JP2009075582A - Terminal device, language model creation device, and distributed speech recognition system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a terminal device, a language model creation device and a distributed speech recognition system, which easily improve the accuracy of speech recognition with respect to different notation depending on a context. <P>SOLUTION: A mobile phone 200 has a speech data transmission unit 208 for transmitting speech data to a speech recognizer which performs speech recognition processing by using a language model, and a mail processing unit 205 for transmitting a mail text of normal transmission mail to the language model creation device which creates the language model as the mail for language model creation. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a terminal device that transmits voice data to a speech recognition device that performs voice recognition and performs speech recognition using a language model, a language model creation device that creates a language model, and a distributed type using these devices The present invention relates to a speech recognition system.

  In recent years, in various terminal devices such as mobile phones, character strings are created by voice input using a microphone instead of character input using a key switch (see, for example, Patent Document 1 and Patent Document 2). .

  In the techniques described in Patent Document 1 and Patent Document 2, a speech recognition device that converts speech data into text by speech recognition is arranged on a network. The terminal device transmits voice data to the voice recognition device. The voice recognition device creates a character string from the voice data by referring to a voice recognition database including an acoustic model, a dictionary, and a language model, and returns the character string to the terminal device. As a result, the terminal device can easily create a character string such as a mail text.

  Patent Documents 1 and 2 describe techniques for updating a speech recognition database.

  In the techniques described in Patent Literature 1 and Patent Literature 2, the speech recognition device refers to the dictionary of the speech recognition database, converts speech data received from the terminal device into a character string, and returns the character string to the terminal device. The terminal device accepts user correction for the character string created by the voice recognition device, and transmits the corrected character string to the voice recognition device. The voice recognition device creates and transmits a mail having the received character string as a mail text, and corrects the dictionary of the voice recognition database based on the correction portion. Thereby, the recognition accuracy of the word notation with respect to the reading kana can be improved.

  Furthermore, Patent Literature 1 and Patent Literature 2 describe a technique for creating a dictionary for each user.

In the technologies described in Patent Literature 1 and Patent Literature 2, the terminal device transmits voice data and a character string to the voice recognition device in association with the caller number. The voice recognition device creates a plurality of dictionaries associated with the caller telephone number. The voice recognition apparatus performs voice recognition using a dictionary corresponding to the caller telephone number of the voice data transmission source and corrects the dictionary corresponding to the caller telephone number of the character string transmission source. Thereby, the dictionary for every user reflecting the tendency of different word notation for every user attribute can be created, and the precision of voice recognition can be improved.
JP 2002-215615 A JP 2001-309049 A

  By the way, the language model included in the speech recognition database is usually created by performing predetermined statistical information processing on a character string prepared as a learning target. The language model is obtained by converting the appearance probability and connection probability into data for each word described in the dictionary.

  Since context features differ from user to user, the appearance probability and connection probability of each word also vary from user to user. Therefore, in order to improve the accuracy of speech recognition, it is desirable to perform speech recognition in consideration of such differences.

  However, with the techniques described in Patent Document 1 and Patent Document 2, since the dictionary is only corrected from the corrected portion of the character string, a language model cannot be created in consideration of the difference in context for each user. That is, with the techniques described in Patent Literature 1 and Patent Literature 2, it is difficult to improve the accuracy of speech recognition for different notations depending on the context.

  Therefore, it is conceivable to create a language model for each user using the character strings sent from the terminal device in the speech recognition devices of Patent Literature 1 and Patent Literature 2. Thereby, it is possible to easily obtain a sufficient amount of character strings from each user to create a language model.

  However, when the techniques described in Patent Literature 1 and Patent Literature 2 are used, a system configuration in which the voice recognition device is positioned on the route of the outgoing mail, such as providing a voice recognition device in the mail server, is required. It is difficult to apply to the system. In other words, since it takes time and effort to construct a system, it is difficult to improve the accuracy of speech recognition for different notations depending on the context.

  The present invention has been made in view of such points, and a terminal device, a language model creation device, and a distributed speech recognition system that can easily improve the accuracy of speech recognition for different notations depending on the context. The purpose is to provide.

  The terminal device according to the present invention includes voice data transmitting means for transmitting voice data to a voice recognition device that performs voice recognition processing using a language model, and a mail text of a normal outgoing mail as a language model creation mail, A configuration having a mail transmitting means for transmitting to a language model creating apparatus for creating a language model is adopted.

  The language model creation device of the present invention is a language model creation device that creates a language model used for speech recognition processing using a language model creation mail received from a terminal device, and includes the ID information and a mail text. Mail receiving means for receiving language model creation mail, mail processing means for extracting mail text and ID information from the received language model creation mail, learning the extracted mail text, and for each ID information And a language model creating means for creating the language model.

  The distributed speech recognition system of the present invention includes a speech recognition device that performs speech recognition processing on speech data using a language model, a terminal device that transmits speech data to the speech recognition device, and the language model by learning a character string. A distributed speech recognition system comprising: a language model creation device that creates a language model creation email by editing a destination of a normal transmission email, and the language model creation device The language model creation device learns the received mail body of the language model creation email to create the language model, and the speech recognition device applies the speech data received from the terminal device to the speech data The speech recognition process is performed using the language model.

  According to the present invention, since the mail body of the outgoing mail is collected by mail, a sufficient amount of character strings for creating a language model for each user is collected from each user without changing the existing system. be able to. Thereby, the accuracy of speech recognition for different notations depending on the context can be easily improved.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

(Embodiment 1)
FIG. 1 is a system configuration diagram showing an example of a configuration of a speech recognition system as a distributed speech recognition system according to Embodiment 1 of the present invention. The present embodiment is an example in which the present invention is applied to a system for creating a mail text on a mobile phone using voice recognition.

  In FIG. 1, the voice recognition system 100 includes mobile phones 200-1 to 200 -M, a mail server 300, and a voice recognition server 400. These devices are connected to the communication network 500 by radio or wire. The mobile phones 200-1 to 200-M have the same configuration and include the terminal device of the present invention. The speech recognition server 400 includes the language model creation device of the present invention and a speech recognition database.

  The mobile phone 200 has a mobile phone function and a mail transmission / reception function, and further has a voice input function for creating a mail text by voice input using the voice recognition server 400. Specifically, the cellular phone 200 transmits voice data indicating the characteristics of the input voice to the voice recognition server 400 in association with ID (identifier) information of the own device. Then, the cellular phone 200 receives character string data (hereinafter referred to as “text data”) as a result of the voice recognition by the voice recognition server 400. The mobile phone 200 transmits the mail text created by the voice input function in this way to the arbitrary destination via the mail server 300 by mail.

  In addition, each time the mobile phone 200 transmits a mail, the mobile phone 200 transmits the text string of the mail body of the transmitted mail to the voice recognition server 400 by mail in association with the ID information of the own device. This mail text is used in the speech recognition server 400 to create a language model for each user of the mobile phone 200 (hereinafter referred to as “user language model”).

  The mail server 300 manages mail transmission / reception of the mobile phone 200.

  The voice recognition server 400 has a voice recognition database, and performs a predetermined voice recognition process based on the voice recognition database. The voice recognition server 400 receives the voice data in association with the ID information of the mobile phone 200, and when there is a corresponding user-specific language model, performs voice recognition processing using the user-specific language model. Then, the voice recognition server 400 returns the text data created as a result of the voice recognition to the voice data transmission source.

  The voice recognition server 400 receives a character string sent from the mobile phone 200 in association with the ID information of the mobile phone 200 by mail. Then, the speech recognition server 400 learns the received character string and creates a user-specific language model associated with the ID information of the mobile phone 200.

  The communication network 500 is, for example, the Internet. The mobile phone 200, the mail server 300, and the voice recognition server 400 communicate with each other via the communication network 500 using a communication protocol such as TCP / IP (transmission control protocol). In addition, the mobile phone 200, the mail server 300, and the voice recognition server 400 perform mail transmission / reception using SMTP (simple mail transfer protocol) and POP3 (post office protocol version 3).

  According to such a voice recognition system 100, each time a mail is transmitted from the mobile phone 200, a character string having the same content as the mail text of the transmitted mail is transmitted to the voice recognition server 400 by mail. As a result, the same character string as the mail text created by the user of the mobile phone 200 is automatically transmitted to the voice recognition server 400. That is, a sufficient amount of text data for creating a user-specific language model can be collected in the speech recognition server 400 without causing the user to have special consciousness or operation and without complicated processing. In addition, as the user creates a mail text using the voice recognition function, the user can gradually learn the features of the mail text unique to the user together with the context information, thereby improving the voice recognition accuracy.

  Next, the configuration of the mobile phone 200 will be described.

  FIG. 2 is a block diagram showing a configuration of the mobile phone 200.

  As shown in FIG. 2, the mobile phone 200 includes an ID storage unit 201, a radio unit 202, an antenna unit 203, an operation unit 204, a mail processing unit 205, a microphone 206, a feature amount extraction unit 207, a voice data transmission unit 208, a text A data receiving unit 209, a display 210, a speaker 211, and a control unit 212 are included. The mail processing unit 205 includes a BCC (blind carbon copy) generation unit 213.

  The ID storage unit 201 stores ID information unique to the mobile phone 200. The ID storage unit 201 is, for example, a SIM (subscriber identity module) installed in many mobile phones 200. The ID information in this case is, for example, a subscriber identification number (IMSI: International mobile subscriber identity) or an identification number (SIMNO) assigned to the SIM.

  The wireless unit 202 performs wireless communication with a wireless base station (not shown) arranged in the communication network 500 via the antenna unit 203 and connects to the communication network 500.

  The operation unit 204 includes a key switch (not shown) and accepts various user operations including a character input operation and an instruction operation for starting mail creation by voice recognition.

  The mail processing unit 205 creates a mail body based on the user's character input operation and the text data received from the voice recognition server 400. Then, the mail processing unit 205 transmits the created mail text to the mail server 300 via the wireless unit 202 by designating an arbitrary mail address as a destination.

  Each time a mail is transmitted, the BCC generation unit 213 associates the same character string with the mail text of the transmitted mail with the ID information stored in the ID storage unit 201 and transmits the same to the voice recognition server 400 by mail. To do. Specifically, the BCC generation unit 213 uses the BCC function to add the voice recognition server 400 to the destination of the mail body of the outgoing mail, and sends a copy mail of the outgoing mail to the voice recognition server 400.

  The microphone 206 inputs the voice including the user's uttered voice and converts it into a voice signal.

  The feature quantity extraction unit 207 analyzes the voice signal output from the microphone 206 and extracts the feature quantity used for voice recognition by the voice recognition server 400. Specifically, the feature amount extraction unit 207 performs frame processing on the speech signal, performs predetermined processing including Fourier analysis for each frame, and performs speech feature amounts such as cepstrum parameters (hereinafter simply referred to as “feature amount”). Extract). Then, the feature amount extraction unit 207 detects a speech section including the user's voice from the analysis result, and generates time-series data based only on the feature amount of the speech section.

  The voice data transmission unit 208 establishes a session for transmitting and receiving voice data and text data to and from the voice recognition server 400 via the wireless unit 202 when creation of a mail body by voice recognition is started. The ID information stored in the ID storage unit 201 is transmitted to the voice recognition server 400. Then, the voice data transmission unit 208 packetizes the voice data output from the feature amount extraction unit 207 and transmits it to the voice recognition server 400.

  The text data receiving unit 209 receives text data returned from the speech recognition server 400 as a speech recognition result for the speech data via the wireless unit 202.

  The display 210 displays a text document as a mail body candidate based on the character input operation by the user and the text data received from the speech recognition server 400. The display 210 displays various information related to the operation of the mobile phone 200.

  The speaker 211 outputs, for example, voice data transmitted from the other party in the mobile phone function.

  The control unit 212 includes a CPU (central processing unit), a storage medium such as a ROM (read only memory) storing a control program, a working memory such as a RAM (random access memory), and the like. Control each part. Each unit of the mobile phone 200 includes, for example, an ASIC (Application Specific Integrated Circuit) and a communication circuit.

  Next, the configuration of the voice recognition server 400 will be described.

  FIG. 3 is a block diagram showing a configuration of the voice recognition server 400.

  As illustrated in FIG. 3, the speech recognition server 400 includes a network interface (I / F) unit 410, a user-specific language model creation unit 420, a speech recognition database (DB) 430, and a speech recognition unit 440. .

  The network interface unit 410 is connected to the communication network 500 by wire.

  The user-specific language model creating unit 420 receives the character string associated with the ID information of the mobile phone 200 sent from the mobile phone 200 via the mail server 300 via the network interface unit 410. Then, the user-specific language model creation unit 420 analyzes the received character string, creates a user-specific language model in association with the ID information, and updates the speech recognition database 430. The user-specific language model creation unit 420 includes a mail reception unit 421, a database (DB) switching unit 422, and a language model creation unit 423. The mail reception unit 421 includes a document extraction unit 424 and an ID extraction unit 425.

  The mail receiving unit 421 receives mail addressed to its own device. Specifically, the mail receiving unit 421 acquires mail from the mail server 300 that is addressed to the domain name of the voice recognition server 400.

  The document extraction unit 424 extracts the mail text of the received mail as a character string to be learned and outputs it to the language model creation unit 423.

  The language model creation unit 423 performs predetermined statistical information processing on the mail text extracted by the mail reception unit 421 to create a user-specific language model. Specifically, when an already existing user-specific language model is a processing target, a language model is created according to the statistical information processing result of the newly received mail body, and the user-specific language model 434 is corrected. . When a new user-specific language model is created, the created language model is registered in the user-specific language model 434 in association with the ID information extracted from the base received mail.

  On the other hand, the ID extraction unit 425 extracts the ID information of the transmission source mobile phone 200 associated with the received mail, and outputs it to the database switching unit 422. Specifically, the ID extraction unit 425 extracts the source ID information from the account name of the destination address of the received mail. The configuration of the destination address when the mobile phone 200 transmits mail to the voice recognition server 400 will be described later.

  The database switching unit 422 switches the language model for each user that is the target of the registration process and the update process by the language model creation unit 423. Specifically, the database switching unit 422 selects a language model for each user corresponding to the ID information extracted from the received mail by the ID extraction unit 425 from the speech recognition database 430 as a processing target of the language model creation unit 423.

  Here, the language model will be described. The language model is obtained by modeling a context pattern in a character string to be learned by the number of related vocabularies, and includes N-gram models such as a trigram model, a bigram model, and a unigram.

  An N-gram is a combination of N consecutive words. The N-gram model is obtained by extracting N-grams from a character string to be learned and calculating the appearance probability for each of the extracted N-grams. In actuality, in the creation of an N-gram model, the vocabulary handled as a learning target is restricted, such as ignoring function words and proper nouns, and N-grams with a low appearance frequency are cut off. Thereby, the amount of calculation can be reduced. Moreover, smoothing of the probability in consideration of N-grams that do not appear is performed. Thereby, it is possible to prevent the voice recognition accuracy from being deteriorated due to a lack of statistics in the initial stage of learning.

  That is, since the language model for each user is created based on only the character string created by the user for each user, the user language model more reflects the context pattern of the user. For example, when a mail including a character string “Contact Ms. Oda” is transmitted several times from a certain mobile phone 200, “Contact Ms. Oda” or “Contact Ms. Oda” in the corresponding language model 434 for each user. The appearance probability of the character string “Contact Oda” is higher than the appearance probability of.

  The voice recognition database 430 stores various data used by the voice recognition unit 440 in a predetermined voice recognition process. The speech recognition database 430 stores an acoustic model 431, a dictionary 432, a common language model 433, and a user-specific language model 434. Here, a state is shown in which M language-specific language models 434-1 to 434 -M corresponding to the mobile phones 200-1 to 200 -M have already been created by the language model creation unit 423.

  The acoustic model 431 is obtained by converting the probabilistic association between the feature amount obtained from the input voice by the mobile phone 200 and the phonetic symbol into data. The acoustic model 431 is created, for example, based on a reading voice database in which character strings of newspaper articles and voices when reading the newspaper articles are converted into data and stored.

  The dictionary 432 is obtained by converting phonetic symbols of words to be speech-recognized into data. The dictionary 432 is created based on, for example, the above-described reading voice database.

  The common language model 433 is obtained by converting general appearance probabilities and connection probabilities for each word described in the dictionary 432 into data. For example, the common language model 433 is created from text data in the above-described speech database. Here, the acoustic model 431, the dictionary 432, and the common language model 433 are described as being commonly used for all users, but of course, a configuration in which some or all of these are provided for each user. It is good.

  The user-specific language model 434 is a language model created by the user-specific language model creation unit 420 based on the received mail from the mobile phone 200, that is, based on the character string of the mail text included in the sent mail of the mobile phone 200. The user-specific language model 434 is created for each user of the mobile phone 200 as described above, and is associated with the ID information of the mobile phone 200, respectively.

  The voice recognition unit 440 performs predetermined voice recognition processing on the voice data received from the mobile phone 200 using the acoustic model 431, the dictionary 432, and the common language model 433. Then, the voice recognition unit 440 returns the text data created as the voice recognition result to the voice data transmission source. In addition, the voice recognition unit 440 receives voice data in association with the ID information of the mobile phone 200, and when there is a corresponding user-specific language model 434, the user-specific language model 434 is also used. The voice recognition unit 440 includes a voice data reception unit 441, a database switching unit 442, a sentence creation unit 443, and a text data transmission unit 444. The audio data receiving unit 441 includes an ID receiving unit 445.

  The voice data receiving unit 441 establishes a session with the mobile phone 200 in response to a request from the mobile phone 200 and receives the voice data sent from the mobile phone 200.

  The ID receiving unit 445 receives ID information transmitted from the mobile phone 200 when a session is established.

  The database switching unit 442 switches the user-specific language model 434 to be processed by the sentence creation unit 443. Specifically, the database switching unit 442 selects the language model for each user 434 corresponding to the ID information received by the ID receiving unit 445 from the speech recognition database 430 as a processing target of the sentence creating unit 443.

  The sentence creation unit 443 performs predetermined speech recognition processing on the speech data using the acoustic model 431, the dictionary 432, and the common language model 433, and generates text data. Specifically, the sentence creation unit 443 generates the likelihood of each phonetic symbol from the acoustic model 431, the word corresponding to the combination of each phonetic symbol from the dictionary 432, and N-grams from the common language model 433 and the user-specific language model 434. The occurrence probability of each word in the context is calculated. Then, the sentence creation unit 443 performs, for example, a process for searching for a word string that maximizes the product of the likelihood of the phonetic symbol and the word appearance probability, and the text data is extracted from the searched word string. Create

  In addition, when there is a user-specific language model 434 corresponding to the received ID information, the text creation unit 443 also refers to the user-specific language model 434. Specifically, the sentence creation unit 443 weights the common language model 433 and the user-specific language model 434, and adds the value obtained by multiplying each of the appearance probabilities by the weight to the appearance probability of each N-gram. Adopt as. Note that the text creation unit 443 takes into account the lack of statistics in the process of creating the user-specific language model 434, and determines the elapsed time since the user-specific language model 434 was created, the number of updates to the user-specific language model 434, and the like. Accordingly, the weighting may be changed.

  By using the user-specific language model 434 as described above, it is possible to more accurately reflect the user's context pattern and perform speech recognition, and improve the accuracy of speech recognition. For example, in the case of “Contact Ms. Oda” above, “Contact Ms. Oda” rather than “Contact Ms. Oda” or “Contact Ms. Oda” for the voice data “Odasan Nirenraku”. Will be generated. If the user has a friend named Mr. Oda and the number of times he / she is communicated to others by email as some kind of contact information, the voice “Odasan Nirenraku” is likely to mean “contact Mr. Oda”. That is, the user's desired character string can be selected with a higher probability, which indicates that the voice recognition accuracy has been improved.

  The text data transmission unit 444 packetizes the text data output from the text creation unit 443 and transmits it to the mobile phone 200 via the network interface unit 410.

  Although not shown, the speech recognition server 400 includes a CPU, a storage medium such as a hard disk drive (HDD) and a ROM that stores a control program, a working memory such as a RAM, and the like. By executing the control program by the CPU, the functions of the above-described units are realized.

  Hereinafter, operations of the mobile phone 200 and the voice recognition server 400 having the above-described configurations will be described.

  First, the operation of the mobile phone 200 will be described using a flowchart. Here, only operations related to mail creation will be described.

  FIG. 4 is a flowchart showing a flow of operations related to mail creation of the mobile phone 200.

  In step S <b> 1100, the control unit 212 determines whether or not the start of mail creation by voice recognition is instructed by the operation of the operation unit 204 or the like. If the start of mail creation by voice recognition is not instructed (S1100: NO), the process proceeds to step S1200. If the start of mail creation by voice recognition is instructed (S1100: YES), the process is performed. The process proceeds to step S1300.

  In step S1200, control unit 212 determines whether or not an instruction to start mail creation by a character input operation of normal operation unit 204 is given. If the start of normal mail creation is not instructed (S1200: NO), the process proceeds to step S2300 described later. If the start of normal mail creation is instructed (S1200: YES), the process is performed. The process proceeds to step S1400.

  In step S1300, the voice data transmission unit 208 establishes a session for voice recognition processing by starting communication with the voice recognition server 400 using the TCP / IP protocol, reads ID information from the ID storage unit 201, and performs voice recognition. Send to server 400.

  In step S <b> 1500, the control unit 212 starts audio input by the microphone 206 and causes the feature amount extraction unit 207 to input an audio signal output from the microphone 206.

  In step S1600, the feature amount extraction unit 207 extracts a feature amount by analyzing the speech signal, and outputs speech data that is time-series data of the feature amount.

  In step S <b> 1700, the voice data transmission unit 208 packetizes the voice data output from the feature amount extraction unit 207 and transmits the packetized data to the voice recognition server 400. For example, the voice data transmission unit 208 accumulates voice data, and transmits the voice data to the voice recognition server 400 in a batch when a voice input end operation is performed by the operation unit 204.

  In step S1800, the text data receiving unit 209 waits for reception of a speech recognition result from the speech recognition server 400. If the speech recognition result is not received (S1800: NO), the process proceeds to step S1900.

  In step S1900, the text data receiving unit 209 determines whether or not a predetermined time has elapsed after the audio data is transmitted from the audio data transmitting unit 208 and timed out. If not timed out (S1900: NO), the process returns to step S1800, and if timed out without receiving a voice recognition result (S1900: YES), the process proceeds to step S2300 described later. At this time, the mobile phone 200 may notify the user using the display 210 that an email cannot be created by voice recognition.

  If the text data receiving unit 209 receives the voice recognition result before the time-out (S1800: YES), the process proceeds to step S2000.

  On the other hand, in step S1400, since the start of normal mail creation has been instructed, the control unit 212 starts normal text data creation by a character input operation using the operation unit 204, and the process proceeds to step S2000.

  In step S2000, the control unit 212 causes the display 210 to display the text data of the speech recognition result or the text data that has been key-inputted on the display 210. At this time, the control unit 212 accepts editing of a character string by operating the operation unit 204 as necessary. When the character string is confirmed as the mail text, the destination mail address is specified, and mail transmission is instructed, the process proceeds to step S2100.

  In step S2100, the mail processing unit 205 prepares to send a language model creation mail. The language model creation e-mail is an e-mail for sending the character string included in the e-mail body of the transmission e-mail to the speech recognition server 400 as a material for creating the user-specific language model 434. The language model creation mail includes the same character string as the transmission mail, the destination is the voice recognition server 400, and includes ID information that uniquely identifies the sender.

  FIG. 5 is a diagram showing a configuration of a normal transmission mail and a configuration of a language model creation mail generated corresponding to the normal transmission mail.

  As shown in FIG. 5, in a normal transmission mail 610, the mail address designated by the user in TO is described as the destination, and the confirmed text is described as the mail body. Here, a case where nothing is described in CC (copy carbon) and BCC is shown as an additional destination. On the other hand, in the mail 620 including the language model creation mail, the BCC generation unit 213 describes the destination address of the language model creation mail (hereinafter referred to as “language model creation address”) in the BCC as an additional address. The

  The address for creating the language model describes the domain name of the voice recognition server 400 and the ID information of the mobile phone 200 in the domain part and account part of the mail address. Here, the ID information stored in the ID storage unit 201 is “01”, and the domain name of the speech recognition server 400 is “SRserver.ne.jp”. In other words, the language model creation address enables the mail body of the outgoing mail to be transmitted to the voice recognition server 400 in association with the ID information. In addition, since the language model creation address is designated in the BCC, the terminal “AAA@bbb.ne.jp”, which is the original destination of the sent mail, does not make the user aware of sending the language model creating mail.

  In step S2200 of FIG. 4, the mail processing unit 205 transmits the normal mail and the language model creation mail to the original destination and the voice recognition server 400 via the mail server 300, respectively.

  In step S2300, mobile phone 200 determines whether or not an instruction to end processing relating to mail creation is given by a user operation or the like. When the termination of the process is not instructed (S2300: NO), the mobile phone 200 returns to step S1100, and when instructed (S2300: YES), the series of processes is terminated.

  As described above, the mobile phone 200 transmits the voice data to the voice recognition server 400 in association with the ID information at the time of executing the voice recognition, and corresponds to the mail text included in the transmitted mail with the ID information at the time of mail transmission. At the same time, it is transmitted to the voice recognition server 400. Thereby, the character string created by the user can be used for the creation of the language model very easily, and the created language model can be used for the speech recognition processing.

  Further, the character string is transmitted by copy mail transmission, and the ID information is associated with the character string by describing the ID information in the account part of the destination address. As a result, the functions of existing facilities and application software can be utilized to the maximum, and the cost of system construction can be reduced.

  Further, the SIM identification information is easy to read, but normally cannot be changed by the user. The SIM is used as a storage medium storing information unique to the user, such as subscriber identification information, in place of the terminal in use. That is, the user often uses the same SIM continuously even after changing the model. Therefore, if SIM identification information is used as the ID information, user identification can be performed more easily and reliably, and the reliability of voice recognition accuracy is improved.

  Next, the operation of the voice recognition server 400 will be described using a flowchart.

  FIG. 6 is a flowchart showing the operation flow of the voice recognition server 400.

  In step S3100, the audio data receiving unit 441 determines whether a session with the mobile phone 200 has been established in response to a request from the mobile phone 200. If a session has not been established (S3100: NO), the process proceeds to step S3200. If the session is established (S3100: YES), the process proceeds to step S3300. Here, when the ID receiving unit 445 receives the ID information, the received ID information is output to the database switching unit 442.

  In step S3200, mail receiving unit 421 makes a mail acquisition request to mail server 300, receives mail addressed to itself, and determines whether a language model creation mail is received from mobile phone 200 or not. If a language model creation mail has been received (S3200: YES), the process proceeds to step S3400. If not received (S3200: NO), the process proceeds to step S4100 described later. Note that the user-specific language model creation process requires less immediacy than the voice recognition process, and therefore the process of step S3200 may be executed at predetermined intervals.

  In step S3300 and subsequent steps, voice recognition processing is executed. First, the voice recognition unit 440 determines a language model to be referred to by the text creation unit 443 according to the ID information acquired by the ID reception unit 445. That is, any user-specific language model 434 is selected using the ID information as an index. Note that the processing performed by the database switching unit 442 may be part of the processing performed when the text creation unit 443 refers to the speech recognition database 430.

  In step S3500, audio data receiving unit 441 determines whether audio data has been received from mobile phone 200 or not. If the voice data has not been received (S3500: NO), the process proceeds to step S3600. If the voice data has been received (S3500: YES), the process proceeds to step S3700, and the sentence of the received voice data Input to the creation unit 443 is started.

  In step S3600, audio data receiving unit 441 determines whether or not a predetermined time has elapsed without receiving audio data since a session with mobile phone 200 has been started, and time-out has occurred. If not timed out (S3600: NO), the process returns to step S3500, and if timed out without receiving audio data (S3600: YES), the process proceeds to step S4100 described later.

  In step S3700, the sentence creation unit 443 performs predetermined speech recognition processing with reference to the speech recognition database 430 to create text data. Then, the sentence creation unit 443 outputs the created text data to the text data transmission unit 444. At this time, when the database switching unit 442 determines to use any one of the user-specific language models 434, the sentence creation unit 443 also uses the corresponding user-specific language model 434.

  In step S3800, text data transmission unit 444 transmits the input text data as a voice recognition result to mobile phone 200 that is the transmission source of the voice data. Then, the process proceeds to step S4100 described later.

  On the other hand, in step S3400 and subsequent steps, language model creation processing is executed. The mail receiving unit 421 extracts text data and ID information of the mail body from the received language model creation mail, and outputs them to the language model creation unit 423 and the database switching unit 422, respectively.

  In step S3900, the database switching unit 422 determines the user-specific language model 434 to be created by the language model creation unit 423 using the input ID information as an index. The process performed by the database switching unit 422 may be part of the process performed when the language model creation unit 423 creates the user-specific language model 434.

  In step S4000, the language model creation unit 423 performs predetermined statistical information processing on the input text data, and updates or creates the user-specific language model 434 determined in step S3900 based on the statistical information processing result. To do.

  In step S4100, the voice recognition server 400 determines whether or not an instruction to end processing related to voice recognition is given by a user operation or the like. If the termination of the process is not instructed (S4100: NO), the voice recognition server 400 returns to step S3100, and if instructed (S4100: YES), the series of processes is terminated.

  As described above, the speech recognition server 400 creates the user-specific language model 434 associated with the ID information of the transmission source based on the mail text included in the language model creation mail received from the mobile phone 200. The voice recognition server 400 performs voice recognition on the voice data received from the mobile phone 200 with reference to the user-specific language model 434 using the ID information of the transmission source as an index. Thereby, the character string created by the user can be collected as a material for creating a voice model for the user by making the best use of the functions of existing facilities and application software. As a result, it is possible to create a language model for each user with the system construction cost suppressed.

  Hereinafter, the processing and communication flow of each device in the speech recognition system 100 will be described with an example.

  FIG. 7 is a sequence diagram showing an example of the processing and communication flow of each device in the speech recognition system 100. Here, for simplification of explanation, it is assumed that the speech recognition unit 440, the user-specific language model creation unit 420, and the speech recognition database 430 are handled separately.

  Each time the mobile phone 200 creates a mail text (S5100), the normal mail and a language model creation mail with ID information attached are sent to the mail server 300 (S5200), and the mail server 300 sends these mails. Store (S5300). In this state, when the user-specific language model creation unit 420 of the voice recognition server 400 makes a mail acquisition request to the mail server 300 (S5400), the mail server 300 describes the domain of the voice recognition server 400 in the destination address. The language model creation mail is returned to the user-specific language model creation unit 420 (S5500).

  The user-specific language model creation unit 420 switches the user-specific language model 434 based on the received ID information attached to the language model creation mail, and performs predetermined statistical information processing on the character string of the mail body (S5600). ), The voice recognition database 430 is updated (S5700). As a result, the user-specific language model 434 associated with the ID information of the mobile phone 200 is updated to reflect the context of the mail text created by the mobile phone 200 (S5800).

  Thereafter, when the mobile phone 200 establishes a session with the voice recognition unit 440 of the voice recognition server 400 (S5900) and transmits its own ID information to the voice recognition unit 440 (S6000), the mobile phone 200 receives voice input and The feature amount extraction from the voice data is started (S6100), and the voice recognition unit 440 switches the user-specific language model 434 based on the received ID information (S6200). Then, the voice recognition unit 440 receives voice data from the mobile phone 200 (S6300), refers to the updated voice recognition database 430 (S6400), and generates text data by a predetermined voice recognition process (S6500). . Then, the voice recognition unit 440 returns the generated text data to the mobile phone 200 (S6600).

  The mobile phone 200 displays the character string of the voice recognition result as a mail text candidate, accepts editing of the displayed character string (S6700), and transmits it to the mail server 300 by normal mail and language model creation mail (S6700). S6800). The processes in steps S6700 and S6800 correspond to the processes in steps S5100 and S5200 described above.

  As described above, in the speech recognition system 100, the mail text of the mail transmitted from the mobile phone 200 is also transmitted to the user-specific language model creating unit 420 in association with the ID information, and used for creating the user-specific language model. .

  As described above, according to the present embodiment, the mobile phone 200 associates the character string included in the outgoing mail with the ID information of the own device, and the language model creation unit 420 for each user of the speech recognition server 400. Send by email. In addition, the mobile phone 200 transmits the voice data to the voice recognition unit 440 of the voice recognition server in association with the ID information of the own device. The user-specific language model creation unit 420 learns the received character string and creates the user-specific language model 434 associated with the transmission source ID information. The voice recognition unit 440 performs voice recognition on the received voice data using the user-specific language model 434 associated with the transmission source ID information. Since the mail body of the outgoing mail is collected by e-mail, the character string created by the user can be used to create a language model very easily without changing the existing system. It can be used for recognition processing. That is, it is possible to easily improve the accuracy of speech recognition for different notations depending on the context.

  Note that the language model for each user and the type of speech recognition process are not limited to the above contents, but various language models created with contents corresponding to the context from a character string and various speech recognition processes using the same. Of course, it can be applied.

  In addition, the association of the ID information with the character string and the voice data is not performed by using the ID information directly, but by using other identification information such as an e-mail address associated with the ID information in advance. May be.

  In addition, the mobile phone 200 transmits the character string of the mail text to the voice recognition server 400 at once or at a timing designated by the user at a time, not every time a normal mail is transmitted. It may be. In this case, for example, the cellular phone 200 may accumulate transmission mail, accept selection from the user, edit the transmission destination of the selected mail, and create a language model creation mail. More specifically, for example, a language model creation address is written at the destination of TO or CC, and a language model creation mail is created in which the mail text of the selected outgoing mail is stored. As a result, it is possible to exclude a transmission mail whose context has been intentionally changed by the user from the learning target, and to further improve the accuracy of voice recognition.

  Further, even when the mobile phone 200 does not store the outgoing mail, the mobile phone 200 may copy the mail text and send the mail with the language model creation address as the TO or CC destination separately from the original outgoing mail. Good.

  Further, when the language model creation address is described as the TO destination, the mobile phone 200 may not include the original destination of the outgoing mail in the mail. Thereby, it is possible to prevent the original destination of the outgoing mail from leaking to the voice recognition server 400 side, and it is possible to protect the privacy of the user of the mobile phone 200 and the other party of the outgoing mail.

(Embodiment 2)
Next, a terminal device used in the distributed speech recognition system according to Embodiment 2 of the present invention will be described. The difference from the first embodiment is that the received mail is added to the language model creation mail to be created to further improve the accuracy of voice recognition.

  When emails are exchanged between family members and close friends on the same topic in a conversational manner, the frequency of reception from the other party and the probability of replying are high. In such mail communication, the context of the mail text is often similar to each other. If more character strings whose context is similar to the character string created by the user can be collected, the accuracy of speech recognition can be improved in a shorter time. Therefore, a case will be described in which received mail from another user with high reception frequency is added as language model creation mail.

  Mobile phone 200 according to Embodiment 2 has the same configuration as mobile phone 200 shown in FIG. 2 of Embodiment 1, for example. However, the mail processing unit 205 executes a received mail transfer process described below in addition to the process described in the first embodiment.

  In the received mail transfer process, the mail processing unit 205 records the source address of the received mail every time mail is received from another device. In addition, each time a mail is received from another device, the mail processing unit 205 determines whether or not the frequency of receiving the received mail from the other party is high, based on past recorded data, by comparison with a threshold value or the like. To do. When the mail reception frequency is high, the mail processing unit 205 creates a language model creation mail describing the character string of the mail text of the received mail from the transmission source, and the destination is the embodiment. The same language model creation address as 1 is specified. Thereby, a copy mail of the received mail is transferred to the voice recognition server 400.

  As described above, according to the present embodiment, not only the transmitted mail but also the mail text of the received mail having a similar context is transmitted to the voice recognition server 400 in association with the ID information. As a result, the user-specific language model creation unit 420 of the speech recognition server 400 can collect sufficient statistics in terms of the same topic and the same sentence expression in a shorter time than when only the outgoing mail is used. it can. That is, the accuracy of voice recognition can be improved in a shorter time.

  Note that received mail may also be selected in units of received mail and sent to the voice recognition server 400 in a batch. This makes it possible to exclude incoming emails whose recipients have intentionally changed the context, or incoming emails that have a high frequency of reception but have exceptionally different contexts, and further improve the accuracy of speech recognition. Improvements can be made.

  In addition, in consideration of mischievous mail and the like, it is also possible to use a destination with a high transmission frequency as a condition for transmitting the mail text.

  In addition, regarding the transmission of the mail text of its own mail to the speech recognition server 400 as a learning target, the mail text may be transmitted only to received mail from a partner who has obtained permission.

(Embodiment 3)
Next, a terminal device used in the distributed speech recognition system according to Embodiment 3 of the present invention will be described. The difference from the first embodiment is that destinations of outgoing mail are grouped and a language model for each user is created for each group.

  For example, in general, the word and style used are different for mail for family members, mail for friends, and mail for work-related people. That is, even in a mail transmitted by one user, the context of the mail text differs depending on the mail transmission partner. Therefore, it is possible to further improve the voice recognition accuracy by grouping destinations having similar email contexts and creating a user-specific language model 434 for each group. Therefore, a case will be described in which a user-specific language model 434 is created for each group of destinations having a similar context of outgoing mail.

  Mobile phone 200 according to Embodiment 3 has the same configuration as mobile phone 200 shown in FIG. 2 of Embodiment 1, for example. However, the mail processing unit 205 executes the partner grouping process described below in addition to the process described in the first embodiment. In addition to the ID information of the mobile phone 200, the mail processing unit 205 and the voice data transmission unit 208 transmit the group ID set for each group in the destination grouping process to the voice recognition server 400.

  The mobile phone 200 has a telephone directory in which telephone numbers and mail addresses are registered for each destination. The destinations registered in the telephone directory are allocated to groups prepared in advance, such as “family”, “friends”, and “company relations” for convenience of search and management.

  In the destination grouping process, the mail processing unit 205 treats the phone book group as a destination group having a similar context in the transmitted mail, and sets a group ID for each of the phone book groups. The group ID may be identification information assigned in advance to each group.

  When transmitting the language model creation mail, the mail processing unit 205 adds the group ID set to the group to which the normal mail transmission destination belongs to the language model creation mail together with the ID information of the own device. For example, the mail processing unit 205 adds a group ID to the account part of the language model creation address described in the first embodiment.

  Also, when establishing a session with the voice recognition server 400, the voice data transmission unit 208 transmits the group ID set to the group to which the normal mail transmission destination belongs to the voice recognition server 400 together with the ID information of the own device. . For example, the voice data transmission unit 208 transmits information obtained by adding the group ID to the ID information of the own device to the voice recognition server 400.

  In this case, the user-specific language model creation unit 420 of the speech recognition server 400 creates the user-specific language model 434 in association with information configured by a combination of ID information and a group ID. Further, the speech recognition unit 440 of the speech recognition server 400 refers to the user-specific language model 434 using information configured by a combination of ID information and group ID as an index.

  As described above, according to the present embodiment, a plurality of ID information is switched according to the transmission destination of the transmission mail and associated with the character string included in the transmission mail. Thereby, speech recognition can be performed in consideration of different contexts for each user and for each mail transmission destination, and the speech recognition accuracy in creating each mail can be improved.

(Embodiment 4)
Next, a language model creation device used in a distributed speech recognition system according to Embodiment 4 of the present invention will be described. The difference from the first embodiment is that, when the received language model creation mail includes a word whose reading is unknown (hereinafter referred to as “unknown word”), the reading of the unknown word is solved. It is that.

  Here, including unknown words, words are information that can be inserted as the body of an e-mail, such as characters, character strings, symbols, symbol strings, images, animations, etc. The concept includes all of the above.

  FIG. 8 is a block diagram showing the configuration of the speech recognition server according to the fourth embodiment of the present invention, and corresponds to FIG. 3 of the first embodiment. The same parts as those in FIG.

  As shown in FIG. 8, the speech recognition server 400a includes an unknown word processing unit 450a.

  The unknown word processing unit 450a solves the reading of unknown words included in the language model creation mail. The unknown word processing unit 450a includes an unknown word detection unit 451a, an inquiry mail transmission / reception unit 452a, and a dictionary registration unit 453a.

  The unknown word detection unit 451a inputs the mail body of the language model creation mail from the document extraction unit 424, and detects an unknown word in the mail body. Specifically, the unknown word detection unit 451a searches the dictionary 432 of the speech recognition database 430 for individual words included in the input mail text. And the unknown word detection part 451a detects the word which does not exist in the dictionary 432 as an unknown word.

  The inquiry mail transmission / reception unit 452a queries the unknown word detected by the unknown word detection unit 451a by mail to the user who transmitted the unknown word, and registers the query result in the dictionary 432. Specifically, the inquiry mail transmission / reception unit 452a creates an email (hereinafter referred to as “inquiry mail”) that indicates an unknown word and inquires about reading of the unknown word, and includes the unknown language. Sent to the sender of the creation email. When the inquiry mail transmission / reception unit 452a receives a mail describing an unknown word reading (hereinafter referred to as “response mail”) as a response to the inquiry mail, it extracts the unknown word reading from the response mail.

  The dictionary registration unit 453a registers the unknown word reading extracted by the inquiry mail transmission / reception unit 452a in the dictionary 432 in association with the unknown word.

  According to such a speech recognition server 400a, when an unknown word is included in the received language model creation mail, reading of the unknown word can be solved. Therefore, an appropriate word can be obtained as a voice recognition result when the corresponding reading voice data is received.

  Hereinafter, the operation of the voice recognition server 400a will be described.

  FIG. 9 is a flowchart showing an operation flow of the voice recognition server 400a, and corresponds to FIG. 6 of the first embodiment. The same steps as those in FIG. 6 are denoted by the same step numbers, and description thereof will be omitted.

  When the text data of the mail body is extracted from the language model creation mail by the document extraction unit 424 (S3400), the process proceeds to step S3810a. At this time, the document extraction unit 424 outputs the extracted text data and the e-mail address of the transmission source to the unknown word detection unit 451a of the unknown word processing unit 450a.

  In step S3810a, the unknown word detection unit 451a refers to the dictionary 432 and determines whether an unknown word exists in the text data input from the document extraction unit 424. If there is no unknown word (S3810a: NO), the process proceeds to step S3900. If an unknown word exists (S3810a: YES), the process proceeds to step S3820a. At this time, the unknown word detection unit 451a outputs the unknown word and the email address of the unknown word transmission source to the inquiry mail transmission / reception unit 452a.

  In step S3820a, the inquiry mail transmission / reception unit 452a transmits the inquiry mail with the mail address input from the unknown word detection unit 451a as a destination. At this time, it is desirable that the inquiry mail transmission / reception unit 452a sets an address for unknown word resolution (hereinafter referred to as “an unknown word resolution address”) different from the language model creation address as a transmission source address. This makes it easy to distinguish between language model creation mail and response mail. Then, the process proceeds to step S3900, where the user language model is updated based on the language model creation mail.

  On the other hand, if the mail receiving unit 421 has not received the language model creation mail (S3200: NO), the process proceeds to step S3210a.

  In step S3210a, inquiry mail transmission / reception unit 452a determines whether a response mail for a past inquiry mail has been received or not. If the inquiry mail transmission / reception unit 452a has not received a response mail (S3210a: NO), the process proceeds to step S4100. When the inquiry mail transmission / reception unit 452a receives the response mail (S3210a: YES), the process proceeds to step S3220a. At this time, the inquiry mail transmission / reception unit 452a extracts the unknown word and the reading of the unknown word described in the received response mail, and outputs them to the dictionary registration unit 453a.

  In step S3220a, the dictionary registration unit 453a registers the reading input from the inquiry mail transmission / reception unit 452a in the dictionary 432 in association with the unknown word input from the inquiry mail transmission / reception unit 452a.

  With such an operation, the voice recognition server 400a can solve reading of unknown words.

  FIG. 10 is a diagram showing an example of description contents of an inquiry mail and a response mail generated corresponding to the inquiry mail. Here, the case where the word “AMI” is detected as an unknown word and the user wishes to read “Ami” with respect to “AMI” is illustrated.

  As shown in FIG. 10, in the inquiry mail 630a, for example, an unknown word resolution address “02@SRserver.ne.jp” different from the language model creation address is described as a transmission source address. In the inquiry mail 630a, for example, an instruction “Please input AMI reading in the text and reply” is written as a subject (SUBJECT). The mobile phone 200 displays the description content of the received inquiry mail 630a.

  When the user follows the instruction of the inquiry mail 630a, “Ami” is described in the response mail 640a as the mail text.

  When the inquiry mail transmission / reception unit 452a receives the response mail 640a, the inquiry mail transmission / reception unit 452a extracts the unknown word described in the subject, that is, the part immediately before “Reading” and excluding “Re:” as an unknown word, The text described in the mail body is detected as an unknown word reading. As a result, a reading “ami” is registered in the dictionary 432 of the speech recognition server 400a in association with the word “AMI”. As a result, “AMI” can also be used to create a language model, and a voice recognition result “AMI” can be obtained when the user utters “Ami”.

  In consideration of the case where the unknown word is long and cannot be entered as the subject, reply with an inquiry e-mail, for example, “a text with“ AMI ”followed by“ AMI ”reading” Please indicate "Please" in the body of the email. In this case, “AMI Ami” is described as the mail body in the response mail. For example, the unknown word and the reading can be separated by searching the unknown word from the front of the mail text. Further, if a predetermined character or symbol such as “:” is inserted between the unknown word and the reading, the unknown word and the reading can be easily separated.

  Further, the destination of the inquiry mail and the unknown word that is the target of the inquiry may be associated with each other, and it may be determined based on this association to which unknown word the response mail is a response. . Thereby, even when unknown words are not correctly described in the response mail, unknown words and readings can be extracted.

  Thus, according to the present embodiment, when an unknown word whose reading is unknown exists in the learning target, the reading of the unknown word is resolved and registered in the dictionary. As a result, even words such as pictograms, emoticons, animations, etc. that are frequently used in e-mails but are not fixed in general reading can be input by voice. That is, the speech recognition rate can be improved when the user utters a word for which general reading is not fixed. In addition, since unknown words are resolved by sequential inquiries by e-mail, the reading can be resolved early after the unknown words are detected, and the accuracy of rapid speech recognition can be improved.

  Note that a background dictionary in which notation and reading are registered as a pair is prepared separately from the dictionary 432 for speech recognition, and the word is not registered in the dictionary 432 for speech recognition and is stored in the background dictionary. Inquiries may be made only when they are not registered. The background dictionary may be stored in the voice recognition server 400a, or may be stored in another server accessible from the voice recognition server 400a.

  Moreover, you may make it arrange | position some or all of the unknown word process parts 450a to the other apparatus on a network.

(Embodiment 5)
Next, a language model creation apparatus used for a distributed speech recognition system according to Embodiment 5 of the present invention will be described. Also in the present embodiment, unknown words are resolved in the same manner as in the fourth embodiment. The difference from the fourth embodiment is that the unknown words are read by a GUI (graphical user interface) prepared on the web. It is to solve it.

  FIG. 11 is a block diagram showing a configuration of the speech recognition server according to the fifth embodiment of the present invention, and corresponds to FIG. 8 of the fourth embodiment. The same parts as those in FIG. 8 are denoted by the same reference numerals, and description thereof will be omitted.

  As shown in FIG. 11, the speech recognition server 400b includes an unknown word processing unit 450b.

  The unknown word processing unit 450b solves the reading of unknown words included in the language model creation mail. The unknown word processing unit 450b includes an unknown word storage unit 454b and a GUI processing unit 455b instead of the inquiry mail transmission / reception unit 452a of the unknown word processing unit 450a of the fourth embodiment.

  The unknown word accumulation unit 454b accumulates the unknown word detected by the unknown word detection unit 451a in association with the ID information of the transmission source of the unknown word until the reading of the unknown word is resolved.

  The GUI processing unit 455b constructs a graphical user interface (hereinafter referred to as “unknown word registration site”) for a user to freely access on the web and register readings of unknown words. This unknown word registration site is constructed using, for example, a common gateway interface (CGI), and can be accessed from the mobile phone 200 via the communication network 500 using a hypertext transfer protocol (HTTP). In the unknown word registration site, the GUI processing unit 455b displays and displays the unknown words extracted from the language model creation mail created by the access source user among the unknown words stored in the unknown word storage unit 454b. Accepts readings for unknown words.

  The dictionary registration unit 453a registers the unknown word reading input at the unknown word registration site in the dictionary 432 in association with the unknown word.

  According to such a speech recognition server 400b, when an unknown word is included in the received language model creation mail, reading of the unknown word can be solved. Also, the unknown word reading can be registered at the timing desired by the user.

  Hereinafter, the operation of the voice recognition server 400b will be described.

  First, the operation of the unknown word registration site will be described, and then the overall operation of the voice recognition server 400b will be described.

  The unknown word registration site first obtains access source ID information as a user login process. This acquisition may be performed by prompting the user to input on the web screen, or may be performed by acquiring information such as a manufacturing number from the mobile phone 200. Then, the unknown word registration site extracts all unknown words associated with the acquired ID information from the unknown word storage unit 454b, and the extracted unknown words can be individually selected on the web screen. Display a list. Then, when any unknown word is selected, the unknown registration site transitions to a reading input screen for inputting an unknown word reading. When an unknown word registration site inputs a reading on the reading input screen and performs a decision operation such as clicking the decision button, the unknown word and the inputted reading are temporarily stored in pairs and a list of unknown words is stored. Return to the display screen. When a registration operation such as clicking a registration button or a logout process of the user is performed, the unknown word registration site acquires the input reading of the unknown word as a registration target.

  The unknown word registration site may be accessible from a terminal such as a personal computer instead of the mobile phone 200. Assuming access from the mobile phone 200, the unknown word registration site considers the visibility on a small display screen and the operability by a limited key switch as described above. It is desirable to switch. However, when an access from a terminal such as a personal computer is assumed, the unknown word registration site may be able to display a list of unknown words and input them on one screen.

  FIG. 12 is a flowchart showing an operation flow of the voice recognition server 400b, and corresponds to FIG. 9 of the fourth embodiment. The same steps as those in FIG. 9 are denoted by the same step numbers, and description thereof will be omitted.

  If there is an unknown word in the text data of the mail body extracted by the document extraction unit 424 (S3810a: YES), the process proceeds to step S3830b. At this time, the unknown word detection unit 451a outputs the unknown word and the ID information of the transmission source of the unknown word to the unknown word storage unit 454b.

  In step S3830b, the unknown word accumulation unit 454b stores the unknown word input from the unknown word detection unit 451a and the ID information in association with each other. Then, the process proceeds to step S3900.

  On the other hand, if the mail receiving unit 421 has not received the language model creation mail (S3200: NO), the process proceeds to step S3230b.

  In step S3230b, the GUI processing unit 455b determines whether or not an operation for registering an unknown word reading by the user has been performed at the unknown word registration site. If the unknown word reading registration operation has not been performed (S3230b: NO), the process proceeds to step S4100.

  On the other hand, when an unknown word reading registration operation is performed (S3230b: YES), the process proceeds to step S3240b. At this time, the GUI processing unit 455b extracts the unknown word for which the registration operation has been performed at the unknown word registration site and the reading of the unknown word, outputs the extracted unknown word to the dictionary registration unit 453a, and proceeds to step S3220a. This unknown word and reading are registered in the dictionary 432 in association with each other. Also, the GUI processing unit 455b deletes the unknown word output to the dictionary registration unit 453a from the unknown word storage unit 454b. As a result, it is possible to prevent an unknown word registration site from inquiring about a word whose reading has already been registered.

  With such an operation, the voice recognition server 400b can solve reading of unknown words.

  As described above, according to the present embodiment, when an unknown word whose reading is unknown exists in the learning target, the reading of the unknown word can be solved and registered in the dictionary, thereby improving the speech recognition rate. be able to. In addition, the user can perform input operations for reading a plurality of unknown words described in his / her mail at a time that is convenient for him / her. Thereby, the trouble of a user's unknown word input can be reduced.

  A part or all of the unknown word processing unit 450b may be arranged in another device on the network. In particular, if the unknown word storage unit 454b and the GUI processing unit 455b are arranged together in another device, the functions of the unknown word registration site can be separated and the burden on the speech recognition server 400b can be reduced. In addition, it is possible to speed up the processing of the unknown word registration site.

  Further, a user interface (IU: user interface) that displays a screen for performing the same operation as that of the unknown word registration site may be prepared in the mail application software of the mobile phone. In this case, for example, the speech recognition server transmits the detected unknown word sequentially or periodically to the mobile phone that is the source of the unknown word, and the mobile phone stores the received unknown word. Then, the application software displays the unknown word and accepts the input of the reading in the same manner as the unknown word registration site, sends the input reading to the speech recognition server in association with the unknown word, and stores it in the dictionary of the speech recognition server. Let me register. Thereby, the processing burden of unknown word resolution can be distributed.

  In addition, the mobile phone may transmit the reading of the word input by the user to the voice recognition server before being prompted to register the reading of the unknown word by the voice recognition server. Such reading is input to, for example, a user dictionary for character conversion, learning information of a Kana-Kanji conversion system, and a telephone directory stored in a mobile phone. Therefore, for example, every time these data are updated, the mobile phone transmits all the updated data or only the updated data portion to the voice recognition server. The voice recognition server that has received the data detects an unknown word from the received data, further obtains an unknown word reading from the received data, and associates the unknown word with the unknown word reading and registers them in the dictionary. This makes it possible to resolve the reading of unknown words before they are detected by the speech recognition server.

  Further, in each of the embodiments described above, the example in which the present invention is applied to a system for creating a mail text using a mobile phone using voice recognition has been described. However, the present invention is not limited to this. For example, in a system including a speech recognition device that performs speech recognition using a language model, and various terminal devices such as a personal computer and a PDA (personal digital assistant) that create a mail text using the speech recognition device Of course, it can be applied.

  Further, although an example in which the user-specific language model creation unit, the speech recognition database, and the speech recognition unit are arranged in the same server has been described, they may be arranged in separate devices on the network.

  A terminal device, a language model creation device, and a distributed speech recognition system according to the present invention are a terminal device, a language model creation device, and a terminal device that can easily improve speech recognition accuracy for different notations depending on context. It is useful as a distributed speech recognition system.

1 is a system configuration diagram showing an example of a configuration of a speech recognition system as a distributed speech recognition system according to Embodiment 1 of the present invention. FIG. 3 is a block diagram illustrating a configuration of a mobile phone including the terminal device according to the first embodiment. 1 is a block diagram showing a configuration of a speech recognition server including a language model creation device according to Embodiment 1. FIG. Flowchart showing an operation flow of the mobile phone according to the first embodiment. The figure which shows the structure of the mail for language model creation in Embodiment 1. The flowchart which shows the flow of operation | movement of the speech recognition server in Embodiment 1. Sequence diagram of voice recognition system according to Embodiment 1 The block diagram which shows the structure of the speech recognition server which concerns on Embodiment 4 of this invention. The flowchart which shows the flow of operation | movement of the speech recognition server in Embodiment 4 of this invention. The figure which shows an example of the description content of the inquiry mail and the response mail in Embodiment 4 of this invention The block diagram which shows the structure of the speech recognition server which concerns on Embodiment 5 of this invention. The flowchart which shows the flow of operation | movement of the speech recognition server in Embodiment 5 of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Speech recognition system 200 Mobile phone 201 ID memory | storage part 202 Radio | wireless part 203 Antenna part 204 Operation part 205 Mail processing part 206 Microphone 207 Feature-value extraction part 208 Voice data transmission part 209 Text data reception part 210 Display 212 Control part 213 BCC generation part 300 mail server 400, 400a, 400b speech recognition server 410 network interface unit 420 user-specific language model creation unit 421 mail reception unit 422 database switching unit 423 language model creation unit 424 document extraction unit 425 ID extraction unit 430 speech recognition database 431 acoustic model 432 Dictionary 433 Common language model 434 User-specific language model 440 Speech recognition unit 441 Speech data reception unit 442 Database switching unit 4 3 sentence creation unit 444 text data transmission section 445 ID receiving section 450a, 450b unknown word processing section 451a unknown word detection unit 452a inquiry mail transmitting and receiving unit 453a dictionary registration unit 454b unknown word storage unit 455b GUI processing unit

Claims (14)

Voice data transmitting means for transmitting voice data to a voice recognition device that performs voice recognition processing using a language model;
A mail transmission means for transmitting a mail body of a normal transmission mail as a language model creation mail to a language model creation device that creates the language model;
A terminal device. The mail sending means
Editing the address of the normal outgoing mail to generate the language model creation mail;
The terminal device according to claim 1. The mail sending means
ID information that is the same as or corresponding to the ID information used when using the speech recognition apparatus is described in a part of the mail, and the language model creation mail is generated.
The terminal device according to claim 2. The mail sending means
Each time an email not addressed to the language model creation device is sent, the language model creation device is added to the destination of the transmitted email.
The terminal device according to claim 2. The mail sending means
The email to be sent as the language model creation email is configured to be selectable in units of outgoing emails and sent in batches.
The terminal device according to claim 1. The mail sending means
The language model creation mail is generated using the language model creation address in which the domain name of the speech recognition device is described in the domain portion and the ID information used when using the speech recognition device is described in the account portion. ,
The terminal device according to claim 1. The mail sending means
Further, the mail body of the received mail is transmitted to the language model creation device as the language model creation mail.
The terminal device according to claim 1. The mail sending means
A plurality of ID information is switched according to the destination of the outgoing mail and associated with the character string.
The terminal device according to claim 4. A reading input means for accepting input of a word reading;
A reading transmission means for transmitting the input reading to the language model creation device as a reading of the word for creating the language model;
The terminal device according to claim 1. An unknown word acquisition means for acquiring an unknown word that is an unknown word in the language model creation device;
The reading input means receives the input of the reading by displaying the acquired unknown word,
The terminal device according to claim 9. A language model creation device that creates a language model used for speech recognition processing using a language model creation email received from a terminal device,
Mail receiving means for receiving the language model creating mail including the ID information and the mail text;
Mail processing means for extracting a mail text and ID information from the received language model creation mail;
Language model creation means for learning the extracted mail text and creating the language model for each ID information;
A language model creation device having Reading acquisition means for acquiring a word reading from the terminal device;
Dictionary registration means for registering the acquired reading in association with the word in the language model creation dictionary;
The language model creation device according to claim 11. An unknown word detecting means for detecting an unknown word which is a word whose reading is unknown from the extracted mail body;
The reading acquisition means includes
Accepting reading of the unknown word from the terminal device;
The language model creation device according to claim 12. A speech recognition device that performs speech recognition processing on speech data using a language model; a terminal device that transmits speech data to the speech recognition device; and a language model creation device that creates the language model by learning a character string. A distributed speech recognition system comprising:
The terminal device
Edit the destination of the normal outgoing mail to generate a language model creation mail, send it to the language model creation device,
The language model creation device includes:
Learn the email body of the received language model creation email to create the language model,
The voice recognition device
Performing speech recognition processing using the language model for the speech data received from the terminal device,
Distributed speech recognition system.

Images