JP6179509B2 - Language model generation apparatus, speech recognition apparatus, language model generation method, and program storage medium - Google Patents

Description

  The present invention relates to a technique for generating a language model used for speech recognition processing.

  A language model is used in the process in which the computer recognizes speech. A language model is a model that defines linguistic constraints. There is a technique for optimizing the language model according to the recognition target.

  For example, in Patent Document 1, the speech recognition apparatus uses, as a language model related to conversation, a database in which a vocabulary set for questions, a vocabulary set for responses, and probabilities associated with them are associated. In Patent Document 2, the speech recognition apparatus uses a result of recognizing a voice (utterance) uttered by one speaker as a recognition target of a conversation between two parties, and uses a language model suitable for the utterance of the other speaker. Generate.

JP 2004-012713 A JP 2011-107314 A

  By the way, in the method described in Patent Document 1, there is a possibility that data as a language model becomes enormous. In other words, in a conversation between a customer service representative and a customer in a contact center or a conversation in a meeting, the question and the response may be repeated multiple times. When such a series of conversations (a plurality of utterances) is to be recognized, the combination of vocabulary candidates (that is, data as a language model) included in the question and the response to the question becomes enormous.

  In the method described in Patent Document 2, only the result of recognizing the speech uttered from one of the talking speakers is used in the process of generating the language model. Language models for words not included in the results are not generated. For this reason, the speech recognition apparatus described in Patent Document 2 cannot generate a language model suitable for the speech of the other speaker.

  The present invention has been made to solve the above problems. That is, the main object of the present invention is when a word that is not included in the utterance of one of the talking speakers is included in the utterance of another speaker responding to the utterance. Is to provide a technique for improving the recognition accuracy of the speech recognition process for recognizing the speech of the other speaker.

In order to achieve the above object, a language model generation apparatus according to the present invention includes:
A first data that is a word or a word string included in the first utterance and a related word related to the first data of the word or the word string included in the second utterance responding to the first utterance. The first episode of the first speaker and other speakers who are exchanging a recognition target conversation with the first data in a group of related words associated with certain second data. Extracting means for collating a word or a word string included in the person's utterance and extracting the second data associated with the first data corresponding to the word or word string;
Model generation means for generating a language model according to the speech of the other speaker using the extracted second data.

The speech recognition apparatus according to the present invention is
Recognition means for executing speech recognition processing for converting speech into text data;
A language model generation device that generates a language model used for the speech recognition processing,
The language model generation device is the language model generation device of the present invention, and is included in the utterance of the first speaker among the first speaker and the other speaker who are exchanging the conversation to be recognized. Based on the word or word string converted into text data by the means, a language model corresponding to the other speaker is generated.

Furthermore, the language model generation method according to the present invention includes:
A first data that is a word or a word string included in the first utterance and a related word related to the first data of the word or the word string included in the second utterance responding to the first utterance. The first episode of the first speaker and other speakers who are exchanging a recognition target conversation with the first data in a group of related words associated with certain second data. Collating a word or word string included in the person's utterance, and extracting the second data associated with the first data corresponding to the word or word string,
A language model corresponding to the speech of the other speaker is generated using the extracted second data.

Furthermore, the program storage medium according to the present invention is:
A first data that is a word or a word string included in the first utterance and a related word related to the first data of the word or the word string included in the second utterance responding to the first utterance. The first episode of the first speaker and other speakers who are exchanging a recognition target conversation with the first data in a group of related words associated with certain second data. A process of collating a word or a word string included in the person's utterance and extracting the second data associated with the first data corresponding to the word or the word string;
A computer program for causing a computer to execute processing for generating a language model according to the speech of the other speaker using the extracted second data is stored.

  The main object of the present invention is also achieved by a language model generation method corresponding to the language model generation apparatus of the present invention. The main object of the present invention is also achieved by a computer program for realizing the language model generation apparatus, the speech recognition apparatus and the language model generation method of the present invention by a computer, and a program storage medium for storing the computer program. Achieved.

  According to the present invention, even when a word that is not included in the utterance of one of the talking speakers is included in the utterance of another speaker who is responding to the utterance, The recognition accuracy of the process for recognizing the utterances of other speakers can be increased.

It is a block diagram which simplifies and represents the hardware constitutions of the speech recognition apparatus of 1st Embodiment which concerns on this invention. It is a block diagram which simplifies and represents the structure of the language model production | generation apparatus of 1st Embodiment. It is a figure showing the example of a related word pair with an image. It is a flowchart showing the operation example in which the language model production | generation apparatus of 1st Embodiment produces | generates a language model. It is a figure showing the example of the attached information linked | related with a related word pair with an image. It is a block diagram which simplifies and represents the structure of the speech recognition apparatus of 2nd Embodiment which concerns on this invention. It is a flowchart showing the operation example in which the speech recognition apparatus of 2nd Embodiment recognizes a speech. It is a block diagram which simplifies and represents the structure of the speech recognition apparatus of 3rd Embodiment which concerns on this invention. It is a flowchart showing the operation example in which the language model production | generation apparatus in 3rd Embodiment produces | generates a related word pair.

  Embodiments according to the present invention will be described below with reference to the drawings.

<First Embodiment>
FIG. 1 is a block diagram showing a simplified hardware configuration of a speech recognition apparatus 1 including a language model generation apparatus according to the first embodiment of the present invention. The voice recognition device 1 includes a CPU (Central Processing Unit) 10, a memory 11 as a storage device, an HDD (Hard Disk Drive) 12 as a storage device, a communication IF (InterFace) 13, an input device 14, and a voice. And an input device 15. The input device 14 is a device that inputs information to the voice recognition device 1. Examples of the input device 14 include a keyboard and a pointing device such as a mouse. The voice input device 15 is a device that takes a voice into the voice recognition device 1 by converting the voice into an electrical signal (voice signal). The voice input device 15 includes, for example, a microphone. The components 10-15 of the speech recognition apparatus 1 are connected to each other through a bus 16 and input / output data.

  FIG. 2 is a block diagram illustrating a configuration example of the language model generation device 18 included in the speech recognition device 1 by a solid line. The language model generation device 18 includes an extraction unit (extraction unit) 100 and a model generation unit (model generation unit) 101. The language model generation device 18 is realized by the CPU 10 executing a computer program (hereinafter also referred to as a program) 19 stored in the HDD 12 (or the memory 11). In other words, it can be said that the language model generation device 18 in the first embodiment is configured by the computer program 19 or the computer-readable storage medium 12 (11) in which the computer program is stored. Note that all or some of the functions of the language model generation device 18 may be realized by hardware provided in the speech recognition device 1.

  In the first embodiment, the storage device (HDD 12) stores a data group 20 in which related word pairs are collected. FIG. 3 is a diagram illustrating an example of the data group 20 in which related word pairs are collected. The related word pair is data related to the conversation, and is data in which the first data based on the first utterance and the second data based on the second utterance responding to the first utterance are associated with each other. is there. The first data is a word or a word string included in the first utterance. The second data is a related word (word or word string) related to the first data among the words or word strings included in the second utterance responding to the first utterance. Specifically, taking a conversation between an operator and a customer as an example, the related word pair includes an operator's utterance (first utterance) and a customer's utterance responding to the operator's utterance (second utterance). It is a combination with a related word that is a word or a word string that co-occurs in.

  More specifically, for example, it is assumed that the customer responds to the utterance (question) of the operator “Can you give me your name?” Such as “I am Suzuki” or “I am Yamada”. Based on such a conversation, “Suzuki” or “Yamada” included in the customer's utterance (second utterance) with respect to the word “name” included in the operator's utterance (first utterance). Data that related words (words) co-occur (high probability of being uttered) can be obtained. In this case, a related word pair in which “name” as the first data and “Suzuki” as the second data are associated, or “name” as the first data and the second data A related word pair associated with “Yamada” is generated.

  Further, it is assumed that the following conversation is exchanged between the operator and the customer.

Operator: “Ctrl”
Customer: “Ctrl, yes”
Operator: “Alt”
Customer: “Alt, yes”
Operator: “Delete”
Customer: “Delete, yes”
Operator: “Press the three buttons simultaneously”
Customer: “Yes”
Operator: “You can then restart”
Customer: “Restart”
Based on such a conversation, for example, a customer's utterance (second utterance) for a word string (a combination of words) “Ctrl, Alt, Delete” included in the operator's utterance (first utterance) Can be obtained that the related word “restart” included in is co-occurring. In this case, a related word pair in which the word string “Ctrl, Alt, Delete” as the first data and the related word (word) “restart” as the second data are generated is generated. . Also, based on the conversation, a related word pair is generated in which the word “restart” as the first data and the related word (word) “restart” as the second data are associated. . Furthermore, based on the above conversation, a word string “Ctrl, Alt, Delete” as the first data, a word string “simultaneously” and “press”, and a related word (word) “restart” as the second data, A related word pair associated with is generated.

  In this example, the operator's utterance is the first utterance, the customer's utterance is the second utterance, the word or word string based on the operator's utterance is the first data, and the word or The word string is the second data. In contrast, the customer utterance is the first utterance, the operator utterance is the second utterance, the word or word string based on the customer utterance is the first data, and the word or word string based on the operator utterance is used. May be the second data.

  The extraction unit 100 extracts the second data from the data group 20 based on a word or a word string included in the utterance of the first speaker among the first speaker and the other speakers who are exchanging the recognition target conversation. It has a function to extract. That is, the extraction unit 100 collates the word or word string included in the utterance of the first speaker with the first data in the data group 20, and if there is first data corresponding to the word or word string, Second data associated with the first data is extracted. When there are a plurality of related word pairs that include the same first data corresponding to the collated word or word string and the second data is different from each other, the extraction unit 100 uses the second data of the related word pairs. To extract. Specifically, in the case where the word “name” is collated with the data group 20 shown in FIG. 3, the extraction unit 100 sets the related word pair “name” and “Suzuki” and “name” Based on the related word pair with “Yamada”, “Suzuki” and “Yamada” are extracted.

  The model generation unit 101 has a function of generating a language model corresponding to the utterance of the other speaker using the second data extracted by the extraction unit 100. Language models include various language model types such as an N-gram language model, a trigger language model, and a hierarchical Bayesian language model. Here, the model generation unit 101 generates a language model selected in consideration of the situation in which speech recognition is performed. Here, description of the method of generating the language model is omitted.

  Next, an operation example of the language model generation apparatus 18 will be described with reference to the flowchart of FIG. FIG. 4 is a flowchart showing an operation example of the language model generation device 18. That is, the flowchart shown in FIG. 4 represents a processing procedure described in a program executed by the language model generation device 18.

  For example, when the language model generation device 18 receives a word or a word string included in the utterance of the first speaker (step S101 in FIG. 4), the extraction unit 100 converts the received word or word string into the data group 20 To the first data. If there is first data corresponding to (corresponding to) the word or word string, the extraction unit 100 extracts second data associated with the first data (step S102).

  After that, using the extracted second data, the model generation unit 101 generates a language model according to the utterances of other speakers who are talking to the first speaker (step S103).

  In the first embodiment, as described above, the language model generation device 18 (speech recognition device 1) is based on the utterances of the first speaker, the words assumed to co-occur with the utterances of other speakers, or A word string (that is, second data) is extracted from the data group 20. And the language model production | generation apparatus 18 (voice recognition apparatus 1) produces | generates the language model according to the utterance of another speaker based on the extracted 2nd data. For this reason, the language model generation device 18 (speech recognition device 1) according to the first embodiment, for example, is a speaker other than the case based on vocabulary candidates included in questions that assume various situations and responses to the questions. The amount of data related to the language model corresponding to the utterance of

  In addition, the language model generation device 18 (voice recognition device 1) of the first exemplary embodiment includes words included in the utterance of the first speaker in the utterances of other speakers that respond to the utterance of the first speaker. Or even when the word string is not included, the voice recognition accuracy of the speech of another speaker can be improved. In particular, it is possible to improve voice recognition accuracy in a series of conversations in which questions and responses are repeated a plurality of times, such as conversations exchanged at a contact center or a conference.

Second Embodiment
Next, a second embodiment according to the present invention will be described. In the description of the second embodiment, the same reference numerals are given to the same name parts as the constituent parts constituting the speech recognition apparatus of the first embodiment, and the duplicate description of the common parts is omitted.

  FIG. 6 is a block diagram illustrating a simplified configuration of the speech recognition apparatus according to the second embodiment. The speech recognition apparatus 1 according to the second embodiment has a speech recognition unit (recognition means) 103. The speech recognition unit 103 has a function of executing speech recognition processing for converting speech into text data. That is, when an electrical signal corresponding to speech is input from the speech input device (microphone) 15, the speech recognition unit 103 converts the electrical signal (speech) into text data. In the second embodiment, among the speakers who are exchanging a conversation to be recognized, a speaker who is assumed to have high recognition accuracy by the speech recognition processing is defined as the first speaker, and the information is stored in the speech recognition apparatus 1. Given. As a result, the speech recognition unit 103 is generated by the model generation unit 101 and stored in the storage device (HDD) 12 when executing speech recognition processing for the utterance of another speaker who is not the first speaker. Use language models.

  Further, the voice recognition unit 103 may output a plurality of conversion candidates (text data) by voice recognition processing. Furthermore, the speech recognition unit 103 associates at least one of conversion reliability, a word part of speech, an acoustic score, a language score, and an N-gram hit rate as probability information with text data obtained by speech recognition processing. Furthermore, when the voice (electrical signal) received from the voice input device 15 is the voice of the first speaker, the voice recognition unit 103 is the voice of the first speaker in the text data obtained by the voice recognition process. Associate information that represents

  When the text data from the speech recognition unit 103 is data based on the utterance of the first speaker, the extraction unit 100 is based on the word or word string included in the text data, as in the first embodiment. The second data is extracted from the data group 20.

  In addition, when the speech recognition unit 103 outputs a plurality of conversion candidates (text data), the extraction unit 100 uses probability information associated with the text data and uses one of the plurality of conversion candidates. The function to select one is provided. For example, the extraction unit 100 selects a conversion candidate that is considered to be most appropriate based on the probability information. Then, the extraction unit 100 extracts the second data from the data group 20 based on the word or word string included in the selected conversion candidate text data, as in the first embodiment.

  Other configurations of the speech recognition device 1 (language model generation device) of the second embodiment are the same as those of the first embodiment.

  Next, an operation example of the speech recognition apparatus 1 in the second embodiment will be described based on the flowchart of FIG. FIG. 7 is a flowchart showing an operation example of the speech recognition apparatus 1 in the second embodiment. That is, the flowchart of FIG. 7 represents a processing procedure described in a program executed by the CPU 10 of the speech recognition apparatus 1.

  In the speech recognition device 1, when speech (electrical signal) is applied from the speech input device 15 to the speech recognition unit 103, the speech recognition unit 103 executes speech recognition processing for converting the speech into text data (FIG. 7). Step S201). In the case where the conversion source voice is the voice of the first speaker's utterance, the extraction unit 100 starts from the data group 20 based on the word or word string included in the text data by the voice recognition process. Two data are extracted (step S202). The model generation unit 101 uses the extracted second data to generate a language model corresponding to another speaker who is talking to the first speaker (step S203). The generated language model is stored in the storage device (HDD 12 or memory 11).

  Thereafter, when a voice (electrical signal) generated by another speaker who responds to the first speaker's utterance is added to the voice recognition unit 103, the voice recognition unit 103 converts the voice into text data. Recognition processing is executed (step S204). At this time, the speech recognition unit 103 uses a language model corresponding to another speaker generated by the model generation unit 101 and stored in the storage device 12 (11).

  Since the speech recognition device 1 (language model generation device 18) of the second embodiment includes the extraction unit 100 and the model generation unit 101 similar to those of the first embodiment, the same effects as those of the first embodiment are obtained. be able to.

<Third Embodiment>
Next, a third embodiment according to the present invention will be described. In the description of the third embodiment, the same reference numerals are given to the same name parts as the constituent parts constituting the speech recognition apparatuses of the first and second embodiments, and the duplicate description of the common parts is omitted. .

  FIG. 8 is a simplified block diagram illustrating a configuration example of the speech recognition apparatus 1 according to the third embodiment. The language model generation device 18 constituting the speech recognition device 1 of the third embodiment includes a data generation unit (data generation means) 104 in addition to the configuration of the second embodiment. A conversation corpus 22 is stored in the storage device (HDD 12). The conversation corpus 22 is a data group obtained by collecting data that associates text data of a conversation as a sample and information related to the conversation. Specifically, the information related to the conversation is information such as the voice feature amount obtained from the voice of the speaker who is having a conversation, the feature of the speaker who spoke, and the time of speaking. For example, the speech feature amount includes a fundamental frequency of speech speech or a feature vector used for speech recognition. Further, as a feature of the speaker who spoke, there is the name or role of the speaker. Examples of roles include the chairman at a conference, the operator and customer at a call center. Further, the utterance time includes the absolute time from the conversation start time to the start of the utterance or the duration of the utterance. In addition to the above information, the conversation corpus 22 may include information such as emotions such as the speaker being angry, utterance style such as the utterance being a question sentence or a statement sentence, etc. as conversation information. .

  The data generation unit 104 is one of functional units of the CPU 10 that is realized by the CPU 10 operating according to the program 19. The data generation unit 104 uses, as first data, a word or a word string included in text data (utterance of the first speaker) by the voice recognition processing of the voice recognition unit 103. Further, the data generation unit 104 has a function of extracting a word or a word string related to the first data from the conversation corpus 22 as second data. Furthermore, the data generation unit 104 has a function of generating a related word pair that associates the first data and the second data, and adding the related word pair to the data group 20.

  Specifically, the data generation unit 104 extracts from the conversation corpus 22 the text data of the conversation including the word or the word string as the first data as described above. Then, the data generation unit 104 specifies the word or word string related to the first data as the second data using the extracted text data of the conversation. Further, the data generation unit 104 generates a related word pair by associating the first data and the second data, and adds the related word pair to the data group 20. Such a data generation unit 104 not only functions when generating the data group 20 in advance, but also has a function of learning (machine learning) the data group 20 while actually executing speech recognition processing for conversation. Can have.

  Next, an operation example in which the data generation unit 104 (CPU 10) generates related word pairs will be described with reference to the flowchart of FIG. FIG. 9 is a flowchart illustrating an operation example in which the speech recognition apparatus 1 according to the third embodiment generates a related word pair. In other words, the flowchart of FIG. 9 represents a processing procedure in which the speech recognition apparatus 1 generates a related word pair.

  When the data generation unit 104 sets a word or a word string included in text data (utterance of the first speaker) by the speech recognition processing of the speech recognition unit 103 as the first data (step S301 in FIG. 9), The first data is collated with the conversation corpus 22. Thereby, the data generation unit 104 extracts the conversation information including the first data from the conversation corpus 22. Based on the extracted conversation information, the data generation unit 104 extracts, as second data, a word or word string that is included in the utterance that responds to the utterance that includes the first data and that is related to the first data. (Step S302). Furthermore, the data generation unit 104 generates a related word pair by associating the first data and the second data (step S303). Then, the data generation unit 104 adds the related word pair to the data group 20 of the storage device 12 (step S304).

  Also in the third embodiment, since the language model generation device 18 similar to that in the first embodiment and the second embodiment is provided, the same effect as in the first embodiment and the second embodiment can be obtained. The data generation unit 104 in the third embodiment uses a conversation corpus when generating related word pairs. Thereby, the data generation part 104 can generate | occur | produce the related word pair according to conversation.

<Other embodiments>
In addition, this invention is not limited to the 1st-3rd embodiment, Various embodiment can be taken. For example, not only the related word pair but also information related to the related word pair may be associated with the data group 20 as attached information as shown in FIG. The attached information includes, for example, the notation of the word or word string as the first data or the second data, the part of speech, the role of the utterer, the time from the start of the conversation to the utterance, the number of utterances and the number of words. One or more of the attached information may be associated with the related word pair of the data group 20.

  When the attached information is associated with the related word pair in this way, the extraction unit 100 accepts not only the first data based on the utterance of the first speaker but also information corresponding to the attached information as reference information. . For example, the extraction unit 100 receives time information related to the conversation as reference information from a time measuring device provided in the voice recognition device 1. In addition, the extraction unit 100 may receive speaker role information or the like input by the input device 14 as reference information.

  The extraction unit 100 extracts second data from the data group 20 based on the first data and the reference information as described above.

  For example, in the case of a conversation in a company contact center, the operator's utterance "Can you name me" is often uttered just before the beginning or end of the conversation in the contact center? It is assumed that such information is associated with related word pairs in the data group 20 as attached information. In this case, when the extraction unit 100 receives an utterance “Can you name me” as the first data, the information (reference information) of the utterance time (the time from the start of the conversation until the utterance is spoken) ), More appropriate second data can be extracted from the first data.

  Further, in a conversation at a contact center, there is a word or a word string that an operator speaks but a customer does not speak. When such a word or word string is added to the extraction unit 100 as the first data, for example, the extraction unit 100 also uses information (reference information) on the role of the speaker from the data group 20. By extracting the second data, it is possible to extract the second data that is more closely related to the first data.

  In the first to third embodiments, the extraction unit 100 extracts second data related to the first data. On the other hand, the extraction unit 100 may extract not only the second data but also related word pairs in which the first data and the second data are associated. In this case, the model generation unit 101 generates a language model using the extracted related word pair.

  While the present invention has been described with reference to the embodiments, the present invention is not limited to the above embodiments. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

  In addition, this application claims the priority on the basis of Japanese application Japanese Patent Application No. 2012-113534 for which it applied on May 17, 2012, and takes in those the indications of all here.

A part or all of the above-described embodiment can be described as in the following supplementary notes, but is not limited thereto.
(Appendix 1)
The word included in the conversation data based on the related word pair that associates the word string included in the utterance of the first speaker and the related word included in the utterance of the second speaker responding to the utterance. A related term extractor for extracting related terms associated with the column;
A language model generation apparatus, comprising: a language model generation unit that generates a language model of the second speaker using the related words extracted by the related word extraction unit.
(Appendix 2)
A speech recognition unit for performing speech recognition on the speech of the first speaker;
The language model generation device according to supplementary note 1, wherein the related word extraction unit extracts a related word associated with the word string included in a result of voice recognition by the voice recognition unit based on the related word pair. .
(Appendix 3)
The language model according to supplementary note 2, wherein the related word extraction unit treats, as a first speaker, a speaker whose speech is recognized by the speech recognition unit with higher accuracy than others among speakers included in a conversation. Generator.
(Appendix 4)
A word string included in the speech recognition result of the first speaker, information converted into text data based on speech in a specific conversation, speech features obtained from the speech or the specific conversation, and Based on a speech corpus or a speech corpus associated with speaker information obtained from the specific conversation, the speech recognition result of the first speaker and the conversation corpus highly related to the first speech recognition result The language model generation device according to Supplementary Note 2 or Supplementary Note 3, further comprising a related word pair generation unit that generates a combination with the text included in the text as the related word pair.
(Appendix 5)
The related word extraction unit associates one or more word strings included in the utterance of the first speaker with one or more word strings included in the utterance of the second speaker responding to the utterance. 5. The language model generation device according to any one of supplementary notes 1 to 4, which is extracted as a word.
(Appendix 6)
The related word extraction unit includes at least one of a notation of a word string included in the utterance text of the first speaker, a part of speech, a role of the speaker, a time from the conversation start time to the utterance, the number of utterances, and the number of words. The language model generation device according to any one of Supplementary Note 1 to Supplementary Note 5, wherein a related word is extracted from the related word pair.
(Appendix 7)
The related word extraction unit uses any one or more of the reliability, the part of speech, the acoustic score, the language score, and the N-gram hit rate included in the speech recognition result of the first speaker. The language model generation device according to any one of supplementary notes 2 to 4, which selects a person's speech recognition result.
(Appendix 8)
The language model according to claim 1, further comprising a speech recognition unit that performs speech recognition on the utterance of the second speaker based on the language model of the second speaker generated by the language model generation unit. Generator.
(Appendix 9)
The word included in the conversation data based on the related word pair that associates the word string included in the utterance of the first speaker and the related word included in the utterance of the second speaker responding to the utterance. Extract related terms associated with a column,
A language model generation method for generating a language model of the second speaker by using an extracted related word.
(Appendix 10)
The word included in the conversation data based on the related word pair that associates the word string included in the utterance of the first speaker and the related word included in the utterance of the second speaker responding to the utterance. Extracting related terms associated with the column;
A computer program that causes a computer to execute processing for generating a language model of the second speaker using the extracted related terms.

  The present invention is effective in the field of using technology related to processing for recognizing speech during conversation.

1 Voice recognition device 10 CPU
18 language model generation device 100 extraction unit 101 model generation unit 103 speech recognition unit 104 data generation unit

Claims (10)

A first data that is a word or a word string included in the first utterance and a related word related to the first data of the word or the word string included in the second utterance responding to the first utterance. The first episode of the first speaker and other speakers who are exchanging a recognition target conversation with the first data in a group of related words associated with certain second data. Extracting means for collating a word or a word string included in the person's utterance and extracting the second data associated with the first data corresponding to the word or word string;
A language model generation device comprising model generation means for generating a language model according to the speech of the other speaker using the extracted second data.   The extraction means includes the same first data corresponding to the collated word or word string, and detects that the plurality of related word pairs having different second data are in the data group. The language model generation apparatus according to claim 1, wherein the second data of the plurality of related word pairs is extracted. The related word pair of the data group includes the first data notation, the part of speech of the first data, the role of the speaker that emits the first data, and the first data after the conversation is started. At least one of the time to issue and the number of words of the first data within a predetermined period is associated as attached information,
The extraction means receives not only a word or word string included in the utterance of the first speaker but also reference information corresponding to the attached information, and extracts the second data from the data group. The language model generation apparatus according to claim 1, wherein the second data is extracted also using reference information. The extraction means receives a conversion candidate of a word or a word string by a speech recognition process for converting speech into text data, and the conversion candidate includes reliability, part of speech, acoustic score, language obtained by the speech recognition process At least one of the score and the N-gram hit rate is associated as probability information,
When the extraction means receives a plurality of the conversion candidates by the voice recognition processing for the utterance of the first speaker, the extraction means uses a plurality of the conversions by using probability information associated with the conversion candidates. 3. The second data is extracted by selecting one of candidates and comparing the selected word or word string of the selected conversion candidate with first data of the data group. Or the language model production | generation apparatus of Claim 3.   The extraction means extracts not only the second data but also the associated first data, and the model generation means associates the extracted first data with the second data. The language model generation device according to any one of claims 1 to 4, wherein a language model corresponding to an utterance of the other speaker is generated using the related word pair. Comprising data generation means for generating the data group;
The data generation means includes a word or word included in a conversation corpus in which words or word strings taken from the first speaker's utterance and used as the first data in the data group and text data representing conversation as a sample are collected. By associating the word or the word string taken as the second data in the data group with high relevance to the word or the word string as the first data in the sequence, the related word pair The language model generation device according to claim 1, wherein the language model generation device generates the data. Recognition means for executing speech recognition processing for converting speech into text data;
A language model generation device that generates a language model used for the speech recognition processing,
The language model generation device is the language model generation device according to any one of claims 1 to 6, wherein the language model generation device includes a first speaker and another speaker who are in a conversation to be recognized. A speech recognition apparatus for generating a language model corresponding to the other speaker based on a word or a word string included in the utterance of the first speaker and converted into text data by the recognition means.   The said recognition means is provided with the function to perform the speech recognition process which converts the speech of said another speaker into text data using the said language model produced | generated by the said language model production | generation apparatus. Voice recognition device. A first data that is a word or a word string included in the first utterance and a related word related to the first data of the word or the word string included in the second utterance responding to the first utterance. The first episode of the first speaker and other speakers who are exchanging a recognition target conversation with the first data in a group of related words associated with certain second data. Collating a word or word string included in the person's utterance, and extracting the second data associated with the first data corresponding to the word or word string,
A language model generation method in which a j computer generates a language model corresponding to the utterance of the other speaker using the extracted second data. A first data that is a word or a word string included in the first utterance and a related word related to the first data of the word or the word string included in the second utterance responding to the first utterance. The first episode of the first speaker and other speakers who are exchanging a recognition target conversation with the first data in a group of related words associated with certain second data. A process of collating a word or a word string included in the person's utterance and extracting the second data associated with the first data corresponding to the word or the word string;
A program storage medium storing a computer program that causes a computer to execute a process of generating a language model according to the speech of the other speaker using the extracted second data.

Images