language model update device for a speech recognition device according to any one of claims 1 to 8, wherein the language score adjusting unit calculates the bonus (B) according to the following formula (IV).

B = b−β [Pi] (IV)

In the formula (IV), b is a constant bonus, β is a coefficient, and [Pi] is a maximum value of the penalty Pi of the similar word or an average value of the penalty of the similar word. The language model update device for a speech recognition device according to any one of claims 1 to 9,
Voice input means for inputting voice;
Feature quantity extraction means for extracting the feature quantity of the input speech;
Language model,
Voice recognition means,
Updating the language model in the language model updating device for the speech recognition device;
The speech recognition apparatus, wherein the speech recognition unit recognizes the speech using the updated language model from the feature amount extracted by the feature amount extraction unit. Using keyword storage means, influence calculation means, language score adjustment means, and language score update means,
A keyword storage step in which the keyword storage means stores a keyword;
An influence degree calculating step in which the influence degree calculating means calculates an influence degree on a similar word by increasing a recognition rate of the keyword;
A language score adjusting step in which the language score adjusting means adjusts at least one of a bonus given to the language score of the keyword and a penalty given to the language score of the similar word based on the degree of influence;
The language score update means performs at least one of the provision of the bonus adjusted by the language score adjustment means for the language score of the keyword and the provision of the penalty adjusted by the language score adjustment means for the language score of the similar word. A language score update method for a speech recognition apparatus, comprising: a language score update step of updating a language score of a language model.   12. The method for updating a language model for a speech recognition apparatus according to claim 11, wherein, in the language score adjustment step, the language score adjustment means decreases the bonus or the penalty according to the increase in the influence degree. .   In the influence degree calculating step, the influence degree calculating means calculates the influence degree based on at least one selected from the group consisting of the appearance frequency of the similar words, the recognition rate of the similar words, and judgment information by the user. The method for updating a language model for a speech recognition apparatus according to claim 11 or 12.   The speech recognition apparatus according to any one of claims 11 to 13, wherein, in the influence degree calculating step, the influence degree calculating unit increases the influence degree in accordance with an increase in appearance frequency of the similar words. To update the language model. The speech recognition device according to any one of claims 11 to 14, wherein, in the language score adjustment step, the language score adjustment unit calculates the penalty (Pi) according to the following formula (I). How to update the language model.

Pi = S (sim (w, wi)) − αlogC (wi) (I)

In the formula (I), w is a keyword, wi is a similar word, sim (a, b) is a similarity between the words a and b, and S (sim) is a function for determining a penalty Pi from the similarity. C (w) is the number of appearances of the word w in the recognition history, and α is a constant coefficient.   The speech recognition according to any one of claims 11 to 15, wherein, in the influence degree calculating step, the influence degree calculating means increases the influence degree as the recognition rate of the similar word decreases. How to update the language model for devices. 17. The speech recognition apparatus according to claim 11, wherein in the language score adjustment step, the language score adjustment unit calculates the penalty (Pi) according to the following formula (II): How to update the language model.

Pi = S (sim (w, wi)) × P (wi) (II)

In the formula (II), P (w) is a recognition rate of the word w. The speech recognition device according to any one of claims 11 to 17, wherein, in the language score adjustment step, the language score adjustment unit calculates the penalty (Pi) according to the following formula (III). How to update the language model.

Pi = S (sim (w, wi)) × T (ui) (III)

In the above formula (III), ui is the influence degree input by the user, and T is the penalty reduction degree. The speech recognition device according to any one of claims 11 to 18, wherein, in the language score adjustment step, the language score adjustment unit calculates the bonus (B) according to the following formula (IV). How to update the language model.

B = b−β [Pi] (IV)

In the formula (IV), b is a constant bonus, β is a coefficient, and [Pi] is a maximum value of the penalty Pi of the similar word or an average value of the penalty of the similar word. Using the speech recognition device according to claim 10,
Updating the language model by the speech recognition language model update method according to any one of claims 11 to 19;
A feature amount extraction step in which the feature amount extraction means extracts the feature amount of the voice input by the voice input means;
A speech recognition method, wherein the speech recognition means includes a speech recognition step of recognizing the speech using a language model from the feature quantity extracted by the feature quantity extraction means.   A computer program capable of executing the method according to any one of claims 11 to 20 on a computer.   A recording medium storing the computer program according to claim 21.

Images