JP4254753B2 - Speaker recognition method - Google Patents

Description

  The present invention relates to a technique for recognizing an individual by voice.

Among the techniques for recognizing an individual using individual characteristics, there is a technique for recognizing an individual by voice (see, for example, Patent Document 1 and Non-Patent Document 1). When performing personal recognition using voice, first, feature quantities such as cepstrum are extracted from personal voice and registered. When recognizing, the degree of similarity between the registered feature amount of speech and the feature amount of speech uttered by the speaker is obtained, and whether the similarity is higher or lower than a certain threshold value. Thus, it is determined whether or not the speaker who uttered the voice is a registered person.
JP-A-9-127973 Tomoko Matsui, “Speaker Recognition by HMM”, Science Technology, The Institute of Electronics, Information and Communication Engineers, January 1996, p17-24

  By the way, when a person utters a voice, even if the same word is uttered several times, a difference may occur. Therefore, when extracting and registering a feature amount from a speaker's voice, the average of the speaker There is a possibility that a feature amount different from the feature amount is registered. If a feature amount that is different from the average feature amount of the speaker is registered in this way, there is a high possibility that the registered individual will not be recognized as a registered person even if he / she speaks. It becomes impossible to recognize an individual accurately.

  The present invention has been made under the above-described background, and an object of the present invention is to provide a technique for registering the original characteristics of a registrant's voice and reducing the possibility of recognition errors.

In order to solve the above-described problem, the present invention provides a feature quantity stored in a storage unit that stores a feature quantity of speech, and a feature quantity stored in the storage unit in association with an input identifier. A speaker recognition method for recognizing that a speaker of an input speech is a registrant when the distance from the feature amount of the input speech is equal to or less than a predetermined threshold, An identifier input step in which an identifier to be identified is input; a voice input step in which a voice of a sound generator is input a plurality of times; and a feature amount of the sound is determined for each voice input a plurality of times in the voice input step; A plurality of feature quantity storage step for storing the obtained plurality of feature quantities in the storage unit, and among the feature quantities stored in the storage unit by the plurality of feature quantity storage steps, the feature quantity is stored together by the plurality of feature quantity storage step. The An erasing step of erasing a feature quantity whose distance from the feature quantity is greater than or equal to a predetermined value from the storage unit, and a step after the erasing step, which is stored in the storage unit by the multiple feature quantity storage step An average feature amount calculating step for obtaining an average value of feature amounts, and a step after the erasing step, for each feature amount stored in the storage unit by the multiple feature amount storing step, the plural feature amount storing step The threshold value calculating step using the maximum value of the calculated distance as a threshold value, the identifier input in the identifier input step, and the average feature value the average value calculated by the calculating step, the associates the determined threshold value by the threshold calculating step, in the storage unit as a feature quantity of the sound's voice Providing speaker recognition method having a feature amount storing step of 憶.

  According to the present invention, the original features of the registrant's voice are stored, and the speaker is recognized using the stored feature amounts, so that the possibility of recognition errors being reduced.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
[Configuration of the embodiment]
FIG. 1 is a block diagram showing a hardware configuration of a main part of the speech collating apparatus according to the present embodiment. As shown in FIG. 1, each unit of the speech collating apparatus is connected to a bus 101, and exchanges data between each unit via this bus 101.

  The voice input unit 106 includes a microphone (not shown), and generates voice data representing the voice input to the microphone. The display unit 108 includes a display device such as a liquid crystal display device, and performs various displays such as characters and graphics under the control of the CPU 102. The information input unit 107 includes an input device such as a keyboard and a mouse (both not shown), and outputs a signal corresponding to the operation content to the CPU 102 in response to a key press or a mouse operation.

  The storage unit 105 includes, for example, a hard disk device (not shown) as a device for permanently storing data. In this hard disk device, a buffer area A1 and a registration area A2 are used as areas for storing various data. And are provided. The buffer area A1 is used as a temporary storage area for data used when the CPU 102 performs various processes. The registration area A2 is used as an area for storing data used when the CPU 102 performs processing for recognizing a speaker. FIG. 2 is a diagram illustrating a format of the registration table TB1 stored in the registration area A2. This registration table TB1 has a field called “identifier”, a field called “threshold”, and a field called “feature”. In these fields, data representing the voice feature amount is stored in the “feature amount” field, and an identifier for uniquely identifying the individual is stored in the “identifier” field. The “threshold” field stores a threshold used when determining whether or not the speaker is a registered individual.

  The ROM 103 stores a control program. The CPU 102 controls each unit according to a control program stored in the ROM 103. FIG. 3 is a functional block diagram showing a functional configuration of processing performed by the CPU 102. The units shown in FIG. 3 are realized by the CPU 102 executing the control program.

  Here, each functional block shown in FIG. 3 will be described. An identifier for uniquely identifying a speaker is input to the information input unit 107. This input identifier is sent to the information acquisition unit 50 when it is determined whether or not the speaker is a registered individual (hereinafter referred to as collation time), and the feature amount of the speaker's voice is determined. At the time of registration (hereinafter referred to as registration), it is sent to the information creation unit 40.

  The voice input to the voice input unit 106 is converted into voice data and sent to the utterance section extraction unit 10. Upon receiving the voice data, the utterance section extraction unit 10 extracts only the voice part of the speaker as the utterance section, removes the silent part and the sound part other than the voice, generates voice data, and outputs the voice data to the feature amount extraction unit 20. send. As will be described later, at the time of registration, the user pronounces the same word a plurality of times, but the utterance section of the word is extracted for each pronunciation and the voice data is sent to the feature amount extraction unit 20.

  The feature amount extraction unit 20 extracts a feature amount of the voice from the voice indicated by the transmitted voice data, and generates feature amount data V indicating the extracted feature amount. Note that the cepstrum is well known as the type of feature quantity, but it is not limited to the cepstrum and may be a feature quantity other than the cepstrum. The feature amount data V generated by the feature amount extraction unit 20 is stored in the buffer area A1 of the storage unit 105 for each pronunciation at the time of registration, and is sent to the feature amount comparison unit 60 at the time of matching.

The utterance selection unit 30 has feature data having a large distance from other feature data in the feature data V stored in the buffer area A1 so that the average feature data of the speaker can be obtained. Erase.
The information creating unit 40 obtains the average feature value data VA obtained by averaging the feature value data V stored in the buffer area A1, and obtains a threshold value t used for matching. Then, the identifier sent from the information input unit 107, the average feature amount data VA, and the obtained threshold value t are associated with each other and stored in the registration table TB1.

When the identifier is sent from the information input unit 107, the information acquisition unit 50 reads the threshold value t and the average feature value data VA stored in the registration table TB1 in association with the sent identifier, and reads the read threshold value t And the average feature amount data VA are sent to the feature amount comparison unit 60.
The feature amount comparison unit 60 obtains a distance between the feature amount indicated by the feature amount data V sent from the feature amount extraction unit 20 and the feature amount indicated by the average feature amount data VA sent from the information acquisition unit 50, and this distance Is larger or smaller than the threshold value t sent from the information acquisition unit 50, it is determined whether or not the speaker is a registered individual. Then, the feature amount comparison unit 60 sends result data indicating the comparison result to the display unit 108. The display unit 108 displays whether or not the speaker is a registered individual based on the result indicated by the transmitted result data.

[Operation of the embodiment]
Next, the operation of this embodiment will be described. In the following description, the operation during registration will be described first, and then the operation during verification will be described.

[Operation during registration]
First, a person who intends to register a voice feature amount (hereinafter referred to as a registrant) operates the information input unit 107 and clicks the “registration button” on the menu screen (see FIG. 8) displayed on the display unit 108. Perform a click operation. When an operation of clicking the registration button is performed, a screen (see FIG. 9) prompting the input of an identifier is displayed on the display unit 108 (FIG. 4: step S10). Thereafter, when an identifier for uniquely identifying the registrant is input and an operation of clicking the enter button displayed on the screen is performed (step S11; YES), a counter indicating the number of times the registrant has pronounced a predetermined word. n is initialized (n = 0) (step S12). Then, a screen (see FIG. 10) requesting to pronounce a predetermined word (for example, the name of the registrant) is displayed (step S13), and the input identifier is stored in the RAM 104.

  After the screen illustrated in FIG. 10 is displayed, the registrant pronounces a predetermined word. When voice uttered by the registrant is input to the voice input unit 106, voice data of the input voice is output from the voice input unit 106. When the voice data is output (step S14; YES), the CPU 102 extracts only the voice part as an utterance section, and generates voice data from which a silent part or a sound part other than the voice is removed (step S15). Then, the CPU 102 extracts the feature amount of the voice represented by the voice data generated in step S15, and generates feature amount data V indicating the extracted feature amount (step S16). Next, the CPU 102 adds 1 to the value of the counter n (step S17), and stores the generated feature amount data V in the buffer area A1 of the storage unit 105 (step S18). Note that when the feature amount data V is stored in the buffer area A1, it is stored in the array V [n] (n is the value of the counter n) as illustrated in FIG.

  When CPU 102 stores feature amount data V in buffer area A1, CPU 102 determines whether or not a registrant has pronounced a predetermined word a predetermined number of times (predetermined value N times) (step S19). The CPU 102 determines whether or not the value of the counter n is a predetermined value N. If the value of the counter n is less than the predetermined value N (step S19; NO), the “until registration is completed” on the screen of FIG. The numerical value of the “number of pronunciations to be performed” is changed to a predetermined value N−the value of the counter n and displayed, and the pronunciation of a predetermined word is requested. On the other hand, when the value of the counter n reaches the predetermined value N (step S19; YES), the following processing is performed.

Next, the CPU 102 obtains a distance from the feature quantity indicated by the other feature quantity data for each feature quantity indicated by the stored feature quantity data V, and performs processing for obtaining an average value of the obtained distances. First, the CPU 102 initializes the value of the counter i (i = 1) (FIG. 5: Step S20), and initializes the value of the distance D [i] (D [i] = 0) (Step S21). Next, the CPU 102 initializes the value of the counter j (j = 1) (step S22), and the feature quantity (i is the value of the counter i) indicated by the feature quantity data V [i] and the feature quantity data V [j] are obtained. The distance from the indicated feature quantity (j is the value of counter j) is obtained, and the obtained distance is added to the value stored in the distance D [i] (step S23).
When the process of step S23 ends, the CPU 102 increments the value of the counter j by 1 (step S24), and determines whether or not the value of the counter j has reached a predetermined value N (step S25). If the value of the counter j is not the predetermined value N (step S25; NO), the CPU 102 repeats the processes of steps S23 and S24 until the value of the counter j reaches the predetermined value N. On the other hand, when the value of the counter j becomes the predetermined value N (step S25; YES), the value of the distance D [i] is the predetermined value N that is the number of feature amount data stored in the buffer area A1. By dividing, an average value of the distance between the feature quantity indicated by the feature quantity data V [i] and the feature quantity indicated by the other feature quantity data is obtained, and the obtained value is stored in the distance D [i] (step S26).

  When the process of step S26 ends, the CPU 102 increments the value of the counter i by 1 (step S27), and determines whether or not the value of the counter i has reached a predetermined value N (step S28). When the value of the counter i is not the predetermined value N (step S28; NO), the CPU 102 repeats the processing from step S21 to step S27 until the value of the counter i reaches the predetermined value N.

On the other hand, when the value of the counter i becomes a predetermined value (step S28; YES), the CPU 102 first initializes the value of the counter i (i = 1) (FIG. 6: step S29). Then, the CPU 102 compares the distance average value D [i] obtained in the processes of steps S21 to S28 with the distance limit threshold T (step S30). Here, the limit threshold T is a predetermined fixed value.
When the average distance value D [i] is less than the limit threshold T (step S30: NO), the CPU 102 advances the process to step S33. On the other hand, when the average distance value D [i] is greater than or equal to the limit threshold T (step S30; YES), the CPU 102 deletes the feature amount data V [i] from the buffer area A1 (step S31).

  Next, the CPU 102 subtracts 1 from the counter n whose value has become N by the processing of steps S13 to S19, and sets the subtraction result as the value of the counter n (step S32). Then, the CPU 102 increments the value of the counter i by 1 (step S33), and determines whether or not the value of the counter i has reached a predetermined value N (step S34). When the value of the counter i is not the predetermined value N (step S34; NO), the CPU 102 repeats the processing from step S30 to step S33 until the value of the counter i reaches the predetermined value N.

  On the other hand, when the value of the counter i becomes the predetermined value N (step S34; YES), it is determined whether the value of the counter n is the same as the predetermined value N (step S35). Here, when the value of the counter n is not the same as the predetermined value N (when the feature value data V stored in the buffer area A1 is deleted in the processing of step S30 to step S34), the CPU 102 returns to FIG. As illustrated, after the feature amount data stored in the buffer area A1 is rearranged (step S36), the process flow is returned to step S13.

On the other hand, when the value of the counter n is the same as the predetermined value N (step S35; YES), the CPU 102 initializes the value of the threshold t (t = 0) (FIG. 7: step S37), and the counter i Is initialized (i = 1) (step S38). Next, the CPU 102 obtains data Va obtained by averaging feature amount data other than the feature amount data V [i] (step S39). Then, a distance D [i] between the feature amount indicated by the feature amount data V [i] and the feature amount indicated by the data Va is obtained (step S40), and it is determined whether the obtained distance D [i] is larger than the threshold value t (step S40). Step S41).
Here, when the value of the threshold value t is less than the value of the distance D [i] (step S41; YES), the CPU 102 sets the threshold value t = distance D [i] (step S42). On the other hand, when the value of the threshold value t is greater than or equal to the value of the distance D [i] (step S41; NO), the CPU 102 advances the process to step S43.

  Next, the CPU 102 increments the value of the counter i by 1 (step S43), and determines whether or not the value of the counter i has reached a predetermined value N (step S44). When the value of the counter i is not the predetermined value N (step S44; NO), the CPU 102 repeats the processing from step S39 to step S43 until the value of the counter i becomes the predetermined value. On the other hand, when the value of the counter i is the predetermined value N (step S44; YES), the CPU 102 obtains average feature value data VA that is an average value of the feature value data V stored in the buffer area A1 ( Step S45). Then, the identifier input from the information input unit 107, the average feature value data VA obtained in step S45, and the threshold value t obtained in the processing in steps S38 to S44 are stored in association in the registration table TB1 (step S46). When the CPU 102 stores the threshold value t and the average feature value data VA, if the identifier input from the information input unit 107 is already in the registration table TB1, the CPU 102 stores it in association with the already stored identifier. When the threshold value t and the average feature value data VA are updated to the newly obtained threshold value t and the average feature value data VA, and the identifier input from the information input unit 107 is not stored in the registration table TB1 Stores the input identifier, the threshold value t, and the average feature amount data VA in the registration table TB1.

  As described above, according to the present embodiment, the average feature amount of the registrant's voice is stored as the feature amount of the registrant's voice. In addition, a threshold used when determining whether or not a speaker is an individual registered in advance is obtained for each registrant based on the feature amount of the registrant's voice.

[Operation during verification]
Next, the operation at the time of collation will be described. First, the speaker operates the information input unit 107 and performs an operation of clicking the “speaker determination button” on the menu screen (see FIG. 8) displayed on the display unit 108. When an operation for clicking the speaker determination button is performed, a screen (see FIG. 9) prompting the input of an identifier is displayed on the display unit 108 (FIG. 13: step S50). Thereafter, when an identifier is input by the speaker and an operation of clicking the enter button displayed on the screen is performed (step S51), the input identifier is stored in the RAM 104.

  Thereafter, the CPU 102 searches the registration table TB1 for the identifier stored in the RAM 104 (step S52). Here, when the same identifier as the identifier stored in the RAM 104 is not found (step S53; NO), the CPU 102 causes the display unit 108 to display a message indicating that the identifier is not registered (step S54). The process ends. On the other hand, when the CPU 102 finds the same identifier as the identifier stored in the RAM 104 (step S53; YES), the threshold t and the average feature amount data VA stored in the registration table TB1 in association with the retrieved identifier. Are read (step S55). Then, a screen requesting pronunciation of a predetermined word (for example, the name of the registrant) is displayed (step S56), and the input of voice is waited (step S57).

  Thereafter, when the speaker pronounces a predetermined word and the voice uttered by the speaker is input to the voice input unit 106 (step S57; YES), the voice data of the input voice is output from the voice input unit 106. Is done. The CPU 102 extracts only the voice part from the voice data as an utterance section, and generates voice data from which a silent part and a sound part other than the voice are removed (step S58). Then, the CPU 102 extracts the feature amount of the voice represented by the voice data generated in step S58, and generates feature amount data V indicating the extracted feature amount (step S59).

  Next, the CPU 102 obtains a distance between the feature amount represented by the feature amount data V and the feature amount represented by the average feature amount data VA (step S60). When the obtained distance is equal to or less than the threshold value t read from the registration table TB1 (when the distance between the feature data V and the average feature data VA is close) (step S61; YES), the CPU 102 registers the speaker. The determination result is displayed on the display unit 108 (step S62). On the other hand, when the obtained distance is larger than the threshold value t read from the registration table TB1 (when the distance between the feature data V and the average feature data VA is far) (step S61; NO), the CPU 102 indicates that the speaker is a registered person. The determination result is displayed on the display unit 108 (step S63).

  As described above, according to the present embodiment, the speaker is recognized using the threshold value unique to the registrant based on the stored average feature amount of the registrant's voice. When a certain person pronounces, the possibility that the person is not registered is reduced.

[Modification]
As mentioned above, although embodiment of this invention was described, this invention is not limited to embodiment mentioned above, It can implement with another various form. For example, the present invention may be implemented by modifying the above-described embodiment as follows.

  A microphone may be connected to the personal computer device, and the above-described program may be executed on the personal computer device to recognize the speaker on the personal computer device. In addition to the personal computer device, the above-described program may be executed by a PDA (Personal Digital Assistance) equipped with a microphone, a mobile phone, or the like to recognize a speaker.

In the above-described embodiment, the threshold value t is calculated from only the voice feature amount, but a predetermined constant may be added to the threshold value t in accordance with the use environment of the voice matching device and the performance of the microphone.
In the above-described embodiment, the speech collation apparatus may calculate only the average feature amount data VA without calculating the threshold value t. In this case, the threshold value t may be input from the information input unit 107.
In the embodiment described above, the process of step S35 may be performed after step S34 without performing the process of step S35.

It is the block diagram which showed the hardware constitutions of the speech collation apparatus which concerns on embodiment of this invention. It is the figure which illustrated the format of registration table TB1. It is the functional block diagram which showed the function structure about the process which CPU102 performs. It is the flowchart which showed the flow of the process which CPU102 performs at the time of registration. It is the flowchart which showed the flow of the process which CPU102 performs at the time of registration. It is the flowchart which showed the flow of the process which CPU102 performs at the time of registration. It is the flowchart which showed the flow of the process which CPU102 performs at the time of registration. 6 is a diagram illustrating a screen displayed on the display unit 108. FIG. 6 is a diagram illustrating a screen displayed on the display unit 108. FIG. 6 is a diagram illustrating a screen displayed on the display unit 108. FIG. It is the figure which illustrated the image of the data memorize | stored in a buffer area. It is a figure for demonstrating rearrangement of the data memorize | stored in the buffer area | region. It is a flowchart which shows the flow of the process which CPU102 performs at the time of collation.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 ... Utterance section extraction part, 20 ... Feature-value extraction part, 30 ... Speech selection part, 40 ... Information preparation part, 50 ... Information acquisition part, 60 ... Feature-value comparison part 101 ... Bus, 102 ... CPU, 103 ... ROM, 104 ... RAM, 105 ... Storage unit, 106 ... Voice input unit, 107 ... Information input unit, 108. ..Display section.

Claims (1)

The distance between the feature quantity stored in the storage unit in association with the input identifier and the feature quantity stored in the storage unit storing the voice feature quantity and the input voice feature quantity is A speaker recognition method for recognizing that an input voice pronunciation person is a registrant when a predetermined threshold value or less,
An identifier input step in which an identifier for uniquely identifying a pronunciation is input;
A voice input step in which the voice of the pronunciation is input multiple times;
A feature amount storage step of obtaining a feature amount of the sound for each sound input a plurality of times in the sound input step, and storing the determined feature amounts in the storage unit;
Among the feature quantities stored in the storage unit by the multiple feature quantity storage step, a feature quantity whose distance from the other feature quantities stored together by the multiple feature quantity storage step is larger than a predetermined value is stored in the storage unit An erasure step to erase from,
An average feature amount calculating step for obtaining an average value of the feature amounts stored in the storage unit by the plurality of feature amount storing steps after the erasing step;
After the erasing step, for each feature quantity stored in the storage unit by the multiple feature quantity storage step, an average value of other multiple feature quantities stored together by the multiple feature quantity storage step and A threshold calculation step using the maximum value of the calculated distance as a threshold,
The identifier input in the identifier input step, the average value obtained in the average feature amount calculating step, and the threshold value obtained in the threshold value calculating step are associated with each other as the sound feature amount of the sound generator. A speaker recognition method comprising: a feature amount storage step stored in the storage unit.

Images