JP2007163967A - Speech recognition device and speech recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve the problem that when the same pronunciation is allocated to objects different with users, it cannot discriminatingly be recognized since a conventional recognition device has the same recognition score for the same pronunciation. <P>SOLUTION: A speech recognition device is equipped with a pronunciation information registration means of registering a plurality of pieces of pronunciation information for an object to be recognized and a sound model registration means of relating sound models when the respective pieces of pronunciation information are used for speech recognition, and is configured to recognize speech of the respective pieces of pronunciation information by using the sound models related to the respective pieces of pronunciation information. Consequently, the object to which the sound models different with users are related although the pronunciation is the same can be recognized discriminatingly by the users. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a speech recognition apparatus and a speech recognition method that can give different pronunciations to the same recognition target for each user.

  Even if the words represent the same object, the words used vary depending on the vocabulary used daily and the relationship with the object. For example, when calling the same A child, the husband of A child calls “A child” and the daughter of A child calls “mother”. On the contrary, even if the same language is used, different objects may be indicated depending on the individual. For example, when A child calls “mother” and when A child's daughter calls “mother”, the called objects are different.

  When using speech recognition, it is convenient for the user if the vocabulary used for speech recognition can also use words familiar to each user. For example, if the family of the previous child A calls the mobile phone of A child with a voice dial using voice recognition from the home phone, the husband is “A child” and the daughter is “mother”. It is convenient if you can make a call with a familiar name.

A speech processing system capable of setting words used for speech recognition for each user as described above is disclosed in Patent Document 1. Patent Document 1 discloses a parent-child phone equipped with a voice recognition device as an example of a voice processing system. This parent-child phone can set a recognition vocabulary for each child device, that is, for each user who occupies the child device, in light of recent usage patterns in which family members occupy child devices. In the previous example, the child machine occupied by A child's husband calls "A child", and the child machine owned by A child's daughter calls "A mother" with the word "Mother". be able to. Document 1 is also configured so that voice recognition can be performed by terminals other than the terminal set with the public recognition vocabulary by disclosing the mutual recognition vocabulary between the slave unit and the master unit or between the slave unit and the slave unit. Therefore, the recognition vocabulary set for each user can be recognized even in terminals other than the handset occupied by the user. By exposing the recognition vocabulary of the child machine of A child's husband to the parent machine, the parent machine can also recognize “A child”. Similarly, if the recognition vocabulary of the child of A child's daughter is disclosed to the parent device, the recognition vocabulary used by the daughter of A child can also be recognized by the parent device.
Japanese Patent Laid-Open No. 2004-72274

  In the speech recognition apparatus disclosed in Document 1, a single speech recognition apparatus can recognize a recognition vocabulary set for each user by a plurality of users, but recognition vocabularies set by other users can also be recognized. Therefore, the recognition accuracy may be deteriorated as compared with the case where only the recognition vocabulary set by the user is used.

  In the speech recognition device disclosed in Document 1, a single speech recognition device can recognize a recognition vocabulary set by a plurality of users for each user. If you assign to different targets, it cannot be handled. For example, if the child A is registered as “Mom” for the phone number of A ’s mother and the daughter of “A” is registered as “Mom” as the cell phone number of A child, the voice recognition device Even if the input "mother" is correctly recognized, it is not possible to determine a single telephone number.

  An object of the present invention is to provide a speech recognition apparatus that solves the above-described problems and improves the convenience for the user.

  In order to solve this problem, the speech recognition apparatus according to the first aspect of the present invention uses pronunciation information registration means for registering a plurality of pronunciation information with respect to a recognition target, and is used when each pronunciation information is recognized by voice. Acoustic model registration means for associating an acoustic model, wherein each of the pronunciation information is recognized by using an acoustic model associated with each of the pronunciation information.

  Further, the speech recognition apparatus according to claim 2, pronunciation information registration means for registering a plurality of pronunciation information for a recognition target, user registration means for registering a user who can utter each of the pronunciation information, Use acoustic model registration means for registering an acoustic model to be used for each user, wherein each of the pronunciation information is voice-recognized using an acoustic model used by a user who can utter the pronunciation information. To do.

  According to the speech recognition apparatus of the present invention, arbitrary pronunciation information can be registered for a recognition target for each user, and an acoustic model to be used for each user can be associated with the registered pronunciation information. It becomes possible. As a result, when different acoustic models are used for each user, erroneous recognition of pronunciation information associated with acoustic models used by other users is reduced, and recognition accuracy is improved.

  Also, different objects can be recognized by the user even with the same pronunciation.

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

  FIG. 1 is a block diagram showing a schematic configuration of a speech recognition apparatus according to an embodiment of the present invention.

  In FIG. 1, 101 is a central processing unit (CPU), 102 is a control memory (ROM), 103 is a memory (RAM), 104 is an operation key such as a keyboard and buttons, 105 is a display device such as a liquid crystal, 106 is a microphone, etc. , An audio output device 107 such as a speaker, a communication device 108 for communicating with an external device, and a data bus 109.

  A control program for realizing the speech recognition apparatus of this embodiment and data used in the control program are recorded in a control memory (ROM) 102.

  These control programs and data are appropriately taken into the memory 103 through the data bus 108 under the control of the central processing unit 101 and executed by the central processing unit 101. The result of execution, that is, the result of speech recognition is displayed on the display device 105, or output from the speech output device 107 using speech synthesis, or output to an external device via a communication device.

  FIG. 2 is a functional block diagram when the speech recognition apparatus of the present invention is applied to a data search system for searching for telephone numbers.

  The data search system includes a database 201 that stores telephone numbers that are search target data and information related to each telephone number. An acoustic model 202 for storing an acoustic model used for speech recognition, a pronunciation information registration unit 203 for registering pronunciation information such as a reading pseudonym of a search keyword used for data search, and an acoustic model registration unit for registering an acoustic model used for speech recognition 204, a recognition grammar creation unit 205 that creates a recognition grammar used for speech recognition, a speech recognition unit 206 that performs speech recognition, and a recognition result process that retrieves data stored in a database and retrieves a telephone number that matches the speech recognition result Unit 207.

  FIG. 3 shows an example of the database 201 stored in the data search apparatus of this embodiment. The database of the present embodiment stores information on ID (301 column), telephone number (302 column), name (303 column), and name reading pseudonym (304 column) for each telephone number. . As shown in the column 304 in FIG. 3, hereinafter, in this embodiment, a case where a reading pseudonym is used as pronunciation information will be described.

  The acoustic model 202 of the present embodiment stores a plurality of acoustic models in advance. In this embodiment, an unspecified speaker model for unspecified speakers, an adult male model for adult men, an adult female model for adult women, and a child model for children are used. However, the present invention is not limited to this, and a user-specific model created using the voice of the user for each user using a speaker adaptation technique or the like may be used.

  Hereinafter, the details of the processing performed in each module will be described by dividing into processing for registering pronunciation information and acoustic model and processing for searching for data by voice recognition.

<Process to register pronunciation information and acoustic model>
FIG. 6 shows a flow of processing for registering pronunciation information and an acoustic model. Hereinafter, the process of registering the pronunciation information and the acoustic model will be described with reference to the flowchart of FIG.

  The user uses the operation key 104 to select data for registering pronunciation information in a predetermined procedure (S601). At this time, the pronunciation information registration unit 203 displays the GUI on the display device 105 as shown in FIG. The user registers the acoustic model used when recognizing desired pronunciation information and the pronunciation information to be registered for the selected data using the operation keys 104 (S602, S603). FIG. 4 shows Hanako's husband, “Hanako”, “Okasan”, Hanako-san, for Hanako Yamada's mobile phone number 099-9999-9999 recorded in the row 305 of the database shown in FIG. Shows a state in which pronunciation information and an acoustic model are registered in order to search for a keyword “Kaitai” and Hanako's child “Okasan”. In FIG. 4, a window 401 displays a name 401, a name reading pseudonym 402, and a telephone number 403. In this embodiment, it is assumed that four keyword reading pseudonyms other than the name reading pseudonym 402 can be registered in 408, 409, 410, and 411. For the registered pronunciation information, the acoustic model displayed in 404 to 407 is selected and registered by the acoustic model registration unit. In FIG. 4, the acoustic model indicated by the black circle is an acoustic model registered. That is, an adult male model is registered for “Hanako”, an adult female model for “Keitai”, an adult male model, and a child model for “Okasan”. In this embodiment, an unspecified speaker model is automatically registered for the name reading kana “Yamada Hanako”. However, as with other reading kana, the user registers the acoustic model. Alternatively, all acoustic models may be registered automatically.

  Next, the registered pronunciation information and the acoustic model to be used are stored in the database 201 (S604). At this time, the registered acoustic model is stored in association with each pronunciation information. The registered pronunciation information and acoustic model may be added and stored in the telephone number data stored in the database 201 together with the telephone number data, or may be stored in the database 201 as data different from the telephone number data. Good. An example of telephone number data when added to the telephone number data is shown in FIG. In the figure, an identification number of an acoustic model to be used is associated with each sound production information (column 501). In each identification number, 1 is an unspecified speaker model, 2 is an adult male model, 3 is an adult female model, and 4 is a child model.

  When registering pronunciation information and acoustic model for other data, the process returns to S601 for selecting data to be registered, and when registration is ended, registration processing of pronunciation information and acoustic model is ended (S605).

  This completes registration of pronunciation information and acoustic models.

<Process to search for data by voice recognition>
Next, the speech recognition process will be described.

  FIG. 7 shows a flow of processing for recognizing and retrieving data. Hereinafter, the flow of processing for recognizing voice and searching for data will be described with reference to the flowchart of FIG.

  When the data search process is started by a predetermined operation, the recognition grammar creation unit 205 first creates a recognition grammar with reference to the database 201 (S701). The recognition grammar describes at least a value output as a recognition result and pronunciation information. The recognition grammar is created so that the correspondence between each pronunciation information and the acoustic model used when recognizing each pronunciation information is recognized. Here, the acoustic model and the pronunciation information are associated by creating each acoustic model that uses the recognition grammar. FIG. 8 shows an example of a recognition grammar created. In the same figure, the pronunciation information described in the recognition grammars (A) to (D) includes (A) an acoustic model 1 (unspecified speaker model), (B) an acoustic model 2 (adult male model), respectively. (C) Acoustic model 3 (adult female model) and (D) acoustic model 4 (child model) are associated with each other. Also, the ID of the telephone number data indicated by 301 in FIG. 3 is output as the recognition result.

  When the user inputs a search keyword by voice through the voice input device 106, the voice recognition unit 206 recognizes the voice using the created recognition grammar (S703). In general speech recognition, for all recognition target words, a recognition score representing the similarity between the acoustic model corresponding to the pronunciation information of each recognition target word and the input speech is calculated, and the most similar recognition target word is recognized by speech recognition. As a result. The speech recognition in S703 also outputs the recognition result by the same processing. However, in calculating the recognition score, the acoustic model registered by the acoustic model registration unit 204 is used for each pronunciation information in accordance with the recognition grammar created by the recognition grammar creation unit 205. For example, when the recognition grammar shown in FIG. 8 is used, the recognition score of “Yama Hanako” shown in 801 is calculated using an unspecified speaker model, and the recognition score of “Okasan” shown in 802 is calculated using an adult male model. To do. A recognition score is obtained for all pronunciation information, and a recognition result having a recognition score most similar to the input speech is output. In this data retrieval system, the recognition grammar is described so that the ID of the telephone number data is output as the recognition result. For example, when “Kaitai” in 804 in FIG. 8 is most similar to the input speech, The output value “001” is output as the recognition result.

  The recognition result processing unit 207 searches for the telephone number data of the ID output as the recognition result from the database 201 (S704). The retrieved telephone number data may be displayed on the display device 105, or the telephone number may be output by voice from the voice output device 107 using voice synthesis or the like. If the data retrieval system is implemented on a telephone, the telephone and the communication person may be made via the communication device 108, and the telephone number registered in the retrieved telephone number data may be called.

  When the search is performed again, the process returns to the voice input of S702, and when the search is ended, the process is ended (S705).

<Effect>
As explained above, by registering arbitrary pronunciation information for each user and changing the acoustic model for calculating the recognition score for each pronunciation information, the recognition accuracy is higher than when the acoustic model is not changed for each pronunciation information. Has the effect of improving. When using a recognition score whose value increases as it is similar to the input speech, for example, when an adult male inputs speech, the recognition score of pronunciation information using the adult male model is high, and conversely, other than the adult male model The recognition score of pronunciation information using an acoustic model can be expected to be low. Thus, when an adult male inputs a voice, an effect of reducing misrecognition to pronunciation information not associated with an adult male model can be expected. For example, in the recognition grammar shown in FIG. 8, the pronunciations of “Hanako” and “Hanayo” are very similar, so when an adult male voice-inputs “Hanako”, only the unspecified speaker model is used. When the recognition score is obtained, the difference is considered to be small. However, according to the present embodiment, when the recognition score of “Hanayo” is calculated with the adult female model, the male voice that is the input speech and the adult female model mismatch, and the recognition score of “Hanayo” is This reduces the possibility of misrecognition. In the present embodiment, the case where an acoustic model prepared in advance is used has been described. However, if a dedicated acoustic model learned by the user's voice is used for each user, the above effect is further improved. Further, the voice recognition device of the present invention can recognize different objects (phone number data in this embodiment) depending on the user even with the same pronunciation. As indicated by reference numerals 803 and 804 in FIG. 8, in the speech recognition apparatus of the present invention, the same “Kaiai” is applied to the output values 002 (phone number of Hanako Yamada) and 001 (phone number of Taro Yamada) of the recognition result. "Is registered. When only one acoustic model is used, both have the same recognition score, so the recognition result cannot be narrowed down to only one by the recognition score. That is, it is not possible to distinguish the recognition result when Hanako Yamada inputs “Keitai” and when Taro Yamada inputs voice as “Keitai”. According to the speech recognition apparatus of the present invention, since the adult female model is associated with 002 “Keitai” and the adult male model is associated with 001 “Kaitai”, Taro Yamada is “Keitai”. , “001”, which calculates the recognition score using the adult male model, has a higher recognition score than “002”, which uses the adult female model, and can solve the above-described problems.

  In addition, since the speech recognition apparatus of the present invention limits the pronunciation information that is easy to be recognized due to the difference in the acoustic model, the speech recognition that narrows down the recognition vocabulary by designating the user in advance and using the user-specific recognition grammar The same effect as the device can be realized without specifying the user.

  In the first embodiment, the direct acoustic model used for each pronunciation information is selected. However, in this embodiment, the user is associated with the pronunciation information, and the relationship between the user and the acoustic model registered in advance is used. A method of associating an acoustic model with pronunciation information will be described.

  FIG. 9 shows a functional block diagram of the data search system of this embodiment.

  The acoustic model registration unit 204 having the configuration of the first embodiment is replaced with a user registration unit 209, and a user information registration unit 208 is further added.

<Process to register user information>
The user information registration unit 208 registers an acoustic model to be used for each user prior to registration of pronunciation information and stores it in the database 201. An example of user information stored in the database 201 is shown in FIG. In the figure, the column 1001 is the ID of the user, and the column 1002 is the identification number of the acoustic model used. In this embodiment, an adult male model, an adult female model, and a child model are used as in the first embodiment. However, the present invention is not limited to this, and the user's voice is used for each user using speaker adaptation technology or the like. Alternatively, a user-specific model created in this way may be used.

<Process to register pronunciation information and user>
Hereinafter, with respect to the process of registering pronunciation information and a user with reference to the flowchart of FIG. 13, only a part different from the process procedure of the first embodiment shown in FIG. 6 will be described.

  Instead of directly registering the acoustic model in the process of S603 performed by the acoustic model registration unit in FIG. 6, in this embodiment, a user who uses the registered pronunciation information is registered (FIG. 13, S606). This is shown in FIG. FIG. 11 shows a state in which the users who use the pronunciation information are registered in 1104-1106 for the pronunciation information “Hanako (1108)”, “Keitai (1109)”, “Okasan (1110)”. The user indicated by is a user who uses each pronunciation information. That is, “Hanako” is registered as user 1, “Keitai” is registered as user 2, and “Okasan” is registered as pronunciation information used by user 3. Further, in this embodiment, the user is not registered for the name reading Kana “Yamada Hanako”. However, the user may be registered in the same manner as other reading Kana, or automatically. All users may be registered.

  Next, the registered user is stored in the database 201 in association with the pronunciation information (S607). The registered pronunciation information and the user may add the phone number data to the phone number data stored in the database 201 together with the phone number data, or store it in the database 201 as data different from the phone number data. Good. An example of telephone number data when added to the telephone number data is shown in FIG. In the figure, the identification number of the user to be used is associated with each sound production information (column 1201). In each identification number, 0 is no specific user, 1 is user 1, 2 is user 2, and 3 is user 3.

<Process to search for data by voice recognition>
Next, the speech recognition process of this embodiment will be described.

  The processing flow for recognizing and retrieving data is the same as that of the first embodiment shown in FIG. This embodiment is different from the first embodiment only in the method for creating a recognition grammar in S701, and only this portion will be described.

  In the first embodiment, the acoustic model used for each pronunciation information is associated with the telephone number data stored in the database 201. However, in this embodiment, the user is associated with each pronunciation information. In order to associate the pronunciation information with the acoustic model to be used, in this embodiment, when creating the recognition grammar, the correspondence relationship between the user and the acoustic model to be used shown in FIG. The acoustic model used for each pronunciation information is associated. For example, since the user 1 and the user 3 are registered in 1106 in the pronunciation information “mother” shown in 1110 in FIG. 11, the acoustic model 2 used by the user 1 with reference to the user information (FIG. 10). The recognition grammar is created so that the (adult male model) and the acoustic model 3 (child model) used by the user 3 are associated with each other.

  The recognition grammar created in this way has the same form as the recognition grammar created in the first embodiment shown in FIG.

  Since the subsequent processing is the same as that of the first embodiment, description thereof is omitted.

<Other embodiments>
When the telephone number data is managed for each user, if the process of selecting a user in S606 is performed first, the process of registering the user for each pronunciation information can be omitted.

<Other embodiments>
A recognition grammar may be created for each user. In this case, the recognition grammar shown in FIG. 8 is the recognition grammar used by (A) no specific user (B) user 1 (C) user 2 (D) user 3. Also, when creating a recognition grammar for each user, when the speech recognition unit 206 recognizes the speech, the user information stored in the database is referred to and the pronunciation described in the recognition grammar of each user is recorded. The information recognition score is calculated using the corresponding acoustic model.

<Effect>
In the present embodiment, the user who uses the pronunciation information is associated with each pronunciation information. Therefore, the pronunciation information and the user can be registered more intuitively than the first embodiment in which the acoustic model is associated with the pronunciation information. The operability of registration work is improved.

  In addition, when telephone number data and pronunciation information are managed for each user, registration can be performed efficiently without registering an acoustic model or user for each pronunciation information.

  In the first and second embodiments, the case where the speech recognition of the present invention is applied to a data search system has been described. However, the present invention is not limited to this, and a function for registering pronunciation information is provided, and a plurality of users are assumed. The present invention can be applied to any voice recognition device or application.

  An object of the present invention is to supply a storage medium recording a program code of software that realizes the functions of the above-described embodiments to a system or apparatus, and the computer (or CPU or MPU) of the system or apparatus stores the storage medium. Needless to say, this can also be achieved by reading and executing the program code stored in.

  In this case, the program code itself read from the storage medium realizes the functions of the above-described embodiments, and the storage medium storing the program code constitutes the present invention.

  As a storage medium for supplying the program code, for example, a flexible disk, a hard disk, an optical disk, a magneto-optical disk, a CD-ROM, a CD-R, a magnetic tape, a nonvolatile memory card, a ROM, or the like can be used.

  Further, by executing the program code read by the computer, not only the functions of the above-described embodiments are realized, but also an OS (operating system) operating on the computer based on the instruction of the program code. It goes without saying that a case where the function of the above-described embodiment is realized by performing part or all of the actual processing and the processing is included.

  Further, after the program code read from the storage medium is written into a memory provided in a function expansion board inserted into the computer or a function expansion unit connected to the computer, the function expansion is performed based on the instruction of the program code. It goes without saying that the CPU or the like provided in the board or the function expansion unit performs part or all of the actual processing, and the functions of the above-described embodiments are realized by the processing.

The block diagram of the speech recognition apparatus of this invention. The module block diagram of the data search system which is one Embodiment of this invention. An example of data used in the data search system of Embodiment 1 An example of GUI which registers pronunciation information and an acoustic model in the data search system of Embodiment 1. FIG. 6 is an example of data in which pronunciation information and an acoustic model used in the data search system of Embodiment 1 are registered. Flowchart of processing for registering pronunciation information and acoustic model in data retrieval system of embodiment 1 Flowchart of speech recognition processing in the data search system of the first embodiment An example of a recognition grammar used in the data search system of Embodiment 1 Flowchart of processing for registering pronunciation information and user in data search system of embodiment 2 An example of user information in the data search system of Embodiment 2 Example of GUI for registering pronunciation information and user in data search system of embodiment 2 An example of pronunciation information used in the data search system of Embodiment 2 and data registered with a user Flowchart of processing for registering pronunciation information and user in data search system of embodiment 2

Explanation of symbols

101 Central Processing Unit 102 Control Memory 103 Memory 104 Operation Key 105 Display Device 106 Audio Input Device 107 Audio Output Device 108 Communication Device 109 Bus

Claims (2)

Pronunciation information registration means for registering a plurality of pronunciation information for a recognition target;
An acoustic model registration means for associating an acoustic model used when recognizing each pronunciation information with speech;
A speech recognition apparatus for recognizing each of the pronunciation information using an acoustic model associated with each of the pronunciation information. Pronunciation information registration means for registering a plurality of pronunciation information for a recognition target;
A user registration means for registering a user who can utter each pronunciation information;
Use acoustic model registration means for registering an acoustic model to be used for each user,
A speech recognition apparatus characterized by recognizing each of the pronunciation information using an acoustic model used by a user who can utter the pronunciation information.

Images