JP2020154061A - Speaker identification apparatus, speaker identification method and program - Google Patents

Abstract

To shorten time required for identification even when there are many identification objects to be registered in speaker identification.SOLUTION: A speaker identification apparatus comprises: a storage section 15 for classifying registered data being voice feature data of all registered persons into groups clustered at similarity of registered data and storing the registered data in correspondence to identification codes given to the respective registered persons; a feature extraction section 11 for extracting collation data being the voice feature data from inputted voice data; a classification determination section 13 for determining into which group the collation data is to be classified; and a speaker identification section 16 for determining the registered person to whom the identification code is given as a speaker of the collation data, the identification code being corresponding to the registered data which has the highest similarity with the collation data and whose similarity with the collation data exceeds a threshold in the registered data classified into the group into which the collation data is to be classified.SELECTED DRAWING: Figure 1

Description

The present invention relates to a speaker identification device, a speaker identification method, and a program for identifying a speaker by collating the voice data of the speaker with the registered voice data.

As a method of authenticating an individual, a voice authentication technique for authenticating a speaker by collating a voice uttered by a speaker with a registered voice has been put into practical use. For example, Patent Document 1 and Patent Document 2 describe a personal authentication system by voice.

The personal authentication system of Patent Document 1 is a personal authentication system that identifies an individual by voice input using a computer. (1) The words to be authenticated are input by voice, and the input voice is voice-analyzed. Means for creating encrypted voice data for authentication, (2) The coded voice data for authentication is classified by words by extracting words by voice recognition analysis using an acoustic model, a language model and a word dictionary. A means for extracting only the data corresponding to the word from the voice pattern database and collating the voice pattern with the encoded voice data for recognition is provided.

The personal authentication system of Patent Document 2 is a personal authentication system that identifies an individual by having the person to be authenticated input voice by voice. (1) A plurality of words to be authenticated and a target of authentication registered in advance in the voice print database. A plurality of words including the non-authentication word are presented to the authenticated person, (2) the voice data input by the authenticated person is taken in, the authentication process is performed on the authentication target word, and the authentication target is not subject to the authentication. A means for registering a word in the voice print database is provided.

Patent Document 3 describes a technique for improving the determination accuracy of the speaker. The speaker determination device of Patent Document 3 has a speaker feature amount of each divided voice section divided into a predetermined time length of the voice section of the voice signal, and a speaker feature amount generated in advance for each person in charge of the counter. A similarity calculation unit that calculates the similarity, a speaker primary determination unit that generates primary determination information representing the speaker ID of each divided audio section from the similarity, and a predetermined number before or after an arbitrary divided audio section. When the degree of similarity between the speaker characteristic amount of the neighboring speaker, which is the most applicable speaker in the divided voice section of, and the speaker characteristic amount of the arbitrary divided voice section satisfies a predetermined condition, the speaker of the neighboring speaker A speaker secondary determination unit that generates secondary determination information by using an ID as secondary determination information for an arbitrary divided audio section, and a story of a divided audio section corresponding to the secondary determination information indicating that the customer is a customer. It includes a speaker clustering unit that clusters a person feature quantity, that is, a set of customer speaker feature quantities to generate a customer speaker ID, and generates tertiary determination information.

Japanese Unexamined Patent Publication No. 2003-323197 Japanese Unexamined Patent Publication No. 2003-302999 JP-A-2019-8131

In the personal authentication by voice, the individual is basically identified by comparing all the registered voices with the collated voices, as in the speaker determination device of Patent Document 3. As the number of registered speakers increases, it takes time to obtain a result in proportion to the number of registrations for determining the similarity. In the personal authentication system of Patent Document 1, comparison is made only with the data corresponding to the spoken word, but if the registered data of the same word increases, the same problem occurs.

In the personal authentication system of Patent Document 2, it is premised that the ID of the person to be authenticated is specified by a credit card or the like and voice authentication is used instead of the password, and the registration data for determining the similarity is limited by the ID. ing. When identifying the speaker by voice including the ID of the person to be authenticated, it is necessary to compare with all the registered data.

The present invention has been made in view of the above circumstances, and an object of the present invention is to shorten the time required for identification even when there are many identification objects registered in speaker identification.

The speaker identification device according to the first aspect of the present invention classifies the registration data, which is the voice feature data of each registered person, into a group clustered according to the similarity between the registered data, and the registered data. A storage unit that stores the identification code assigned to each person in association with the identification code, a feature extraction unit that extracts collation data that is audio feature data from the input audio data, and the collation data are classified into any of the above groups. Among the registered data classified into the group to which the collation data should be classified, the classification determination unit for determining whether or not the collation data should have the highest degree of similarity to the collation data and the collation data. It is provided with a speaker identification unit that determines that the registered person to whom the identification code associated with the registration data whose similarity exceeds the threshold is assigned as the speaker of the matching data.

According to the present invention, it is determined to which of the groups in which the registered data is clustered in advance the collation data should be classified, and the registered data to be determined for similarity is limited to the group to be classified. It is possible to shorten the time required for identification as compared with the case of comparing all.

Preferably, the storage unit stores a phonetic element string, a syllable string, or a character string indicating the vocal content of each of the registered data in association with the registered data, and the speaker identification device stores the input voice data. Further, a voice recognition unit for extracting a phonetic element string, a syllable string, or a character string indicating the voice content is further provided, and the speaker identification unit is classified into the group to which the collation data should be classified, and the voice content is Among the registered data having the same voice content as the collated data, the identification associated with the registered data having the highest similarity with the collated data and having a similarity with the collated data exceeding the threshold value. The registered person to which the code is given is determined to be the speaker of the collation data.

In that case, in the group to which the collation data should be classified, the comparison target is limited to the registered data having the same utterance content as the collation data, so that the time required for identification can be further shortened. Moreover, since only the registered data having the same utterance content is compared, the possibility of erroneous recognition can be reduced.

Preferably, the classification determination unit includes a trained model of a neural network in which the voice feature data of the registered person classified into the group by the clustering is machine-learned as training data.

When the classification is determined by the trained model of the neural network, it is possible to suppress the increase in the time required for identification even if the number of groups in which the registered data is classified increases.

The speaker identification method according to the second aspect of the present invention includes a feature extraction step of extracting matching data which is voice feature data from input voice data, and the matching data is the voice feature of each registered person. A classification determination step for determining which group the registered data, which is data, should be classified into a group in which the registered data is clustered according to the similarity between the registered data, and the registered data are classified into the clustered group. Among the registration data stored in association with the identification code given to each registered person, among the registration data classified into the group to which the collation data should be classified, the collation data The registered person to whom the identification code associated with the registered data having the highest degree of similarity and the degree of similarity with the collated data exceeds the threshold is referred to as the speaker of the collated data. It includes a speaker identification step for determining.

In the program according to the third aspect of the present invention, the computer classifies the registration data, which is the voice feature data of all the registered persons, into a group clustered according to the similarity between the registered data, and the registered data. A storage unit that stores the identification code assigned to each person in association with the identification code, a feature extraction unit that extracts matching data that is voice feature data to be identified from the input voice data, and the matching data is classified into any of the above groups. Among the classification determination unit that determines whether the data should be collated and the registered data that are classified into the group to which the collation data should be classified, the collation data has the highest degree of similarity to the collation data. The registered person corresponding to the identification code of the registered data whose similarity with the data exceeds the threshold value is made to function as a speaker identification unit for determining the speaker of the collation data.

According to the present invention, even when there are many identification targets registered in speaker identification, the time required for identification can be shortened.

A block diagram showing a configuration of a speaker identification device according to an embodiment of the present invention. Block diagram showing the configuration of the voice registration device according to the embodiment The figure which shows the example of the voice feature database which concerns on embodiment The figure which shows the example of the neural network of the classification determination part which concerns on embodiment A flowchart showing an example of the speaker identification operation according to the embodiment. Block diagram of the speaker identification device according to the modified example of the embodiment Flow chart showing an example of speaker identification operation according to a modified example A block diagram showing an example of the hardware configuration of the speaker identification device according to the embodiment.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. The same or corresponding parts in the figure are designated by the same reference numerals.

Embodiment.
FIG. 1 is a block diagram showing a configuration of a speaker identification device according to an embodiment of the present invention. The speaker identification device 1 compares the collation data, which is the voice feature data extracted from the voice input from the microphone 21, with the registered data, which is the voice feature data stored in the storage unit 15, and the speaker identifies the speaker. Identify which of the registered persons. The speaker identification device 1 includes a voice input unit 10, a feature extraction unit 11, a voice recognition unit 12, a classification determination unit 13, a data extraction unit 14, a storage unit 15, and a speaker identification unit 16. The registered data is classified into a group clustered according to the similarity between the registered data, and is stored in the storage unit 15 as a voice feature database associated with the identification code assigned to each registered person. The registered data is associated with a phoneme string, a syllable string, or a character string indicating the content of each utterance.

(Voice registration)
Prior to speaker identification, the storage unit 15 of the speaker identification device 1 stores registration data, which is voice feature data of the registered person, by the voice registration device in advance. FIG. 2 is a block diagram showing a configuration of a voice registration device according to an embodiment. The voice registration device 2 includes a voice input unit 10, a feature extraction unit 11, a voice recognition unit 12, a clustering unit 17, an input unit 18, a data registration unit 19, and a storage unit 15. The same device as the speaker identification device 1 may be used in common for the voice registration device 2. The voice input unit 10, the feature extraction unit 11, the voice recognition unit 12, and the storage unit 15 are the same for the speaker identification device 1 and the voice registration device 2, and the speaker identification device 1 and the voice registration device 2 are the same devices. In the case of, they are common.

The audio input unit 10 samples the audio signal input from the microphone 21 by the registered person at a predetermined frequency, performs A-D conversion, and generates audio data. The voice input unit 10 sends voice data to the feature extraction unit 11 and the voice recognition unit 12.

The feature extraction unit 11 extracts the registered data, which is the voice feature data, from the voice data. The voice feature data is, for example, a speaker information representation based on a GMM super vector (Gaussian Mixture Model Supervector), an i-vector, or a tensor decomposition.

The GMM Supervector (GMM-SV) is a feature quantity in which voice is modeled by a Gaussian Mixture Model (GMM) and the average vector of each Gaussian distribution constituting the GMM is connected in a row. GMM is a probability distribution represented by a weighted linear sum of multiple Gaussian distributions (WM Campbell, DE Sturim, and DA Reynoldes, "Support Vector Machines using GMM Supervectors for Speaker Verification," IEEE Signal Processing Letters, vol. 13, pp. 308-311, 2006.).

The i-vector is a feature quantity obtained by dimensionally compressing a GMM super vector (GMM-SV) based on factor analysis. M, which is a GMM-SV extracted from one utterance, is m, which is a GMM-SV of a universal background model (UBM) that does not depend on the speaker and language, and the utterance content, speaker, and recording environment. Using the projection matrix T to the low-dimensional space that models the variation of speech due to changes, it is decomposed as M = m + Tw. This w is an i-vector (N. Dehak, PJ Kenny, R. Dehak, P. Dumouchel, and P. Ouellet, "Front-End Factor Analysis for Speaker Verification," IEEE Transactions on Audio, Speech, and Language Processing. , vol. 19, no. 4, pp. 788-798, 2011.).

The speaker information representation based on the tensor decomposition is a speaker information representation based on the tensor decomposition, which is an extension of the i-vector approach. One-speech is expressed by a matrix in which rows and columns correspond to each distribution and mean vector of GMM, respectively, a matrix for many speakers is treated as a tensor, and speaker information is expressed by introducing tensor analysis ( Chin Tuan Tou, Daisuke Saito, Nobuaki Minematsu, Keikichi Hirose, "Examination of Speaker Identification Using Speaker Information Expression Based on Tensor Decomposition," Proceedings of the Spring Lectures of the Acoustical Society of Japan, pp. 217-220, 2005. ).

Here, a case where i-vector is used as the voice feature data will be described as an example. The feature extraction unit 11 sends the extracted registered data, which is an i-vector, to the clustering unit 17 and the data registration unit 19.

The voice recognition unit 12 extracts a voice feature from the voice data, searches for a utterance content that is closest to the voice feature, and extracts a phoneme string, a syllable string, or a character string indicating the utterance content. The voice recognition unit 12 sends the obtained phoneme string, syllable string or character string to the data registration unit 19.

The input unit 18 receives the input of the identification code for identifying the registered person who has input the voice from the microphone 21, and sends the input identification code to the data registration unit 19. The input unit 18 includes, for example, a pointing device such as a keyboard, a mouse or a touch panel, a display, an IC card reader, a bar code reader, or a two-dimensional code reader. The identification code is a code input by keyboard operation, selection by a pointing device, or an IC card, bar code, or two-dimensional code. The input unit 18 reads biometric information such as fingerprints, glows, and veins of fingers or hands, collates it with the biometric information registered in advance, and sends the referenced identification code of the registered person to the data registration unit 19. You may.

The data registration unit 19 includes registration data which is voice feature data sent from the feature extraction unit 11, an identification code of the registered person sent from the input unit 18, and phoneme strings and syllables sent from the voice recognition unit 12. A column or a character string is associated with each other and stored in the storage unit 15.

After storing the voice feature data of all the registered persons in the storage unit 15, the clustering unit 17 clusters the registered data, which is the voice feature data of all the registered persons, with the similarity between the registered data. Classify into groups. The similarity is, for example, a cosine similarity or a log-likelihood ratio. For clustering, for example, the k-means ++ method, the k-means method, or Ward's method can be used.

The clustering unit 17 sends the number of the group in which each registered data is classified to the data registration unit 19. The data registration unit 19 associates each registered data with an assigned group number and stores it in the storage unit 15. That is, the registered data is classified into a group clustered according to the similarity between the registered data, and is stored in association with the identification code assigned to each registered person and the utterance content of the registered data.

FIG. 3 is a diagram showing an example of a voice feature database according to the embodiment. The voice feature database is stored in the storage unit 15. The record of the voice feature database is composed of a registrant ID, a group ID, a voice content, and voice feature data. The registrant ID is an identification code of the registered person. The group ID is the number of the group in which each registered data is classified. The utterance content is a phoneme string, a syllable string, or a character string, which is the utterance content of the registered data. The voice feature data column may be the voice feature data itself, or may be a pointer to a file or the like in which the voice feature data is stored.

The voice feature database may contain registration data with different registrant IDs and the same utterance content. In addition, there may be a plurality of registered data having the same registrant ID but different utterance contents. The registration data having the same registrant ID usually has the same group ID.

When the speaker identification device 1 and the voice registration device 2 are different devices, the voice registration device 2 classifies the registration data of all the registered persons into groups and associates them with the identification code and the utterance content of each registered person. The voice feature database stored in the storage unit 15 is transferred to the speaker identification device 1. When the speaker identification device 1 and the voice registration device 2 are common devices, the storage unit 15 of the voice registration device 2 is used as it is as the storage unit 15 of the speaker identification device 1. The voice registration device 2 and the speaker identification device 1 may be able to access the common storage unit 15. This completes voice registration.

(Speaker identification)
The speaker identification device 1 shown in FIG. 1 is composed of collation data which is voice feature data extracted from the voice input from the microphone 21 and voice feature data stored in the storage unit 15 created by the voice registration device 2. Compare with some registration data to identify which of the registered speakers is the speaker. The voice feature database stored in the voice input unit 10, the feature extraction unit 11, the voice recognition unit 12, and the storage unit 15 of the speaker identification device 1 is the same as that of the voice registration device 2.

The voice input unit 10 samples the voice signal input from the microphone 21 by the speaker who is the person to be identified at a predetermined frequency, performs AD conversion, and generates voice data. The voice input unit 10 sends voice data to the feature extraction unit 11 and the voice recognition unit 12.

The feature extraction unit 11 extracts collation data, which is the voice feature data of the person to be identified, from the voice data. The voice feature data is, for example, a speaker information representation based on a GMM (Gaussian Mixture Model) super vector, i-vector, or tensor decomposition. The type of voice feature data of the collation data is the same as the type of voice feature data of the registered data. That is, when the registered data is, for example, an i-vector, the collation data is an i-vector.

Here, a case where i-vector is used as the voice feature data will be described as an example. The feature extraction unit 11 sends the extracted collation data, which is an i-vector, to the classification determination unit 13.

The classification determination unit 13 determines to which group the collation data should be classified into a group in which the registration data, which is the voice feature data stored in the storage unit 15, is clustered according to the similarity between the registration data. .. The registered data is classified into a group clustered in advance by the voice registration device 2 according to the similarity between the registered data. The similarity is, for example, a cosine similarity or a log-likelihood ratio.

For the classification determination, for example, a machine-learned artificial neural network (hereinafter, simply referred to as a neural network) can be used using the registered data classified into groups as training data. The classification determination unit 13 may compare the centroid vector of each group with the collation data, and determine the group of centroids having the highest similarity as the group to which the collation data should be classified. The classification determination unit 13 sends the number of the group to which the collation data should be classified, that is, the most probable group to which the collation data belongs, to the data extraction unit 14.

FIG. 4 is a diagram showing an example of a neural network of the classification determination unit according to the embodiment. A neural network consists of an input layer, an intermediate layer and an output layer, each containing a node composed of artificial neurons (hereinafter, simply referred to as a neuron), and an edge that connects the nodes to each other between adjacent layers. .. The intermediate layer includes one or more n layers. Each element xi of the voice feature data is input to each node i (i = 1 ... k) of the input layer. In the intermediate layer, the outputs of the previous layers are combined, the result calculated by the activation function is transmitted to the subsequent layers, and finally output to the output layer. The output layer has as many nodes as the number of groups in which the registered data is clustered, and each node j (j = 1 ... M) outputs the probability yj that the collation data is classified into the group j.

Machine learning is performed by inputting the registered data classified into groups into the neural network, backpropagating the difference between the group in which the registered data is classified and the output of the neural network, and adjusting each parameter. Keep it.

The classification determination unit 13 inputs the matching data into the trained model of the neural network, and sets the number of the node having the largest output of the output layer, that is, the number of the most probable group, as the number of the group to which the matching data should be classified. To do. The classification determination unit 13 sends the number of the group to which the collation data should be classified to the data extraction unit 14.

When determining the classification of collation data by a neural network, the amount of calculation for the determination is almost determined by the number of nodes in the input layer and the number of layers in the intermediate layer, and is not proportional to the number of nodes in the output layer. Therefore, the amount of calculation is small when the number of groups increases, as compared with the method of determining the classification of groups based on the degree of similarity with the centroid of each group. As a result, even if the number of groups increases, it is possible to suppress an increase in the time required for identification.

The voice recognition unit 12 extracts a voice feature from the voice data, searches for a utterance content that is closest to the voice feature, and extracts a phoneme string, a syllable string, or a character string indicating the utterance content. The voice recognition unit 12 sends the obtained phoneme string, syllable string or character string to the data extraction unit 14.

The data extraction unit 14 reads from the storage unit 15 the registration data having the same utterance content as the utterance content sent from the voice recognition unit 12 among the registration data belonging to the group of numbers sent from the classification determination unit 13. It is sent to the speaker identification unit 16.

The speaker identification unit 16 compares each of the registered data sent from the data extraction unit 14 with the collation data, and if the highest similarity exceeds a set threshold value, the speaker identification unit 16 registers the highest similarity. The person whose data is registered is determined to be the person to be identified who is the speaker. That is, among the registered data in which the collation data is classified into the group to be classified and the utterance content is the same as the utterance content of the collation data, the similarity with the collation data is the highest and the similarity with the collation data is high. The registered person to whom the identification code associated with the registered data exceeding the threshold is given is determined to be the speaker of the collation data.

The type of similarity between the registered data and the collation data is the same as the type of similarity when the registered data is clustered. For example, when the registered data is clustered by the cosine similarity, the similarity between the registered data and the collation data is calculated by the cosine similarity.

Based on the result of determining that the person to be identified is one of the registered persons, the device connected to the speaker identification device 1 can be made to perform the operation permitted by the registered person. For example, the door can be unlocked, the registered person can be permitted to access information unique to the registered person, or the registered person can be made to respond to an AI speaker suitable for the user. The speaker identification device 1 may be incorporated in, for example, a building security system, a customer information management device, an AI speaker, or the like.

FIG. 5 is a flowchart showing an example of the speaker identification operation according to the embodiment. The speaker identification device 1 generates voice data from the voice signal when voice is input from the speaker who is the person to be identified (step S10). The feature extraction unit 11 extracts the matching data from the voice data and sends it to the classification determination unit 13 (step S11). The voice recognition unit 12 recognizes the utterance content from the voice data and sends it to the data extraction unit 14 (step S12). The classification determination unit 13 determines a group to which the collation data should be classified and sends it to the data extraction unit 14 (step S13).

When the data extraction unit 14 reads the registered data having the same utterance content as the utterance content of the collation data among the registered data belonging to the group to which the collation data should be classified from the storage unit 15 and sends it to the speaker identification unit 16. The speaker identification unit 16 selects one of the read registered data (step 14). Then, the degree of similarity between the collation data and the selected registered data is calculated (step S15). If there is registered data that has not been selected yet (step S16; Y), one of the unselected registered data is selected again (step S14), and the similarity is calculated (step S15).

When the speaker identification unit 16 calculates the similarity with the collation data for all the registered data sent from the data extraction unit 14 (step S16; N), the speaker identification unit 16 selects the maximum value of the calculated similarity (step S17). .. If the maximum value of the calculated similarity is larger than the threshold value (step S18; Y), the registered person to whom the identification code associated with the registration data corresponding to the similarity is assigned is referred to as the speaker of the matching data (talk). Person) (step S19). If the maximum value is equal to or less than the threshold value (step S18; N), it is determined that the speaker of the collation data is neither of the registered persons (step S20).

When the speaker of the collation data is identified as one of the registered persons, the speaker identification device 1 can cause the connected device to perform an operation permitted by the registered person. If it is determined that the speaker of the collation data is neither of the registered persons, the identified person can be prompted to speak again. Even if the registered data stored in the storage unit 15 and belonging to the group to which the collation data should be classified does not include the registered data corresponding to the utterance content recognized by the voice recognition unit 12, the person to be identified again Can encourage vocalization.

The speaker identification unit 16 does not compare the maximum value of the similarity with the threshold value, and simply refers to the registered person to which the identification code associated with the registration data corresponding to the similarity of the maximum value is given. It may be determined that the speaker is the speaker. In that case, the threshold can be regarded as the minimum possible value of similarity.

As described above, the speaker identification device 1 according to the embodiment determines which of the groups in which the registered data is clustered in advance the collation data should be classified, and the degree of similarity to the group to be classified. Since the registered data to be determined is limited, the time required for identification can be shortened as compared with the case of comparing with all the registered data. Further, in the group to which the collation data should be classified, the comparison target is limited to the registered data having the same utterance content as the collation data, so that the time required for identification can be further shortened. In addition, since the utterance content that is irrelevant to the characteristics of the speaker is used to compare only the registered data having the same utterance content, the possibility that the registered data to be compared are similar to each other is reduced. , The possibility of misrecognizing the speaker is reduced.

When the classification determination unit 13 determines which group the collation data should be classified into, the largest output of the output layer of the neural network, that is, the maximum value of the probability of being classified into a group is a reference value. If it is smaller, it may be determined that the collation data is not classified into any group, and the speaker of the collation data is not one of the registered persons. The standard value in this case may be determined according to the number of registered data and the number of groups. If the maximum value of the probability of being classified into a group is smaller than the reference value, it is estimated that the similarity of any registered data in that group is smaller than the threshold value. In this case, since it is determined that the speaker is not the registered person without calculating the similarity with the registered data, the time required for speaker identification can be further shortened.

Modification example.
FIG. 6 is a block diagram of a speaker identification device according to a modified example of the embodiment. In the modified example, the voice recognition unit 12 is not provided, and voice recognition is not performed. Other configurations are the same as in the embodiment.

In the modified example, the data extraction unit 14 reads the registration data belonging to the group of numbers sent from the classification determination unit 13 from the storage unit 15 and sends it to the speaker identification unit 16. The registration data sent to the speaker identification unit 16 includes voice feature data having different utterance contents.

In the modified example, the voice feature database stored in the storage unit 15 may be associated with the utterance content or may not include the generated content, as in the embodiment. Even if the utterance content is not included, there may be a plurality of registered data having the same registrant ID but different utterance contents.

The speaker identification unit 16 compares each of the registered data sent from the data extraction unit 14 with the collation data, and if the highest similarity exceeds a set threshold value, the speaker identification unit 16 registers the highest similarity. The person whose data is registered is determined to be the person to be identified who is the speaker. In the modified example, among the registered data classified into the group to which the collated data should be classified, the registered data having the highest similarity with the collated data and the similarity with the collated data exceeds the threshold value is supported. The registered person to which the attached identification code is given is determined to be the speaker of the collation data.

FIG. 7 is a flowchart showing an example of the speaker identification operation according to the modified example. In the modified example, in the operation of the embodiment of FIG. 5, the voice recognition step S12 is omitted. Further, since the data extraction unit 14 reads the registered data belonging to the group to which the collation data should be classified from the storage unit 15 and sends it to the speaker identification unit 16, the speaker identification unit 16 classifies the collation data. Select one of the registered data belonging to the power group (step S14'). Other operations are the same as the flowchart of FIG.

In the modified example, since voice recognition is not performed, the registered data with different generated contents are also compared with the collated data, but since the collated data is limited to the group to be classified, it requires more identification than comparing with all the registered data. You can shorten the time. Moreover, since voice recognition is not performed, the processing time is short accordingly.

FIG. 8 is a block diagram showing an example of the hardware configuration of the speaker identification device according to the embodiment. As shown in FIG. 8, the speaker identification device 1 includes a control unit 41, a main storage unit 42, an external storage unit 43, an operation unit 44, a display unit 45, an input / output unit 46, and a transmission / reception unit 47. The main storage unit 42, the external storage unit 43, the operation unit 44, the display unit 45, the input / output unit 46, and the transmission / reception unit 47 are all connected to the control unit 41 via the internal bus 40.

The control unit 41 is composed of a CPU (Central Processing Unit) and the like, and according to the control program 50 stored in the external storage unit 43, the voice input unit 10, the feature extraction unit 11, and the voice recognition unit 12 of the speaker identification device 1. Each process of the classification determination unit 13, the data extraction unit 14, the storage unit 15, and the speaker identification unit 16 is executed.

The main storage unit 42 is composed of a RAM (Random-Access Memory) or the like, loads the control program 50 stored in the external storage unit 43, and is used as a work area of the control unit 41.

The external storage unit 43 is composed of a flash memory, a hard disk, a non-volatile memory such as a DVD-RAM (Digital Versatile Disc Random-Access Memory) and a DVD-RW (Digital Versatile Disc ReWritable), and processes the speaker identification device 1. The program to be executed by the control unit 41 is stored in advance, and the data stored by this program is supplied to the control unit 41 according to the instruction of the control unit 41, and the data supplied from the control unit 41 is stored.

The operation unit 44 is composed of a pointing device such as a keyboard and a mouse, and an interface device for connecting the keyboard, the pointing device, and the like to the internal bus 40. A voice recognition result selection instruction or the like is input via the operation unit 44 and supplied to the control unit 41.

The display unit 45 is composed of an LCD (Liquid Crystal Display), an organic EL display, or the like, and displays a speaker identification result, a character string of voice-recognized voice content, and the like.

The input / output unit 46 is composed of a serial interface or a parallel interface. The input / output unit 46 connects the microphone 21 and inputs an audio signal. In addition, a speaker (not shown) is connected to play, for example, a message prompting the person to be identified to input voice.

The transmission / reception unit 47 is composed of a network termination device or a wireless communication device connected to the network, and a serial interface or a LAN (Local Area Network) interface connected to them. The transmission / reception unit 47 exchanges data with, for example, a device that uses the speaker recognition result via the network.

The processing of the voice input unit 10, the feature extraction unit 11, the voice recognition unit 12, the classification determination unit 13, the data extraction unit 14, the storage unit 15, and the speaker identification unit 16 of the speaker identification device 1 shown in FIG. 1 is The control program 50 executes the processing by using the control unit 41, the main storage unit 42, the external storage unit 43, the operation unit 44, the display unit 45, the input / output unit 46, the transmission / reception unit 47, and the like as resources.

The configuration of the speaker identification device 1 described in each embodiment is an example, and can be arbitrarily changed and modified. The configuration of the speaker identification device 1 is not limited to all of those shown in the embodiments. For example, as described in the embodiment, the same device may be used in common by the speaker identification device 1 and the voice registration device 2. Further, the storage unit 15 may be installed on the network, and the storage unit 15 may be accessed from the speaker identification device 1 and the voice registration device 2 via the network.

In addition, the above hardware configuration and flowchart are examples, and can be arbitrarily changed and modified.

Speaker identification of speaker identification device 1 composed of voice input unit 10, feature extraction unit 11, voice recognition unit 12, classification determination unit 13, data extraction unit 14, storage unit 15, speaker identification unit 16, and the like. The central part of processing can be realized by using a normal computer system without relying on a dedicated system. For example, a computer program for executing the above operation is stored and distributed in a computer-readable recording medium (USB memory, CD-ROM, DVD-ROM, etc.), and the computer program is installed in the computer. The speaker identification device 1 that executes the above processing may be configured accordingly. Further, the speaker identification device 1 may be configured by storing the computer program in a storage device of a server device on a communication network such as the Internet and downloading it by a normal computer system.

Further, when the speaker identification device 1 is realized by sharing the OS (operating system) and the application program, or by coordinating the OS and the application program, only the application program part is stored in the recording medium or the storage device. You may.

It is also possible to superimpose a computer program on a carrier wave and distribute it via a communication network. For example, the computer program may be posted on a bulletin board system (BBS, Bulletin Board System) on a communication network, and the computer program may be distributed via the network. Then, by starting this computer program and executing it in the same manner as other application programs under the control of the OS, the above processing may be executed.

1 Speaker identification device 2 Voice registration device 10 Voice input unit 11 Feature extraction unit 12 Voice recognition unit 13 Classification judgment unit 14 Data extraction unit 15 Storage unit 16 Speaker identification unit 17 Clustering unit 18 Input unit 19 Data registration unit 21 Microphone 40 Internal bus 41 Control unit 42 Main storage unit 43 External storage unit 44 Operation unit 45 Display unit 46 Input / output unit 47 Transmission / reception unit 50 Control program

Claims (5)

A memory that classifies the registration data, which is the voice feature data of all the registered persons, into a group clustered according to the similarity between the registered data, and stores the registered data in association with the identification code assigned to each of the registered persons. Department and
A feature extraction unit that extracts matching data, which is voice feature data, from the input voice data,
A classification determination unit that determines which group the collation data should be classified into,
Among the registered data classified into the group to which the collated data should be classified, the registered data having the highest similarity with the collated data and having a similarity with the collated data exceeding a threshold value. A speaker identification unit that determines that the registered person to which the identification code associated with the data is assigned is the speaker of the collation data.
A speaker identification device. The storage unit stores a phoneme string, a syllable string, or a character string indicating the utterance content of each of the registered data in association with the registered data.
The speaker identification device further includes a voice recognition unit that extracts a phoneme string, a syllable string, or a character string indicating the utterance content from the input voice data.
The speaker identification unit is classified into the group to which the collation data should be classified, and has the highest degree of similarity to the collation data among the registered data whose utterance contents are the same as the utterance contents of the collation data. The registered person to whom the identification code associated with the registered data whose similarity with the collated data exceeds the threshold is determined to be the speaker of the collated data.
The speaker identification device according to claim 1. The story according to claim 1 or 2, wherein the classification determination unit includes a trained model of a neural network in which the voice feature data of the registered persons classified into the group by the clustering is machine-learned as training data. Person identification device. A feature extraction step that extracts matching data, which is voice feature data, from the input voice data,
A classification determination step for determining which group the collation data should be classified into, which is a group in which the registration data, which is the voice feature data of all the registered persons, is clustered according to the similarity between the registration data.
The group to which the collation data should be classified in the registered data stored in association with the identification code given to each of the registered persons by classifying the registered data into the clustered group. Among the registered data classified in the above, the identification code associated with the registered data having the highest degree of similarity with the collated data and having a degree of similarity with the collated data exceeding the threshold is given. A speaker identification step for determining the registered person as the speaker of the collation data, and
Speaker identification method. Computer,
A memory that classifies the registration data, which is the voice feature data of all the registered persons, into a group clustered according to the similarity between the registered data, and stores the registered data in association with the identification code assigned to each registered person. Department,
Feature extraction unit that extracts matching data that is voice feature data to be identified from the input voice data,
The degree of similarity with the collation data among the classification determination unit for determining which group the collation data should be classified into and the registered data classified into the group to which the collation data should be classified. The speaker identification unit, which determines that the registered person corresponding to the identification code of the registered data whose similarity with the collated data exceeds the threshold value is the speaker of the collated data.
A program that functions as.

Images