JP2010032865A - Speech recognizer, speech recognition system, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To recognize a speech, using either of a speaker adapted acoustic model or a versatile acoustic model, in response to a speech recognition request executed frequency, when adapting the acoustic model to a speaker in each group. <P>SOLUTION: A speech recognizer includes: a usage frequency extraction means 60 for extracting a user ID of a user whose speech recognition request executed frequency is a predetermined threshold value or more, from an access information database 32 for storing access information including a speech recognition request and the user ID; a high-frequency user grouping means 50 for grouping the users, based on the extracted user IDs; a speaker adaptation processing means 70 for performing speaker adaptation processing of the acoustic model for each group; and a speech recognition processing means 20 for recognizing the speech of the user corresponding to the user ID, using the acoustic model to which the speaker adaptation processing is performed for every group. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a technique applied to a speech recognition system used by an unspecified number of users.

  2. Description of the Related Art Conventionally, in a speech recognition system, a technique for automatically adapting an acoustic model using a character string as a recognition result and input speech data is known in order to improve recognition accuracy (Patent Document 1). . In addition, based on user profile information, a technique is also known in which users are divided into a plurality of groups and the acoustic model is speaker-adapted for each group (Patent Documents 2 and 3). In Patent Document 2, a plurality of speaker classes are set based on speech feature data obtained from a large number of unspecified speakers, and unspecified stories are based on the speech feature data of speakers belonging to each speaker class. An author codebook Then, a centroid vector sequence for each speaker adaptation word is calculated from the speech feature vector sequence of the speakers belonging to each speaker class and stored. The speech feature vector sequence of the input speaker and the barycentric vector sequence are associated with each other to determine to which speaker class the speech of the input speaker belongs, and the corresponding unspecified speaker codebook is selected. At the time of voice recognition, the voice of the input speaker is coded by the selected code book, and voice recognition is performed using the corresponding voice model.

  In Patent Document 3, an optimal output function is determined for each category using speech samples classified for each speaker characteristic, the overall output function is determined using the output function, and the hidden Markov model of each classification is determined. Speech recognition that matches the attributes of the speaker is performed from the output probability and the attribute probability of the speaker. In addition, the recognition based on the speaker classification is used to calculate the estimated score. Further, the calculation of the output probability and the calculation of the estimated score are speeded up using the fuzzy reasoning and the region segmentation method / variable separation method.

Moreover, in patent document 4, a recognition process is performed using the acoustic model corresponding to a group among several acoustic models.
JP 09-034485 A JP 09-258769 A JP-A-10-282986 JP 2002-182682 A

  However, when the speaker adaptation processing of the acoustic model is performed for each group as in the above-described prior art, if the number of users reaches several thousand, the processing load for grouping becomes very large. In this way, when the user scale is large, increasing the number of groups increases the number of speaker adaptation processes. Conversely, decreasing the number of groups results in an increase in the number of users belonging to the group, and speaker adaptation. Rather than expecting an improvement in recognition accuracy even if the acoustic model is used, there is a concern that the recognition accuracy may be lowered for a specific speaker.

  The present invention has been made in view of such circumstances, and when adapting an acoustic model for each group to a speaker, the acoustic model adapted for the speaker or a general-purpose one is used depending on the frequency of the voice recognition request. An object of the present invention is to provide a speech recognition apparatus, a speech recognition system, and a program that can perform speech recognition using any one of acoustic models.

  (1) In order to achieve the above object, the present invention takes the following measures. That is, the speech recognition device of the present invention is a speech recognition device that performs speech recognition using either a speaker-adapted acoustic model or a general-purpose acoustic model according to the frequency of requesting speech recognition. A use frequency extracting means for extracting a user ID of a user whose frequency of making a voice recognition request is equal to or higher than a predetermined threshold from an access information database storing access information including a voice recognition request and a user ID; High-frequency user grouping means for grouping users based on the extracted user ID, speaker adaptation processing means for performing speaker adaptation processing for acoustic models for each group, and speaker adaptation for each group Voice recognition processing means for performing voice recognition of the user corresponding to the user ID using the processed acoustic model. It is characterized.

  As described above, the user IDs of the users whose frequency of the voice recognition request is equal to or higher than a predetermined threshold are extracted, and the users are grouped based on the extracted user IDs, and the acoustic model for each group. Because the speaker adaptation processing is performed, even if the number of users reaches several thousand, the speaker adaptation of the acoustic model for realizing high speech recognition accuracy is performed while suppressing the number of times the speaker adaptation processing is performed. It becomes possible. As a result, the user can use a highly accurate speech recognition service while making a request for speech recognition. In addition, it becomes possible for the provider of the voice recognition service to enclose the user and make it regular so as to promote use and reuse the regular.

  (2) In the speech recognition apparatus of the present invention, the high-frequency user grouping means performs grouping based on user profile information.

  In this way, since grouping is performed based on user profile information, effective and efficient grouping can be performed.

  (3) Moreover, in the speech recognition apparatus of the present invention, the high-frequency grouping unit performs grouping by mapping the extracted user IDs to a group having the maximum speech recognition score obtained from the group-specific acoustic model. It is characterized by performing.

  As described above, grouping is performed by mapping the extracted user ID to the group having the maximum speech recognition score obtained from the acoustic model for each group, so that effective and efficient grouping can be performed. It becomes.

  (4) In the speech recognition apparatus of the present invention, the high-frequency grouping means performs grouping based on a distribution tendency of speech recognition scores obtained from a group-specific acoustic model.

  In this way, since grouping is performed based on the distribution tendency of the speech recognition scores obtained from the acoustic model for each group, effective and efficient grouping can be performed.

  (5) Further, in the voice recognition device of the present invention, the voice recognition processing means may perform general-purpose when there is a voice recognition request from a user whose frequency of making the voice recognition request is equal to or less than a predetermined threshold. It is characterized by performing speech recognition using an acoustic model.

  As described above, when a voice recognition request is made by a user whose frequency of making a voice recognition request is equal to or less than a predetermined threshold value, voice recognition is performed using a general-purpose acoustic model. It becomes possible to respond.

  (6) Moreover, the server apparatus of this invention is comprised from the server apparatus which has the speech recognition apparatus in any one of Claims 1-5, and at least 1 portable terminal device, The said portable terminal device The server apparatus makes a voice recognition request to the server apparatus, and the server apparatus performs voice recognition and transmits a voice recognition result to the portable terminal apparatus.

  With this configuration, a user ID of a user whose frequency of making a voice recognition request is equal to or greater than a predetermined threshold is extracted, and the users are grouped based on the extracted user ID, and an acoustic model for each group. Because the speaker adaptation processing is performed, even if the number of users reaches several thousand, the speaker adaptation of the acoustic model for realizing high speech recognition accuracy is performed while suppressing the number of times the speaker adaptation processing is performed. It becomes possible. As a result, the user can use a highly accurate speech recognition service while making a request for speech recognition. In addition, it becomes possible for the provider of the voice recognition service to enclose the user and make it regular so as to promote use and reuse the regular.

  (7) Further, the program of the present invention is a program for performing speech recognition using either a speaker-adapted acoustic model or a general-purpose acoustic model according to the frequency at which a speech recognition request is made, A process of extracting a user ID of a user whose frequency of making a voice recognition request is equal to or higher than a predetermined threshold from an access information database storing access information including a voice recognition request and a user ID; and the extracted user Based on the ID, a process for grouping users, a process for performing speaker adaptation processing for the acoustic model for each group, and a user ID using the acoustic model for which speaker adaptation processing has been performed for each group A series of processes including voice recognition processing for users corresponding to the computer, which can be read and executed by a computer It is characterized by a door.

  As described above, the user IDs of the users whose frequency of the voice recognition request is equal to or higher than a predetermined threshold are extracted, and the users are grouped based on the extracted user IDs, and the acoustic model for each group. Because the speaker adaptation processing is performed, even if the number of users reaches several thousand, the speaker adaptation of the acoustic model for realizing high speech recognition accuracy is performed while suppressing the number of times the speaker adaptation processing is performed. It becomes possible. As a result, the user can use a highly accurate speech recognition service while making a request for speech recognition. In addition, it becomes possible for the provider of the voice recognition service to enclose the user and make it regular so as to promote use and reuse the regular.

  According to the present invention, user IDs of users whose voice recognition request frequency is equal to or higher than a predetermined threshold are extracted, and the users are grouped based on the extracted user IDs. Speaker adaptation of the acoustic model to achieve high speech recognition accuracy while reducing the number of speaker adaptation processes, even when the number of users reaches several thousand because the speaker adaptation processing of the acoustic model is performed Can be performed. As a result, the user can use a highly accurate speech recognition service while making a request for speech recognition. In addition, it becomes possible for the provider of the voice recognition service to enclose the user and make it regular so as to promote use and reuse the regular.

  Next, embodiments according to the present invention will be described with reference to the drawings. FIG. 1 is a diagram illustrating the concept of the speech recognition apparatus according to the present embodiment. In this embodiment, a user whose voice recognition request (hereinafter referred to as “usage frequency”) is N times (N is a natural number) or more is assigned to an arbitrary group based on the usage frequency and the recognition score. Speech recognition is performed by using a group-specific acoustic model that is adapted for each group. This provides a highly accurate voice recognition service. As shown in FIG. 1, the group is divided into a frequently used group and a less frequently used group. For a group that is frequently used, speech recognition processing is performed using a speaker-adapted acoustic model. This makes it possible to provide a highly accurate voice recognition service for so-called regulars. On the other hand, for a group with low usage frequency, speech recognition processing is performed using a general-purpose acoustic model for unspecified speakers. This makes it possible to target a wide range of users.

  FIG. 2 is a block diagram showing a schematic configuration of the voice recognition system according to the present embodiment. The recognition process request receiving means 10 receives a voice recognition request and voice data from the user terminal, and instructs the voice recognition processing means 20 to perform a recognition process. In addition to returning the recognition result to the user terminal, the input voice, the recognition result, and the access information are stored in each DB. The group ID search means 11 acquires the group ID to which the user ID received together with the recognition request belongs from the user ID / group ID table.

  The speech recognition processing means 20 performs recognition processing using the general acoustic model 21, the group-specific acoustic model 22, and the language model 23 in accordance with the recognition request received from the recognition processing request receiving means 10. Then, the recognition result is returned to the recognition process request receiving means 10. In the recognition request, the group ID is also included, and recognition processing is performed using the acoustic model of the corresponding group. The stored voice DB 33 stores the input voice. The audio data may be spectrum data, cepstrum data, or VQ data in addition to audio data in the PCM format or the like.

  The recognition result DB 31 holds a recognition result character string and a recognition score. The recognition score may be further divided into acoustic likelihood and language probability. The access information profile information DB 32 holds the time information when the recognition is requested, the user ID, the voice data name stored in the corresponding accumulated voice DB 33, and the recognition result file name stored in the recognition result DB 31. The user ID / group ID table 40 is a mapping table of user IDs and assigned group IDs. After the grouping is performed by the high-frequency user grouping means 50 and the speaker adaptation model is updated, this table is updated.

  The usage frequency extraction unit 60 extracts the usage frequency of each user ID from the access information profile information DB 32 and transmits information on the user IDs used N times or more to the high frequency user grouping unit 50. The high-frequency user grouping means 50 groups users that have been used N times or more. As a grouping method, there are a method based on profile information and a method based on the maximum score of the speaker adaptation model or the distribution of scores. This grouping method will be described later.

  The speaker adaptation processing means 70 performs speaker adaptation processing of the acoustic model for each group using the general language model, the accumulated speech data, and the recognition result. The data acquisition means 71 acquires voice data and recognition results used for speaker adaptation from the accumulated voice DB 33 and the recognition result DB 31 and transmits them to the acoustic model processing means 72. The acoustic model processing means 72 performs speaker adaptation processing for the acoustic model and updates the completed acoustic model.

  The above is the schematic configuration of the voice recognition system according to the present embodiment, but the recognition processing request receiving means 10, the voice recognition processing means 20, the general-purpose acoustic model 21 and the language model 23 are mounted on the mobile terminal device, and the other configurations. Elements can be implemented in server devices.

  FIG. 3 is a flowchart showing a procedure for speaker adaptation. First, the usage frequency extraction means 60 calculates the usage frequency of the user who made the voice recognition request (step S1). Next, the high frequency user grouping means 50 groups users (step S2). Here, for example, users whose usage frequency is N times or more are divided into M groups. M is an arbitrary natural number of 1 or more, and the maximum value is the number of users whose usage frequency is N or more. Next, the speaker adaptation processing means 70 performs speaker adaptation processing for the acoustic model (step S3). Here, the data acquisition means 71 acquires the voice data of the user ID assigned to each group from the accumulated voice DB 33 and acquires the recognition result from the recognition result DB 31. Then, the acoustic model processing means 72 performs an adaptive process on the general-purpose speaker model for unspecified speakers. Note that this adaptive processing can use a conventionally known technique such as MLLR. Next, the speaker adaptation processing unit 70 updates the user ID / group ID mapping table via the acoustic model processing unit 72 (step S4).

  FIG. 4 is a flowchart showing the procedure of the voice recognition process. When the recognition process request receiving means 10 receives a recognition request from the user (step S11), the group ID search means 11 searches for a group ID (step S12). In this search, the user ID is used as a key. Next, the speech recognition processing means 20 performs speech recognition processing using the acoustic model corresponding to the group ID (step S13). Here, for example, 0 is assigned to the unspecified speaker as the group ID.

  FIG. 5 is a diagram showing a grouping method using an acoustic model performed by the high-frequency user grouping means 50. As shown in FIG. 5, the high-frequency user grouping unit 50 acquires all voices of a specific user ID and matches them with a group-specific acoustic model. For example, the average value of the recognition scores is calculated for the group-specific acoustic models whose group IDs are “1”, “2”,. Here, for example, when the group ID is “1”, the normalized score is 64.1, when the group ID is “2”, the normalized score is 39.4, and when the group ID is “M”, the normalized score is 64.1. Assume that the conversion score is 72.9. In this case, the group with the highest score is selected. Alternatively, grouping may be performed based on the score distribution. Note that the high-frequency user grouping means 50 can also group users who are frequently used based on profile information such as the user's age, sex, and origin.

  As described above, according to the speech recognition system according to the present embodiment, the user ID of a user whose frequency of requesting speech recognition is equal to or greater than a predetermined value (N) is extracted, and the extracted user Based on the ID, the users are grouped, and the speaker adaptation processing of the acoustic model is performed for each group. Therefore, even when the number of users reaches several thousand, the number of times of speaker adaptation processing is suppressed, It is possible to perform speaker adaptation of an acoustic model for realizing high speech recognition accuracy. As a result, the user can use a highly accurate speech recognition service while making a request for speech recognition. In addition, it becomes possible for the provider of the voice recognition service to enclose the user and make it regular so as to promote use and reuse the regular.

  The characteristic operation of this embodiment can be performed by causing a computer to execute a program. That is, the program according to the present embodiment is a program for performing speech recognition using either a speaker-adapted acoustic model or a general-purpose acoustic model in accordance with the frequency at which a speech recognition request is made. A process of extracting a user ID of a user whose frequency of making a voice recognition request is equal to or higher than a predetermined threshold from an access information database storing access information including a recognition request and a user ID, and the extracted user ID Based on the above, a process for grouping users, a process for performing speaker adaptation processing for acoustic groups for each group, and an acoustic model for which speaker adaptation processing has been performed for each group are used for the user ID. A command that allows a computer to read and execute a series of processes including voice recognition of a corresponding user It is characterized in that the.

  As a result, even when the number of users reaches several thousand, it is possible to perform speaker adaptation of an acoustic model for realizing high speech recognition accuracy while suppressing the number of times of speaker adaptation processing. As a result, the user can use a highly accurate speech recognition service while making a request for speech recognition. In addition, it becomes possible for the provider of the voice recognition service to enclose the user and make it regular so as to promote use and reuse the regular.

It is a figure which shows the concept of the speech recognition apparatus which concerns on this embodiment. It is a block diagram which shows schematic structure of the speech recognition system which concerns on this embodiment. It is a flowchart which shows the procedure of speaker adaptation. It is a flowchart which shows the procedure of a speech recognition process. It is a figure which shows the method of grouping using the acoustic model which the high frequency user grouping means 50 performs.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Recognition process request | requirement reception means 11 Group ID search means 20 Speech recognition processing means 21 General purpose acoustic model 22 Acoustic model according to group 23 Language model 40 User ID / group ID table 50 High frequency user grouping means 60 Usage frequency extraction means 70 Speaker Adaptive processing means 71 Data acquisition means 72 Acoustic model processing means

Claims (7)

A speech recognition device that performs speech recognition using either a speaker-adapted acoustic model or a general-purpose acoustic model according to the frequency at which a speech recognition request is made,
Usage frequency extraction means for extracting a user ID of a user whose frequency of making a voice recognition request is equal to or higher than a predetermined threshold from an access information database storing access information including a voice recognition request and a user ID;
A high-frequency user grouping means for grouping users based on the extracted user ID;
Speaker adaptation processing means for performing speaker adaptation processing of the acoustic model for each group;
A speech recognition apparatus comprising speech recognition processing means for performing speech recognition of a user corresponding to the user ID using an acoustic model on which speaker adaptation processing has been performed for each group.   The speech recognition apparatus according to claim 1, wherein the high-frequency user grouping means performs grouping based on user profile information.   The voice according to claim 1, wherein the high-frequency grouping means performs grouping by mapping the extracted user ID to a group having a maximum voice recognition score obtained from a group-specific acoustic model. Recognition device.   The speech recognition apparatus according to claim 1, wherein the high-frequency grouping unit performs grouping based on a distribution tendency of speech recognition scores obtained from a group-specific acoustic model.   The speech recognition processing means performs speech recognition using a general-purpose acoustic model when a speech recognition request is made by a user whose frequency of making the speech recognition request is equal to or less than a predetermined threshold. The speech recognition apparatus according to claim 1. A server device having the voice recognition device according to any one of claims 1 to 5;
At least one portable terminal device, and
A voice recognition system, wherein a request for voice recognition is made from the portable terminal device to the server device, the server device performs voice recognition and transmits a voice recognition result to the portable terminal device. A program for performing speech recognition using either a speaker-adapted acoustic model or a general-purpose acoustic model according to the frequency of the speech recognition request,
A process of extracting a user ID of a user whose frequency of making a voice recognition request is equal to or higher than a predetermined threshold from an access information database storing access information including a voice recognition request and a user ID;
A process of grouping users based on the extracted user ID;
Processing to perform speaker adaptation processing of the acoustic model for each group;
A series of processes including a process of performing speech recognition of a user corresponding to the user ID using an acoustic model subjected to speaker adaptation processing for each group, and converted into a command that can be read and executed by a computer A program characterized by

Images