JP2005208483A - Device and program for speech recognition, and method and device for language model generation - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a device and program for speech recognition whose speech recognition rate is improved by optimizing candidates of speech recognition, and a method and device for language model generation used for the speech recognition device. <P>SOLUTION: The speech recognition device 1 is equipped with a language model storage part 21 which stores a language model generated from a fixed-form pattern of utterance, a sound model storage part 22 which stores a sound model including sound characteristics of a speech, and a speech processing part 31 which takes a sound analysis of a speech signal by reference to the language model and sound model and converts it into character information. The language model is generated from the fixed-form pattern of utterance including character information obtained at the site specified by a URL of the organization that the speaker belongs to. Further, the sound model is a model learnt with a telephone speech. Such a language model and a sound model are employed, so a proper candidate can be converted into character information to improve the speech recognition rate. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a speech recognition device that recognizes speech and converts it into character information, a speech recognition program, a language model generation method that generates a language model used in the speech recognition device, and a language model generation device.

  In recent years, a speech recognition technique for converting speech into characters using a computer has begun to be used in various systems and scenes such as car navigation. In a speech recognition apparatus using speech recognition technology, large vocabulary continuous speech recognition is generally performed with reference to an acoustic model and a language model. The acoustic model includes information such as consonants and vowels that are basic sound units for use in speech recognition. In addition, the language model includes an n-gram language model and a word dictionary in which co-occurrence information is calculated by morphological analysis of the corpus (for example, Non-Patent Documents 1, 2, and 3).

In such a speech recognition technology, in order to improve the speech recognition rate, research that approaches from the improvement of the language model has also been made. For example, Non-Patent Document 4 proposes field adaptation of a language model using word clustering, and Non-Patent Document 2 proposes speech recognition using utterance prediction in speech dialogue. In the speech recognition technology proposed in Non-Patent Document 4, it is difficult to collect linguistic data specialized for the target field. Therefore, the vocabulary contained in the small corpus of the target field is regarded as a class, and it is easy to collect Words that behave similarly from the same class. In the speech recognition technique proposed in Non-Patent Document 2, utterance prediction using state transition is performed.
Satoshi Ogata, Yasuo Ariki, “Syllable-Based Acoustic Modeling for Japanese Spoken Speech Recognition”, IEICE Transactions, Japan, 2003, D-II, Vol. J86-D-II, no. 11, pp. 1523-1530 Takayuki Tamai, Ikuo Horiuchi, Atsushi Ichikawa, “Speech recognition using utterance prediction in spoken dialogue system” Information Processing Society of Japan, Research Report, “Spoken Language Information Processing”, Japan, 2002, No. 43-1 Kiyohiro Shikano, Katsunobu Ito, Tatsuya Kawahara, Kazuya Takeda, Mikio Yamamoto, “IT Text Speech Recognition System” Ohm, 2001 Shinsuke Mori, Nobuyasu Ito, Masafumi Nishimura, “Application of probabilistic language model by word clustering” Information Processing Society of Japan, Research Report “Spoken Language Information Processing”, Japan, 2003, No. 45-15

  However, the above-described speech recognition apparatus has the following problems. That is, in the speech recognition technique proposed in Non-Patent Document 4, even if there is a lot of data applied to the field, there are many recognition candidates depending on the corpus that generated the language model, and the speech is erroneously converted. Probability is high. Further, depending on the utterance content for which speech recognition is performed, the utterance prediction using the state transition proposed in Non-Patent Document 2 may not be suitable.

  Therefore, an object of the present invention is to provide a speech recognition device, a speech recognition program, a language model generation method used for the speech recognition device, and a language model generation device with an improved speech recognition rate.

  In describing the means for solving the above problems, first, a preliminary experiment will be described in which a plurality of language models and a plurality of acoustic models are combined and the speech recognition rate in each combination is confirmed by a speech recognition apparatus.

As acoustic models used in the preliminary experiment, the following five types of acoustic models recorded in the 2001 version of the Julius consortium were used.
・ Standard adult model ・ High-precision adult model (64 mixed)
・ High-precision adult model (128 mixed)
・ High-precision adult model (256 mixed)
・ Telephone acoustic model

  Here, the standard adult model is a model learned by the “newspaper reading speech corpus”. The high-accuracy adult model is a model learned from the “ATR multi-speaker speech database”, and has three types of mixed distributions (64, 128, 256 mixtures). The telephone acoustic model is a model learned by telephone voice.

Moreover, the following language model was used as a language model used for a preliminary experiment.
・ Web20k
・ MNK100k
・ Original language model 1
・ Original language model 2

  Here, the Web 20k is a language model recorded in the 2001 version of the Julius consortium, and is a language model learned from 20,000 vocabulary texts collected from Web pages. The MNP 100k is a language model recorded in the 2001 version of the Julius consortium, and is a language model learned from 100,000 vocabularies collected from newspaper article data. The self-made language model 1 is a language model in which transcription data for about 200 minutes at a call center is used as learning data. The self-made language model 2 is a language model using, as learning data, transcription data for 154 utterances excluding a portion where utterances overlap among utterances for 10 minutes in a call center. Note that the learning data of the self-made language model 2 includes transcription data for 154 utterances learned by the self-made language model 1.

  When each of the 20 combinations of the four language models and the five acoustic models is applied to a speech recognition apparatus, and the speech recognition apparatus recognizes speech in a call center, the result shown in FIG. 1 is obtained. FIG. 1 shows combinations of acoustic models and language models, and correct word recognition rates (voice recognition rates) of the respective combinations. As a result, the correct word recognition rate (speech recognition rate) was the highest in the combination of the self-made language model 2 generated using the transcription of the speech data 154 utterance as the learning data and the telephone acoustic model. Next, a correct word recognition rate (speech recognition rate) is shown in which the combination of the self-produced language model 1 generated using the transcription data of about 200 minutes including the speech data 154 utterance as learning data and the telephone acoustic model is good. From this, it was found that speech recognition can be performed with a high correct word recognition rate (speech recognition rate) by using a small amount of an appropriate language model rather than using a large amount of language models. Even if a large-scale language model is used to solve the problem of unknown words that cannot be recognized correctly when unknown words that are not included in the language model are present in the speech to be analyzed, The number of selections increases and conversion errors are likely to occur. This also shows that an appropriate language model with a small amount of text information in the corpus is effective in improving the accuracy of speech recognition.

  Based on the above results, in order to solve the above problems, the following means may be provided. That is, the speech recognition device according to the present invention is a speech recognition device that converts a speech signal obtained by converting a spoken speech into character information, and changes in a regular pattern of an utterance having a changeable portion and a fixed portion. Language model storage means for storing a language model representing a connection relationship of one or a plurality of sentence words each having one or a plurality of terms estimated to be spoken by a speaker in a possible part, and an acoustic including an acoustic characteristic of speech An acoustic model storage unit that stores a model, and a voice processing unit that converts a speech signal into character information with reference to a language model and an acoustic model are provided.

  According to this speech recognition apparatus, since the language model is generated from a fixed pattern of utterances, the speech processing means refers to such a language model and an acoustic model and converts appropriate candidates into character information. be able to. In addition, since the language model is generated from sentences generated by applying terms estimated to be spoken by the speaker to the changeable part of the fixed pattern, the speech processing means Can be converted from character candidates into character information, and the speech recognition rate is improved.

  Here, the term may be acquired from a site specified by a URL related to the speaker. Terms that are estimated to be spoken by speakers often include names of companies, organizations, corporations, etc. related to the speakers, business information, technical terms, etc. It frequently appears on the site specified by the URL related to the person. Therefore, by including a term acquired from a site specified by a URL related to the speaker in the language model, the speech processing means converts speech uttered by the speaker from appropriate candidates to character information. And the speech recognition rate is improved.

  The acoustic model may be a model learned by telephone voice. As a result, the voice processing means can appropriately convert voice spoken by telephone, such as conversation at a call center, into character information, and the voice recognition rate is improved.

  The speech recognition apparatus also includes a basic data storage unit that stores basic data in which an utterance is transcribed, a first corpus storage unit that stores a first corpus including character information related to a speaker, The second corpus storage means for storing a second corpus having a larger amount of character information than the amount of character information of the corpus, and a language model generation means for generating a language model, the language model generation means, First comparison means for comparing an appearance frequency related value related to the appearance frequency of a word appearing in the basic data and an appearance frequency related value related to the appearance frequency of the word appearing in the second corpus; Based on the comparison result of the comparison means, a fixed pattern generation means for generating a fixed pattern from the basic data in which the words appearing in the basic data with a predetermined frequency of occurrence or more are converted from the basic data, and the first corpus A second comparing means for comparing an appearance frequency related value related to the appearance frequency of the present word and an appearance frequency related value related to the appearance frequency of the word appearing in the second corpus; Based on the comparison result, a term extracting means for extracting terms appearing in the first corpus with a predetermined frequency of occurrence and a sentence in which the terms extracted by the term extracting means are applied to the variableized portion of the standard pattern are generated. It is preferable to have a sentence generation means for generating a connection relation and a connection relation generation means for generating a connection relation of words of the sentence.

  According to this configuration, the first comparison means compares the appearance frequency related value of the word in the basic data with the appearance frequency related value of the word in the second corpus, and based on the comparison result, the fixed pattern The generating means extracts words appearing in the basic data with a predetermined appearance frequency or more, and generates a fixed pattern. Thereby, a generalized fixed pattern can be obtained. Also, the second comparison means compares the appearance frequency related value of the word in the first corpus with the appearance frequency related value of the word in the second corpus, and based on the comparison result, the term extraction means Terms that appear in the first corpus with a frequency of appearance or higher are extracted. Therefore, the sentence generation means applies a term related to the speaker appearing in the first corpus at a predetermined appearance frequency or more to the variable part of the fixed pattern, so that an expected sentence of the speaker is generated. it can. Further, the connection relation generating means generates a connection relation between words appearing in the sentence based on the sentence. Therefore, the speech processing means can convert speech uttered by the speaker from proper candidates to character information, and the speech recognition rate is improved.

  Here, the first corpus may be character information acquired from a site specified by a URL related to the speaker, and the second corpus may be character information collected from the WWW. According to this configuration, since a term that is estimated to be spoken by a speaker frequently appears on a site specified by a URL related to the speaker, such a term can be easily extracted. . Also, since a lot of text information is posted on the WWW site, it is suitable as a language resource for the second corpus. Therefore, the voice recognition rate of the voice processing means is improved.

  The language model generation means may apply a term acquired from a site specified by the URL related to the speaker to the variableized portion of the fixed pattern based on the co-occurrence information of the second corpus. Since the co-occurrence information of the second corpus collected from the WWW language resources is used, the term obtained from the site specified by the URL related to the speaker can be easily converted into a variable part of the fixed pattern. And can be applied efficiently.

  Further, the voice recognition device stores a telephone / URL correspondence storage means for storing the telephone number of the speaker in association with the URL related to the speaker, and a telephone / URL correspondence storage means for receiving the information indicating the incoming telephone number. The URL stored in association with the incoming telephone number is read, the site specified by the read URL is accessed, character information is acquired from the accessed site, and stored in the first corpus storage means. And a WWW access means. As a result, the URL related to the other party's speaker specified by the incoming telephone number is read out by the telephone / URL correspondence storage means, and the site specified by the URL can be accessed by the WWW access means. The corpus can be obtained. As a site specified by a URL related to a speaker, for example, a website of a company, an organization, a corporation, etc. to which the speaker belongs can be cited. From such a site, for example, the name of the company, business information, professional Since characteristic words such as terms are included in the first corpus, an appropriate first corpus can be generated according to the incoming call from the other party's telephone, and the speech recognition rate is improved. .

  The voice recognition device further includes character editing means for correcting the character information converted by the voice processing means based on an input instruction or change instruction for character information from the input means, and the character recognition means has corrected the character information. The character information is preferably basic data stored in the basic data storage means. With this configuration, after the speech is converted into character information by the speech processing means, it is calibrated by the character editing means using the input means, and the calibrated character information is used as basic data, so that a more appropriate language model is generated as a language model. Can be generated by means.

  The speech recognition program according to the present invention is a speech processing program for converting speech uttered by a speaker into character information, and the computer uses a utterance fixed pattern having a changeable portion and a fixed portion. A language model storage means for storing a language model representing a connection relation of one or a plurality of sentences each having one or a plurality of terms estimated to be spoken by a speaker in a changeable portion; and an acoustic characteristic of speech An acoustic model storing means for storing an acoustic model, a language model and an acoustic model are referred to function as speech processing means for converting a speech signal into character information.

  According to this speech recognition program, since the language model is generated from a fixed pattern of utterances, it is possible to convert appropriate candidates into character information, and the speech recognition rate is improved. In addition, since the language model is generated from sentences generated by applying terms estimated to be spoken by the speaker to the changeable part of the fixed pattern, the speech processing means Can be converted from character candidates into character information, and the speech recognition rate is improved.

  In addition, the language model generation method according to the present invention generates a language model that generates a language model to be referred to together with an acoustic model including an acoustic characteristic of speech when the speech signal is acoustically analyzed and converted into character information in the speech recognition apparatus. A method of generating a fixed pattern of an utterance having a changeable part and a fixed part from basic data obtained by transcription of spoken speech data, and a spoken to a changeable part of the fixed pattern A language model generating step of generating one or a plurality of sentences fitted with one or a plurality of terms estimated to be uttered by the person and generating a language model representing a connection relation of words of the sentences .

  According to this language model generation method, the fixed pattern used for generating the language model is extracted from the basic data that transcribes the spoken speech data, so the language model suitable for the utterance content along the fixed pattern Can be generated. In addition, the fixed pattern has a changeable part and a fixed part, and after generating a sentence in which a term estimated to be spoken by the speaker is generated in the changeable part, the words in the sentence are connected. Since the language model representing the relationship is generated, the language model is suitable for conversion into appropriate character information in accordance with the speech uttered by the speaker.

  The term may be acquired from a site specified by a URL related to the speaker. Since terms that are estimated to be uttered by the speaker frequently appear on the site specified by the URL related to the speaker, such a term can be easily extracted.

  In this language model generation method, it is preferable to use, as basic data, character information in which character information converted by the speech recognition device is calibrated by editing. In this way, a more appropriate language model can be generated by calibrating the character information whose speech has been converted by the speech recognition apparatus and using it as basic data.

  The language model generation apparatus according to the present invention generates a language model that generates a language model to be referred to together with an acoustic model including an acoustic characteristic of speech when the speech signal is acoustically analyzed and converted into character information in the speech recognition apparatus. A basic data storage means for storing basic data in which an utterance is transcribed, a first corpus storage means for storing a first corpus including character information related to the speaker, and a first corpus A second corpus storage means for storing a second corpus having a character information amount larger than the character information amount of the character information, and a language model generation means for generating a language model, wherein the language model generation means includes basic data Comparing the appearance frequency related value related to the appearance frequency of the word appearing in the word with the appearance frequency related value related to the appearance frequency of the word appearing in the second corpus Based on the comparison result of the comparison means and the first comparison means, a fixed pattern generation means for generating a fixed pattern from the basic data, in which words appearing in the basic data with a frequency of occurrence of a predetermined frequency or more are generated from the basic data, and the first corpus A second comparing means for comparing an appearance frequency related value related to the appearance frequency of the appearing word and an appearance frequency related value related to the appearance frequency of the word appearing in the second corpus; Based on the comparison result, a term extracting means for extracting terms appearing in the first corpus with a predetermined frequency of occurrence and a sentence in which the terms extracted by the term extracting means are applied to the variableized portion of the standard pattern are generated. And a connection relation generating means for generating a connection relation of words of the sentence.

  According to this language model generation device, the first comparison unit compares the appearance frequency related value of the word in the basic data with the appearance frequency related value of the word in the second corpus, and based on the comparison result The standard pattern generation means extracts words that appear in the basic data at a predetermined appearance frequency or higher, and generates a standard pattern. Thereby, a generalized fixed pattern can be obtained. Also, the second comparison means compares the appearance frequency related value of the word in the first corpus with the appearance frequency related value of the word in the second corpus, and based on the comparison result, the term extraction means Terms that appear in the first corpus with a frequency of appearance or higher are extracted. Therefore, the sentence generation means applies a term related to the speaker appearing in the first corpus at a predetermined appearance frequency or more to the variable part of the fixed pattern, so that an expected sentence of the speaker is generated. it can. Further, the connection relation generating means generates a connection relation between words appearing in the sentence based on the sentence. Therefore, the language model generation apparatus can convert speech uttered by a speaker from appropriate candidates to character information, and can generate a language model that improves the speech recognition rate.

  ADVANTAGE OF THE INVENTION According to this invention, the candidate in speech recognition can be optimized and a speech recognition rate can be improved.

  Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the description of the drawings, the same or corresponding elements are denoted by the same reference numerals, and redundant description is omitted.

<Configuration of communication line system>
FIG. 2 is a schematic diagram showing a communication line system 2 using the speech recognition apparatus 1 of the present embodiment. The call center 3 is provided with a voice recognition device 1 and a telephone 4, and the telephone 4 can communicate with a telephone 7 on the other side via a communication network 6. The telephone 4 and the voice recognition apparatus 1 are connected by an adapter 9 established on the communication line 8, and the telephone 4 and the telephone 7 generate a voice signal that is spoken by a speaker, and the voice recognition apparatus 1 Send. In addition, a server 11 is installed as a partner who makes a call using the telephone 7, and is connected to the communication line network 6 through the adapter 10. The voice recognition device 1 is a device that converts the conversation of the speaker into character information, and can access a site established by the server 11 via the communication network 6 so that the site can be referred to. It has become.

  Although FIG. 1 shows one telephone 4 in the call center 3 and one external telephone 7, a plurality of telephones 4 corresponding to a plurality of operators are installed in the actual call center 3. It corresponds to the other party's telephones coming from a plurality of telephones 7 from the outside. In addition, the voice recognition device 1 can access each site established by a large number of servers via the communication line network 6. Each site is established for each company, organization, various corporation, etc., and the name of the company, business information, technical terms, etc. are frequently displayed on the site.

<Configuration of voice recognition device>
The voice recognition device 1 is physically configured to include a CPU, a memory such as a RAM, a display device such as a display, an interface, a keyboard and a mouse input unit, and the like as hardware. FIG. 3 is a configuration diagram showing the speech recognition apparatus 1 for each functional block. The speech recognition apparatus 1 includes a language model storage unit 21, an acoustic model storage unit 22, a basic data storage unit 23, a large-scale corpus storage unit 24, a small-scale corpus storage unit 25, and a telephone / URL correspondence storage unit 27 as storage units. Is provided. Further, the voice recognition device 1 includes an interface I / F 29. The interface I / F 29 receives a voice signal of a voice conversation between the speakers via the telephone 4 and the telephone 7, and outputs the voice signal to the voice processing unit 31, and also to an external site specified by the URL. It is also an interface for access. Further, the speech recognition apparatus 1 includes a speech processing unit 31, a character editing unit 32, a language model generation unit 33, a WWW access unit 34, and a display unit 37 as processing units.

  The language model storage unit 21 stores language models. This language model is a keyword obtained from a site specified by a URL pattern of an organization (company, organization, etc.) to which a predicted or planned speaker belongs, for example, an organization name, a handling product, business information, A predicted utterance expression is generated from business terminology, etc., and is generated from the utterance expression. By morphological analysis of the predicted utterance expression, a word that appears frequently at the beginning of a sentence or a word that appears next to a certain word is connected. Represents a relationship. The acoustic model storage unit 22 stores an acoustic model that is acoustic information for each phonetic symbol such as a vowel or a consonant, and this acoustic model is learned by telephone speech.

  The basic data storage unit 23 stores basic data that is text data obtained by transcription of call center voice data. The large-scale corpus storage unit 24 stores a 12 GB large-scale corpus (second corpus) including character information randomly collected from the language resources of the Internet WWW 40 by recursively following links on the WWW. To do. Unnecessary symbols and garbled characters are deleted from the large corpus. The small corpus 25 stores character information acquired from a site specified by the URL of the organization to which the speaker belongs as a small corpus (first corpus).

  The audio data storage unit 26 stores audio data input via the interface I / F 29. The telephone / URL correspondence storage unit 27 stores the telephone number and the URL of the organization to which the speaker calling by the telephone number is associated with each other.

  The voice processing unit 31 receives the voice signal output from the interface I / F 29 and performs acoustic analysis such as frequency analysis of the voice signal. Furthermore, the speech processing unit 31 refers to the language model stored in the language model storage unit 21 to list words that are likely to appear at the beginning of the sentence, and the result of acoustic analysis of the input speech signal and the acoustic model storage unit The acoustic model stored in 22 is collated. Next, a list of language models referring to word candidates that can be connected to the word at the beginning of the sentence is listed, and the result of acoustic analysis of the speech signal is compared with the acoustic model. In this way, the input voice signal is sequentially recognized and converted into character information. Then, the converted character information is output to the display unit 37.

  The character editing unit 32 gives an input instruction or change instruction from the input unit 36 composed of a keyboard and a mouse when there is a character converted in error in the character information output by the voice processing unit 31. Receive and proofread and edit text information. The character editing unit 32 may be a general text creation application. The character editing unit 32 stores the edited character information in the basic data storage unit 23.

  The language model generation unit 33 uses the basic data stored in the basic data storage unit 23, the large-scale corpus stored in the large-scale corpus storage unit 24, and the small-scale corpus stored in the small-scale corpus storage unit 25. A language model to be stored in the language model storage unit 21 is generated. That is, the language model generation unit 33 generates a fixed pattern by variableizing words whose difference between the information amount of words appearing in the basic data and the information amount of words appearing in the large-scale corpus is greater than or equal to a predetermined value. In addition, the language model generation unit 33 identifies a word whose difference between the information amount of the word appearing in the small corpus and the information amount of the word appearing in the large corpus is a predetermined value or more by the URL related to the speaker. Character information acquired from a site is extracted as a frequent keyword. Further, the language model generation unit 33 generates a language model based on a sentence in which the character information is applied to a variable part of the standard pattern, and outputs the language model to the language model storage unit 21. Here, when the language model generation unit 33 applies the character information acquired from the site specified by the URL to the variableized portion of the fixed pattern, it applies it based on the co-occurrence information of the large-scale corpus.

The information amount I (x) is a value related to the appearance frequency of words, and is calculated statistically as follows. That is,
I (x) = − log ₂ P (x) (Formula 1)
However,
P (x) = F (x) / N (Formula 2)
Here, N is the total number of words, and F (x) is the number of occurrences of the word x. The smaller the occurrence number F (x), the larger the information amount I (x). In order to perform a logarithmic operation, basic data may be included in a large-scale corpus so that P (x) does not become zero.

  When the WWW access unit 34 receives the incoming telephone number via the interface I / F 29, the WWW access unit 34 refers to the telephone / URL correspondence storage unit 27 and stores the URL stored in association with the incoming telephone number. Read and access the site specified by the read URL. The WWW access unit 34 acquires character information from the accessed site and stores it in the small-scale corpus storage unit 25.

  The display unit 37 receives the character information output from the voice processing unit 31, displays the character information on the screen, displays the character information edited by the character editing unit 32, and also displays the character information edited by the character editing unit 32. Display sites accessed by.

<Language model generation method>
Here, a method of generating a language model stored in the language model storage unit 21 will be described with reference to FIGS. 4 and 5. In the speech recognition device 1, the language model is performed by the language model generation unit 33. The language model generation unit 33 includes a first comparison unit, a fixed pattern generation unit, a second comparison unit, a term extraction unit, a sentence generation unit, and a connection relationship generation unit that function as described below.

  Basic data, which is text data obtained by writing call center voice data in advance, is created and stored in the basic data storage unit 23 (S101). A fixed pattern at the call center is extracted from the basic data (S102). For extraction of the fixed pattern, a large-scale corpus stored in the large-scale corpus storage unit 24 is used. The large-scale corpus is a 12 GB corpus collected from the WWW 40 language resources of the Internet as described above. The extraction of the fixed pattern is as follows. First, the first comparison unit obtains a difference between the information amount of each word appearing in the basic data and the information amount of each corresponding word in the large-scale corpus, and descends in descending order based on the magnitude of the difference in the information amount. Sort. Then, the fixed pattern generation unit first selects a word whose information amount difference is a predetermined value or more (for example, 8.0 or more). FIG. 6 shows an example of a word having a difference in information amount of 8.0 or more. Furthermore, the fixed pattern generation unit converts the selected word into a variable (“@VAR”), thereby extracting a fixed pattern as shown in FIG. As described above, the fixed pattern has a variable part and a fixed part which can be changed.

  Next, the second comparison unit obtains a difference between the information amount of each word appearing in the small corpus and the information amount of each corresponding word in the large corpus for comparison determination. Then, the term extraction unit extracts words whose information amount difference is equal to or greater than a predetermined value as keywords (S103). In keyword extraction, more appropriate keywords can be extracted by limiting the part of speech that can be input to the variableized portion of the fixed pattern. In addition, an appropriate keyword can be extracted by selecting a word having the same characteristics as the variable part of the word using a thesaurus. This keyword is extracted from the character information acquired from the site specified by the URL of the organization to which the predicted or planned speaker belongs. Keywords include the name of the organization (eg, company, organization) to which the speaker belongs, the speaker's name, speaker attribute data, products handled by the organization, business terms, business information, technical terms, etc. Words that can be expected. This keyword is a word that is assumed to change depending on the speaker. The small-scale corpus stored in the small-scale corpus storage unit 25 is the site accessed by the WWW access unit 34 accessing the server 11 (see FIG. 2) that establishes a URL related to the speaker. Character information obtained from

  Further, the sentence generation unit determines whether the keyword selected in S103 can be substituted into the fixed pattern extracted in S102 using the co-occurrence information of the large-scale corpus, and assigns an appropriate keyword to the fixed pattern, A predicted sentence is generated (S104). In addition, the connection relation generation unit analyzes the generated predicted sentence by morpheme (word, part of speech), finds the probability of the word appearing at the beginning of the utterance, the connection probability of the word appearing next to each word, and the connection relation Generate a language model. The speech recognition apparatus 1 may be configured by storing the language model generated in this way in the language model storage unit 21.

  Further, the speech recognition apparatus 1 performs speech recognition (S106), the converted character information is proofread (S107), the proofread character information is used as the basic data in S101, and the language is renewed by the steps S102 to S105. A model may be generated and stored in the language model storage unit 21. In this proofreading process, the operator re-listens the voice data stored in the voice data storage unit 26 and uses the character editing unit 32 and the input unit 36 to proofread the sentence in which the voice is converted into character information. In this case, since the character information whose dialogue has been proofread is used as the basic data, a more appropriate language model can be generated by the language model generation unit 33.

  The acquisition of a small-scale corpus used in keyword extraction in S103 can also be performed as follows. In other words, when receiving a call from the telephone 7 on the other side, the WWW access unit 34 receives the incoming telephone number from the telephone 7 via the interface I / F 29. Then, the WWW access unit 34 refers to the telephone / URL correspondence storage unit 27 and reads out URL information stored in association with the incoming telephone number. Further, the WWW access unit 34 accesses a site specified by the read URL, acquires character information from the accessed site, and stores it in the small-scale corpus storage means. In this way, a small corpus may be acquired.

  The language model is generated from the latest predetermined amount of transcription data, and it is preferable not to use past data after the predetermined amount. Or you may make it produce | generate a language model using the content with high utterance frequency in transcription data. In this way, the basic data used to generate the language model is always recent data or frequently used data, so that candidates for speech recognition can be optimized, and the speech recognition rate is improves.

  A speech recognition program according to an embodiment of the present invention will be described. FIG. 8 is a configuration diagram of a recording medium on which a voice recognition program is recorded. The storage area 210 of the recording medium 200 records the speech recognition program according to the embodiment (hereinafter referred to as “this program”), and includes the speech processing module 211, the character editing module 212, the language model as its constituent elements. A generation module 213 and a WWW access module 214 are provided. The functions realized by operating these modules are the same as the functions of the speech processing unit 31, the character editing unit 32, the language model generation unit 33, and the WWW access unit 34 in the speech recognition apparatus 1 described above. . The language model data 215, acoustic model data 216, basic data 217, large-scale corpus data 218, small-scale corpus data 219, and telephone / URL correspondence data 220 stored in the storage area 210 are the languages in the speech recognition device 1. The language model stored in the model storage unit 21, the acoustic model stored in the acoustic model storage unit 22, the basic data stored in the basic data storage unit 23, the large-scale corpus stored in the large-scale corpus storage unit 24, the small This is the same as the data stored in the small corpus stored in the large corpus storage 25 and the telephone / URL correspondence storage 27.

  As mentioned above, although preferred embodiment of this invention was described in detail, it cannot be overemphasized that this invention is not limited to the said embodiment. For example, in the above embodiment, the fixed pattern is extracted from the basic data and the large corpus from which the voice data is transcribed. However, the present invention is not limited to this, and a fixed pattern uttered by the operator at the call center is determined in advance. The fixed pattern may be written into text data. In the above embodiment, keywords related to the speaker are extracted from the small corpus and the large corpus. However, the present invention is not limited to this, and the site is specified directly by referring to the site specified by the URL of the organization to which the speaker belongs. Necessary keywords may be selected. Moreover, the speech recognition apparatus 1 in the above embodiment can be applied not only to Japanese but also to other languages such as English. The calculation method of the appearance frequency related value is not limited to Formulas 1 and 2, and other methods may be used.

  Such a voice recognition device 1 can be applied to a call center that accepts company product orders, a call center in a securities company, a consumer consultation center, etc., as well as voice communication between police and fire department reporting centers, defense-related facilities, aviation It can also be applied to voice communication between controllers and pilots, voice analysis of voice recorders, etc.

It is a figure which shows the combination of the acoustic model in a preliminary experiment, and a language model, and its positive word recognition rate. 1 is a schematic diagram of a communication line system according to an embodiment. It is a functional block diagram of the speech recognition apparatus shown in FIG. It is a flowchart which produces | generates a language model. It is a conceptual diagram for demonstrating the production | generation method of a language model. It is a figure for demonstrating the difference of the information content in a corpus. It is a figure which shows the example of the variable-ized basic data. It is a block diagram of the recording medium with which the speech recognition program was recorded.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Voice recognition apparatus, 21 ... Language model storage part, 22 ... Acoustic model storage part, 23 ... Basic data storage part, 24 ... Large scale corpus storage part, 25 ... Small corpus storage part, 27 ... Telephone / URL correspondence storage Part 31 sound processing part 32 character variation part 33 language model generation part 34 WWW access part

Claims (13)

In a speech recognition device that converts speech signals obtained by converting spoken speech into character information,
A language that represents a connection relationship of one or a plurality of sentence words each having one or a plurality of terms estimated to be spoken by a speaker in the changeable part in a typical pattern of utterances having a changeable part and a fixed part Language model storage means for storing the model;
Acoustic model storage means for storing an acoustic model including acoustic characteristics of speech;
A speech recognition device comprising speech processing means for converting speech signals into character information with reference to the language model and the acoustic model.   The speech recognition apparatus according to claim 1, wherein the term is acquired from a site specified by a URL related to a speaker.   The speech recognition apparatus according to claim 1, wherein the acoustic model is a model learned by telephone speech. Basic data storage means for storing basic data in which utterances are transcribed;
First corpus storage means for storing a first corpus including character information relating to a speaker;
Second corpus storage means for storing a second corpus having a character information amount larger than that of the first corpus;
Language model generation means for generating the language model,
The language model generation means includes
First comparison means for comparing an appearance frequency related value related to an appearance frequency of a word appearing in the basic data and an appearance frequency related value related to the appearance frequency of the word appearing in the second corpus;
Based on the comparison result of the first comparison unit, a fixed pattern generation unit that generates the fixed pattern from the basic data, which is a variable of words that appear in the basic data at a predetermined frequency of appearance,
Second comparing means for comparing an appearance frequency related value related to an appearance frequency of a word appearing in the first corpus and an appearance frequency related value related to the appearance frequency of the word appearing in the second corpus; ,
Term extracting means for extracting the term that appears in the first corpus at a predetermined frequency of appearance based on the comparison result of the second comparing means;
Sentence generating means for generating a sentence in which the term extracted by the term extracting means is applied to the variable part of the fixed pattern;
The speech recognition apparatus according to claim 1, further comprising a connection relationship generation unit configured to generate a connection relationship between words of the sentence. The first corpus is character information acquired from a site specified by the URL related to a speaker;
The speech recognition apparatus according to claim 4, wherein the second corpus is character information collected from the WWW.   The speech recognition apparatus according to claim 4, wherein the language model generation unit applies the term to a variable portion of the fixed pattern based on the co-occurrence information of the second corpus. A telephone / URL correspondence storing means for storing the telephone number of the speaker in association with the URL related to the speaker;
Receiving information indicating the incoming telephone number, referring to the telephone / URL correspondence storage means, reading the URL stored in association with the incoming telephone number, and accessing the site specified by the read URL The speech recognition apparatus according to claim 5, further comprising: a WWW access unit that acquires character information from the accessed site and stores the character information in the first corpus storage unit. Further comprising character editing means for correcting the character information converted by the voice processing means based on an input instruction or change instruction of the character information from the input means,
8. The speech recognition apparatus according to claim 5, wherein the character data modified by the character editing unit is used as the basic data stored in the basic data storage unit. In a speech recognition program for converting a speech signal obtained by converting spoken speech into character information,
Computer
A language that represents a connection relationship of one or a plurality of sentence words each having one or a plurality of terms estimated to be spoken by a speaker in the changeable part in a typical pattern of utterances having a changeable part and a fixed part Language model storage means for storing the model;
Acoustic model storage means for storing an acoustic model including acoustic characteristics of speech;
A speech recognition program that functions as speech processing means for converting speech signals into character information with reference to the language model and the acoustic model. A language model generation method for generating a language model to be referred to together with an acoustic model including an acoustic characteristic of speech when an acoustic signal is acoustically analyzed and converted into character information in a speech recognition device,
A fixed pattern generation step for generating a fixed pattern of an utterance having a changeable part and a fixed part from basic data obtained by transcribing spoken voice data;
A language that generates one or a plurality of sentences in which one or a plurality of terms estimated to be spoken by a speaker in the changeable part of the fixed pattern is generated, and generates a language model that represents a connection relation of words in the sentences A language model generation method comprising: a model generation step.   The language model generation method according to claim 10, wherein the term is acquired from a site specified by a URL related to a speaker.   The language model generation method according to claim 10 or 11, wherein character information obtained by calibrating character information converted by the voice recognition device is used as the basic data. A language model generation device that generates a language model that is referred to together with an acoustic model including an acoustic characteristic of speech when an acoustic signal is acoustically analyzed and converted into character information in a speech recognition device,
Basic data storage means for storing basic data in which utterances are transcribed;
First corpus storage means for storing a first corpus including character information related to a speaker;
Second corpus storage means for storing a second corpus having a character information amount larger than the character information amount of the first corpus;
Language model generation means for generating the language model,
The language model generation means includes
First comparison means for comparing an appearance frequency related value related to an appearance frequency of a word appearing in the basic data and an appearance frequency related value related to the appearance frequency of the word appearing in the second corpus;
Based on the comparison result of the first comparison unit, a fixed pattern generation unit that generates the fixed pattern from the basic data by converting words that appear in the basic data with a frequency of appearance higher than a predetermined frequency,
Second comparing means for comparing an appearance frequency related value related to an appearance frequency of a word appearing in the first corpus and an appearance frequency related value related to the appearance frequency of the word appearing in the second corpus; ,
A term extracting means for extracting the term that appears in the first corpus at a predetermined appearance frequency or more based on a comparison result of the second comparing means;
Sentence generating means for generating a sentence in which the term extracted by the term extracting means is applied to the variable part of the fixed pattern;
A language model generation apparatus comprising: a connection relationship generation unit configured to generate a connection relationship between words of the sentence.

Images