JP4490384B2 - Method and apparatus for evaluating speech similarity - Google Patents

Description

  The present invention relates to a technique for evaluating speech similarity using a so-called interband correlation matrix, and a technique for speaker authentication and speaker identification using this technique.

In speaker authentication and speaker identification, it is necessary to evaluate how similar a voice uttered by a speaker and a voice obtained in advance from a specific speaker are. In general, in order to evaluate the similarity of speech, the features of the speech to be compared are quantified by some means to obtain feature amounts, and the feature amounts obtained from the speeches are compared. This is a method of evaluating the similarity of speech. Non-Patent Document 1 discloses a technique for performing speaker identification using an interband correlation matrix as such a voice feature amount. This inter-band correlation matrix is a matrix whose elements are correlation values between envelope bands of each component of the band-divided speech. The contents of the interband correlation matrix obtained from the voices of a plurality of speakers are not significantly affected by the utterance contents of each speaker and depend on the speakers themselves. If the speakers are the same, an interband correlation matrix having similar contents can be obtained regardless of the utterance contents. Therefore, speaker authentication and speaker identification can be performed by using this interband correlation matrix as a feature amount of speech.
Paper presented by Michiko Kazama, Mikio Higashiyama, and Yoshio Yamazaki at the Institute of Electronics, Information and Communication Engineers in March 2002, “Speaker Identification Using Bandwidth Correlation Matrix of Envelope”

  According to the technique disclosed in Non-Patent Document 1 described above, speaker authentication and identification can be performed regardless of the utterance content. However, even in the case of using this technique, if noise is superimposed on the input voice in the process of acquiring the input voice from the speaker, the speaker authentication or speaker identification will be hindered.

  The present invention has been made in view of the above-described circumstances, and an object thereof is to improve resistance to noise in speech similarity evaluation performed using an interband correlation matrix.

In the present invention, an N-by-N interband correlation matrix is input as an audio feature quantity, with the correlation value of the envelope of the audio component between N (N is a plurality) bands arranged on the frequency axis as elements. Generate from voice. When there are two voices to be compared, at least one type of n satisfying 1 ≦ n ≦ N is selected, and for each n, two interband correlation matrices obtained for each voice to be compared Substitute one row n and column n in the other row n and column n, or decimate row n and column n from both of the two interband correlation matrices, and after this substitution or decimation The similarity calculation processing for calculating the similarity between the two band correlation matrices is executed, and the speech similarity is evaluated based on the execution result.
Here, when one of the two interband correlation matrices to be compared is acquired from the input speech in an environment where noise occurs, the elements of the interband correlation matrix may be affected by noise. . For example, when an input speech is acquired in a state where components in a specific band are buried in noise energy, and an interband correlation matrix is generated from the input speech, a row related to the band to which the noise belongs in the interband correlation matrix and The noise effect appears in the column. However, according to the present invention, the similarity between the interband correlation matrices can be calculated by excluding the influence of such noise by the above-described row and column substitution or thinning process.
Typical examples of the technology for evaluating the similarity of speech according to the present invention include speaker authentication and speaker identification. The details will be clarified in “Best Mode for Carrying Out the Invention” described below.

Embodiments of the present invention will be described below with reference to the drawings.
<First Embodiment>
FIG. 1 is a block diagram showing the configuration of a speaker authentication apparatus according to an embodiment of the present invention. This speaker authentication apparatus is roughly divided into a registration unit 10, a template DB (database) 20, and an authentication unit 30.

  The configuration of the registration unit 10 is as follows. First, the ID input unit 11 is a device that receives an ID of a user who has a right to receive a certain service, and includes a keyboard or the like. The voice input unit 12 is a device that receives the voice of a user having the above-mentioned legal title. For example, the voice input unit 12 includes a microphone and an A / D converter that converts an analog voice signal output from the microphone into voice data in a digital format. ing. The utterance section detection unit 13 is a device that monitors the level of the voice data output from the voice input unit 12 and passes the voice data in the utterance section whose level is equal to or greater than a threshold value.

  The feature amount extraction unit 14 is a device that extracts feature amounts from voice data of an utterance section that has passed through the utterance section detection unit 13. In the present embodiment, a plurality of types of sentences are uttered for each user, and the feature amount extraction unit 14 extracts feature amounts from the voices of the respective sentences. The registration information creation unit 15 associates the feature amount obtained by averaging the feature amounts with the user ID received by the ID input unit 11 and registers the feature amount in the template DB 20 as registration information.

  The configuration of the authentication unit 30 is as follows. First, the ID input unit 31 is a device that receives an ID of a user to be authenticated, and includes a keyboard or the like. The voice input unit 32 is a device that receives the voice of the user to be authenticated, and is configured by a microphone, an A / D converter, and the like, similar to the voice input unit 12. A speech segment detection unit 33 similar to the speech segment detection unit 13 is provided following the voice input unit 32.

  The feature amount extraction unit 34 is a device that extracts a feature amount from voice data that has passed through the utterance section detection unit 33 by the same extraction method as the feature amount extraction unit 14. The registration information selection unit 35 is a device that reads the feature amount associated with the user ID received by the ID input unit 31 from the template DB 20 and outputs the feature amount together with a predetermined threshold value.

  The feature amount comparison unit 36 compares the feature amount extracted by the feature amount extraction unit 34 with the feature amount read by the registered information selection unit 35, and uses the similarity between the voices represented by both feature amounts as a comparison result. It is a device that outputs. The authentication result output unit 37 compares the similarity output from the feature amount comparison unit 36 with the threshold output from the registration information selection unit 35, thereby generating the user who is the source of the audio received by the audio input unit 32. Is an apparatus that authenticates whether or not the user corresponds to the ID received by the ID input unit 31 and outputs an authentication result.

  Each of the feature quantity extraction units 14 and 34 has an interband correlation matrix generation unit shown in FIG. The frequency analysis unit 51 performs FFT (Fast Fourier Transform) on the voice data that has passed through the utterance section detection unit 13 or 33 in units of a frame of a certain time length, and outputs an amplitude spectrum obtained by the FFT.

  The band-specific envelope generation unit 52 extracts each component belonging to each of the first to Nth bands arranged on the frequency axis from the FFT result of the frequency analysis unit 51, and generates an envelope of each component of the sound belonging to each band. Device. More specifically, the band-specific envelope generator 52 performs the following processing. First, for each frame, those belonging to the first to Nth bands on the frequency axis are extracted from the amplitude spectrum obtained from the frequency analysis unit 51, and converted into a dB value format. Then, for each frequency band, the transition of the amplitude spectrum of the dB value expression obtained for each frame is obtained, and the envelopes E-i (i = 1 to N) of the sound components belonging to the first band to the Nth band are obtained. Output. Note that the envelope E-i (i = 1 to N) of the sound components for each band is obtained by dividing the band of the input sound by a filter bank composed of a plurality of BPFs (band pass filters) having different pass bands. You may obtain | require by performing processes, such as a half-wave rectification and smoothing, to the signal of each band which passed through.

The correlation value calculation unit 53 performs the j-th band envelope Ej and the k-th band envelope Ek for all (j, k) combinations in the range of j = 1 to N and k = 1 to N. Correlation value ajk is calculated, and an N-row N-column interband correlation matrix having correlation values ajk (j = 1 to N, k = 1 to N) as elements is output.
The details of the configuration of the feature quantity extraction units 14 and 34 have been described above.

  The feature amount comparison unit 36 receives the inter-band correlation matrix as described above from the feature amount extraction unit 34 and the registered information selection unit 35, and executes a similarity calculation process for calculating the similarity between the two band correlation matrices. To do. Then, the authentication result output unit 37 is a user corresponding to the ID received by the ID input unit 31 by the user who is the source of the voice received by the voice input unit 32 based on the execution result of the similarity calculation process. Whether or not authentication processing is performed.

  By the way, when a user to be authenticated utters a voice in an environment where noise is generated, a component of a relatively narrow specific band in the voice may be buried in the noise energy. In this case, for example, when a frequency component belonging to the qth (1 ≦ q ≦ N) band of the first to Nth bands in the user's voice is buried in the noise energy, the influence is caused by the feature amount extraction unit 34. It appears in the q-th row and q-th column of the output interband correlation matrix.

  The feature of this embodiment is that the similarity calculation process of the feature amount comparison unit 36 and the authentication process of the authentication result output unit 37 improved so that a correct authentication result can be obtained even in a situation where the influence of noise appears. In the embodiment. There are the following two aspects of the improved similarity calculation processing and authentication processing.

<< First Aspect >>
In this aspect, the feature quantity comparison unit 36 selects at least one type of n satisfying 1 ≦ n ≦ N (all n from n = 1 to N in this example), and each n is illustrated in FIG. As described above, one of the inter-band correlation matrix of the input speech generated by the feature amount extraction unit 34 and the inter-band correlation matrix read from the template DB 20 by the registration information selection unit 35 (in this example, read from the template DB 20). The n-th row and the n-th column of the inter-band correlation matrix) are respectively substituted into the n-th row and the n-th column of the other (in this example, the inter-band correlation matrix of the input speech), and the inter-band correlation matrix after the substitution A similarity calculation process for calculating the similarity between the two is executed. And the authentication result output part 37 authenticates based on the execution result of this similarity calculation process. In a preferred mode, the authentication result output unit 37 obtains the number of times that the similarity is equal to or greater than a threshold in the similarity calculation process executed for each n, and performs authentication based on whether or not the number is equal to or greater than a predetermined value.

<< Second Aspect >>
In this aspect, the feature quantity comparison unit 36 selects at least one kind of n satisfying 1 ≦ n ≦ N (in this example, all n from n = 1 to N), and each n is illustrated in FIG. As described above, the n-th row and the n-th column are respectively obtained from both the inter-band correlation matrix of the input speech generated by the feature amount extraction unit 34 and the inter-band correlation matrix read from the template DB 20 by the registration information selection unit 35. A similarity calculation process is performed to calculate the similarity between the decimation and the correlation matrix between both bands after the decimation. And the authentication result output part 37 authenticates based on the execution result of this similarity calculation process.

There are various modes for calculating the similarity between the interband correlation matrices. In a preferred embodiment, the feature quantity comparison unit 36 obtains, for example, correlation values between elements of the correlation matrix between the two bands for each column, and averages the correlation values for all the columns. Output the value as a similarity. In another preferred embodiment, the feature amount comparison unit 36 calculates the similarity D according to the following equation (1) when the two interband correlation matrices to be compared are A and B.
D = √ (trace (AB ⁻¹ ) · trace (BA ⁻¹ )) (1)
However, in the above formula (1), trace () means the sum of the diagonal elements of the square matrix in parentheses.

  In the first aspect, when the nth row and the nth column are substituted between different types of interband correlation matrices, the inverse matrix is obtained by losing the regularity of the interband correlation matrix after the substitution. Of course, the similarity D may not be calculated by the above equation (1). Therefore, in the first aspect, it is preferable to calculate the similarity between the interband correlation matrices by other arithmetic processing using the properties of the correlation matrix.

  Hereinafter, a first operation example of this embodiment in which speaker authentication is performed according to the first aspect and a second operation example of this embodiment in which speaker authentication is performed according to the second aspect will be described.

<< First Operation Example >>
FIG. 5 is a flowchart showing a first operation example of this embodiment. In the authentication process of a user, first, the ID of the user is acquired by the ID input unit 31, and the voice data in the utterance section of the user is acquired by the voice input unit 32 and the utterance section detection unit 33 ( Step S101).

  Next, the feature amount extraction unit 34 generates an interband correlation matrix A from the input speech, that is, the speech data acquired in step S101 (step S102), and the registration information selection unit 35 sets the ID acquired in step S101. The associated interband correlation matrix B is read from the template DB 20 (step S103). Then, the authentication result output unit 37 initializes the point register P indicating a certain degree of positive authentication result to “0” (step S104).

  When the number of bands (that is, the number of rows and columns) in the interband correlation matrix is N, in the present embodiment, the loop processing of steps S105 to S110 is repeated for each n from n = 1 to N. Execute. First, the feature quantity comparison unit 36 substitutes the nth row and the nth column of the interband correlation matrix B corresponding to the user ID into the nth row and the nth column of the interband correlation matrix A of the input speech (step) S106a). Next, the feature amount comparison unit 36 calculates a similarity D between the interband correlation matrix A after the substitution and the interband correlation matrix B corresponding to the user ID (step S107a). Then, the authentication result output unit 37 compares the similarity D with the threshold th1 (step S108). The authentication result output unit 37 adds a point having a predetermined size to the point register P only when the similarity D is equal to or greater than the threshold th1 (steps S108 and S109).

  When the loop processing of steps S105 to S110 is completed for each n from n = 1 to N, the authentication result output unit 37 determines whether or not the content of the point register P is greater than or equal to the threshold th2 (step S111). If the determination result is affirmative, the user who is the speaker is determined to be the person corresponding to the user ID (step S112). Otherwise, it is determined that the user is a spoofer (step S113). Output the result.

<< second operation example >>
FIG. 6 is a flowchart showing a second operation example of the present embodiment. In the second operation example, steps S106a and S107a in the first operation example are replaced with steps S106b and S107b. In this second operation example, in step S106b, the feature amount comparison unit 36 thins out the n-th row and the n-th column from both the inter-band correlation matrix A of the input speech and the inter-band correlation matrix B corresponding to the user ID. . Next, in step S107b, the feature amount comparison unit 36 calculates the similarity D between the interband correlation matrices A and B (both are (N-1) rows (N-1) columns matrix) after this thinning. The contents of other processes are the same as those in the first operation example.

  According to the present embodiment described above, when the voice input unit 32 acquires a user's voice, for example, if a voice component in the q-th band (1 ≦ q ≦ N) is buried in noise energy, the feature amount is extracted. The influence appears in the q-th row and the q-th column of the interband correlation matrix A output by the unit 34. However, in the present embodiment, the feature amount comparison unit 36 reads out the interband correlation matrix A generated by the feature amount extraction unit 34 and the registered information selection unit 35 for each n in the range of 1 ≦ n ≦ N. The n-th row and the n-th column of the inter-band correlation matrix B are respectively substituted into the other n-th row and the n-th column (step S106a), or the n-th row and the inter-band correlation matrices A and B are both The nth column is thinned (step S106b), and similarity calculation processing is performed to calculate the similarity D between the correlation matrices A and B after the substitution or after the thinning (steps S107a and S107b). And the authentication result output part 37 authenticates based on the frequency | count that the similarity D became more than threshold value th1 in the similarity calculation process performed about each n. Here, even if the influence of noise appears in, for example, the q-th row and the q-th column of the interband correlation matrix A output by the feature amount extraction unit 34, the source of the input voice is a legitimate user corresponding to the user ID. In some cases, a high similarity D is highly likely to be obtained in the similarity calculation process executed for n = q. Therefore, according to the present embodiment, authentication is performed as to whether or not the source of the input voice is a legitimate user corresponding to the user ID even in a situation where the input voice is acquired in an environment where noise occurs. Can be done accurately.

<Second Embodiment>
FIG. 7 is a block diagram showing a configuration of a speaker identification device according to an embodiment of the present invention. The speaker identification device includes a registration unit 10 and a template DB 20 similar to those in the first embodiment (FIG. 1) and an identification unit 40.

  The configuration of the identification unit 40 is as follows. First, the voice input unit 41 is a device that receives voice from a user to be identified and outputs voice data. The utterance section detection unit 42 is a device that detects and outputs data in the utterance section among the voice data output from the voice input unit 41. The feature amount extraction unit 43 is a device that extracts a feature amount from the voice data that has passed through the utterance section detection unit 42. The feature amount comparison unit 44 is a device that executes a similarity calculation process on the feature amount extracted by the feature amount extraction unit 43 and the feature amount of each user registered in the template DB 20. The identification result output unit 45 receives the voice input unit 41 among the users who generate the feature amounts stored in the template DB 20 based on the execution result of the similarity calculation process in the feature amount comparison unit 44. This is a device that selects a user who is estimated to be a source of voice and outputs the ID of the user.

  The configuration of the feature quantity extraction units 14 and 43 in the present embodiment and the contents of the interband correlation matrix which is the feature quantity extracted by these are as described in the first embodiment. In addition, the similarity calculation performed by the feature amount comparison unit 44 and the speaker identification performed by the identification result output unit 45 include the following two modes as in the first embodiment.

<< First Aspect >>
In this speaker identification system, the feature quantity comparison unit 44 compares the inter-band correlation matrix of the input speech generated by the feature quantity extraction unit 43 with each inter-band correlation matrix in the template DB 20. In the first mode, the feature quantity comparison unit 44 is at least one type of n satisfying 1 ≦ n ≦ N (in this example, all n from n = 1 to N) as in the first embodiment. For each n, the first one of the inter-band correlation matrix of the input speech generated by the feature quantity extraction unit 43 and the inter-band correlation matrix in the template DB 20 (in this example, the inter-band correlation matrix in the template DB 20) is selected. Substituting n-th row and n-th column into the n-th row and n-th column of the other (in this example, the inter-band correlation matrix of the input speech), and calculating the similarity between the inter-band correlation matrices after this substitution The degree calculation process is executed. And the identification result output part 45 performs identification based on the execution result of this similarity calculation process. In a preferred embodiment, the identification result output unit 45 obtains, for each n, the one in which the maximum similarity is obtained by the similarity calculation process from among the inter-band correlation matrices stored in the template DB 20, and is stored in the template DB. The user who is the generation source of the inter-band correlation matrix having the largest number of times of obtaining the maximum similarity among the inter-band correlation matrices is estimated as the user who is the generation source of the voice received by the voice input unit 41.

<< Second Aspect >>
In this aspect, for each n, the feature quantity comparison unit 44 has an n-th row and an n-th column of both the inter-band correlation matrix of the input speech generated by the feature quantity extraction unit 34 and the inter-band correlation matrix in the template DB 20. And a similarity calculation process for calculating the similarity between the correlation matrices between the two bands after the decimation. The processing contents of the identification result output unit 45 are the same as those in the first aspect.

  Next, a first operation example of the present embodiment for performing speaker identification according to the first aspect and a second operation example of the present embodiment for performing speaker identification according to the second aspect will be described.

<< First Operation Example >>
FIG. 8 is a flowchart showing a first operation example of this embodiment. In the identification processing of a certain user, first, voice data in the speech segment of the user to be identified is acquired by the speech input unit 41 and the speech segment detection unit 42 (step S201).

  Next, the feature quantity extraction unit 43 generates an interband correlation matrix A from the audio data acquired in step S201 (step S202). At this time, the identification result output unit 45 initially sets the point register P (i) (i = 1 to M) provided for each user who is a subscriber of the speaker identification service according to the present embodiment to “0”. Turn into. Here, the point register P (i) is a register for storing information indicating the degree to which the i-th user is estimated to be the generation source of the audio data acquired in step S201.

  When the number of bands in the interband correlation matrix is N, in the present embodiment, the loop processing of steps S203 to S213 is repeatedly executed for each n in the range of n = 1 to N. In this loop process, first, the maximum similarity Dmax is initialized to 0 (step S204). When the number of interband correlation matrices registered in the template DB 20 is M, the loop processing of steps S205 to S211 is repeatedly executed for each i from i = 1 to M.

  First, the feature amount comparison unit 44 reads the i-th user's interband correlation matrix Bi from the template DB 20 (step S206). Next, the feature quantity comparison unit 44 adds the nth row and the nth column of the interband correlation matrix Bi read in step S206 to the nth row and the nth column of the interband correlation matrix A obtained in step S202. Substitute (step S207a). Next, the feature amount comparison unit 44 calculates the similarity D between the interband correlation matrix A and the interband correlation matrix Bi after this substitution (step S208a). Then, the identification result output unit 45 compares the similarity D obtained in this way with the maximum similarity Dmax (step S209), and only when D> Dmax, Dmax = D and imax = i (Step S210).

  As a result of the above processing being repeated for each i from i = 1 to M, the interband correlation matrix A obtained from the input speech among the M interband correlation matrices Bi (i = 1 to M) in the template DB 20. An index imax of the interband correlation matrix Bimax having the highest correlation value is obtained. The identification result output unit 45 adds a point having a predetermined size to the contents of the point register P (imax) corresponding to the index imax (step S212).

  The above processing is repeated for each N from n = 1 to N. Thereafter, the identification result output unit 45 obtains the index i of the maximum one of the point registers P (i) (i = 1 to M), and outputs the user ID corresponding to this index i as the identification result (step S214). ).

<< second operation example >>
FIG. 9 is a flowchart showing a first operation example of the present embodiment. In the second operation example, steps S207a and S208a in the first operation example are replaced with steps S207b and S208b. In the second operation example, in step S207b, the feature amount comparison unit 44 determines that the nth row and the nth column from both the interband correlation matrix A of the input speech and the interband correlation matrix Bi read from the template DB 20. Thin out. Next, in step S208b, the feature amount comparison unit 44 calculates the similarity D between the interband correlation matrices A and Bi (both are (N-1) rows (N-1) columns matrix) after the thinning. The contents of other processes are the same as those in the first operation example.

  The above is the operation of this embodiment. Also in this embodiment, the same effect as the first embodiment can be obtained.

While various embodiments of the present invention have been described above, other embodiments are possible for the present invention. For example:
<Other embodiments>
(1) In the speaker authentication device in the first embodiment, the registration unit 10 is provided as a device different from the authentication unit 30. However, the authentication unit 30 also has the function of the registration unit 10 so that the user can You may comprise so that the correlation matrix between the bands of one's own voice used for authentication can be registered into template DB20 via the authentication part 30. FIG. The speaker identification device in the second embodiment is the same, and the identification unit 40 may have the function of the registration unit 10 together.
(2) In the first embodiment, the user is authenticated based on the number of times the similarity is greater than or equal to the threshold th1, but the highest similarity among the similarities obtained for various types of n. The user may be authenticated based on the above.
(3) In the second embodiment, after the substitution or decimation of the n-th row and the n-th column, the inter-band correlation matrix Bi whose similarity to the inter-band correlation matrix A of the input speech is first (maximum). For only (i = imax), points were added to the point register P (imax) corresponding to the interband correlation matrix Bimax (step S212 in FIGS. 8 and 9), but the similarity to the interband correlation matrix A is For each inter-band correlation matrix Bi that is second or higher, a point having a size corresponding to the rank of similarity may be added to the corresponding point register P (i).
(4) In each of the above embodiments, when the interband correlation matrix is N rows and N columns, the interband correlation with substitution or decimation of the nth row and the nth column is performed for N types of n from 1 to N. Although the similarity between the matrices was calculated, the similarity between the interband correlation matrices with substituting or thinning out the nth row and the nth column was calculated for n types smaller than N types, and the calculation result A final decision (authentication or identification) may be made based on this. For example, it is conceivable to calculate the similarity between the interband correlation matrices with substitution or thinning of the n-th row and the n-th column for only odd or even n.
(5) In each of the above embodiments, in the similarity calculation process according to the first aspect, the nth row and the nth column of the interband correlation matrix B read from the template DB 20 are used as the interband correlation matrix A of the input speech. The nth row and the nth column are substituted for the nth row and the nth column. Conversely, the nth row and the nth column of the interband correlation matrix A of the input speech are nth of the interband correlation matrix B read from the template DB 20. Substituting in the row and the n-th column, the similarity between both band correlation matrices after this substitution may be calculated.

It is a block diagram which shows the structure of the speaker authentication apparatus which is 1st Embodiment of this invention. It is a block diagram which shows the structure of the correlation matrix production | generation part between bands used as a feature-value extraction part in the same embodiment. It is a figure which shows the 1st aspect of the similarity calculation process in the embodiment. It is a figure which shows the 2nd aspect of the similarity calculation process in the embodiment. It is a flowchart which shows the 1st operation example of the embodiment. It is a flowchart which shows the 2nd operation example of the embodiment. It is a block diagram which shows the structure of the speaker identification device which is 2nd Embodiment of this invention. It is a flowchart which shows the 1st operation example of the embodiment. It is a flowchart which shows the 2nd operation example of the embodiment.

DESCRIPTION OF SYMBOLS 10 ... Registration part, 20 ... Template DB, 30 ... Authentication part, 40 ... Identification part, 12, 32, 41 ... Voice input part, 11, 31 ... ID input part, 13, 33, 42 ... ... Speech section detection unit, 14, 34, 43 ... feature amount extraction unit, 15 ... registration information creation unit, 35 ... registration information selection unit, 36,44 ... feature amount comparison unit, 37 ... output of authentication result Part, 45... Identification result output part.

Claims (10)

Computer
An inter-band correlation matrix generating process for generating an N-row N-column inter-band correlation matrix from input speech, the element being a correlation value of an envelope of speech components between N (N is plural) bands arranged on the frequency axis When,
Between each of N types of n satisfying 1 ≦ n ≦ N , or for each of M (N> M) types of n determined in advance among the N types, the two voices to be compared between the bands Substituting one n-th row and n-th column of the two inter-band correlation matrices obtained by the correlation matrix generation process into the other n-th row and n-th column, the similarity between the inter-band correlation matrices after this substitution A similarity evaluation process for executing similarity calculation processing for calculating the degree of evaluation and evaluating the similarity of speech based on the execution result
A method for evaluating the similarity of speech. Computer
An inter-band correlation matrix generating process for generating an N-row N-column inter-band correlation matrix from input speech, the element being a correlation value of an envelope of speech components between N (N is plural) bands arranged on the frequency axis When,
Between each of N types of n satisfying 1 ≦ n ≦ N , or for each of M (N> M) types of n determined in advance among the N types, the two voices to be compared between the bands Executing a similarity calculation process for thinning out the nth row and the nth column from the two interband correlation matrices obtained by the correlation matrix generation process, and calculating the similarity between the interband correlation matrices after the decimation, Similarity evaluation process for evaluating speech similarity based on the execution result
A method for evaluating the similarity of speech. An inter-band correlation matrix generating means for generating an N-row N-column inter-band correlation matrix having an element as a correlation value of an envelope of speech components between N bands (N is a plurality) arranged on the frequency axis from input speech. When,
Between each of N types of n satisfying 1 ≦ n ≦ N , or for each of M (N> M) types of n determined in advance among the N types, the two voices to be compared between the bands The n-th row and the n-th column of one of the two inter-band correlation matrices obtained by the correlation matrix generating means are substituted into the other n-th row and the n-th column, and the similarity between the inter-band correlation matrices after this substitution A speech similarity evaluation apparatus comprising: similarity evaluation means for executing similarity calculation processing for calculating a degree and performing evaluation related to speech similarity based on the execution result. An inter-band correlation matrix generating means for generating an N-row N-column inter-band correlation matrix having an element as a correlation value of an envelope of speech components between N bands (N is a plurality) arranged on the frequency axis from input speech. When,
Between each of N types of n satisfying 1 ≦ n ≦ N , or for each of M (N> M) types of n determined in advance among the N types, the two voices to be compared between the bands Executing a similarity calculation process for thinning out the nth row and the nth column from the two interband correlation matrices obtained by the correlation matrix generating means, and calculating the similarity between the interband correlation matrices after the decimation, A speech similarity evaluation apparatus comprising: similarity evaluation means for evaluating speech similarity based on the execution result. As the feature amount of each sound obtained from one or a plurality of users, N rows and N columns each having a correlation value of an envelope of sound components between N (N is a plurality) bands arranged on the frequency axis as elements. A database that stores the interband correlation matrix in association with the user ID;
An ID input unit for receiving an ID of a user to be authenticated;
A voice input unit for receiving the voice of the user to be authenticated;
A feature amount extraction unit that generates an N-by-N interband correlation matrix having the correlation value of an envelope of a speech component between N bands arranged on the frequency axis as elements, from the speech received by the speech input unit; ,
A registration information selection unit that reads an interband correlation matrix associated with the user ID received by the ID input unit from the database;
Between each of N types of n satisfying 1 ≦ n ≦ N , or between M (N> M) types of n determined in advance among the N types, between the bands generated by the feature amount extraction unit The correlation matrix and the n-th row and the n-th column of the inter-band correlation matrix read by the registered information selection unit are respectively substituted into the other n-th row and the n-th column, and the inter-band correlation after the substitution A feature amount comparison unit that executes similarity calculation processing for calculating the similarity between matrices;
Based on the execution result of the similarity calculation process, the user who is the source of the voice received by the voice input unit is authenticated whether or not it is a user corresponding to the ID received by the ID input unit. A speaker authentication apparatus comprising: an authentication result output unit for outputting a result. As the feature amount of each sound obtained from one or a plurality of users, N rows and N columns each having a correlation value of an envelope of sound components between N (N is a plurality) bands arranged on the frequency axis as elements. A database that stores the interband correlation matrix in association with the user ID;
An ID input unit for receiving an ID of a user to be authenticated;
A voice input unit for receiving the voice of the user to be authenticated;
A feature amount extraction unit that generates an N-by-N interband correlation matrix having the correlation value of an envelope of a speech component between N bands arranged on the frequency axis as elements, from the speech received by the speech input unit; ,
A registration information selection unit that reads an interband correlation matrix associated with the user ID received by the ID input unit from the database;
Between each of N types of n satisfying 1 ≦ n ≦ N , or between M (N> M) types of n determined in advance among the N types, between the bands generated by the feature amount extraction unit Similarity calculation for thinning out the n-th row and the n-th column from both the correlation matrix and the inter-band correlation matrix read out by the registered information selection unit, and calculating the similarity between the inter-band correlation matrices after this thinning A feature amount comparison unit that executes processing;
Based on the execution result of the similarity calculation process, the user who is the source of the voice received by the voice input unit is authenticated whether or not it is a user corresponding to the ID received by the ID input unit. A speaker authentication apparatus comprising: an authentication result output unit for outputting a result.   The speaker according to claim 5 or 6, wherein the authentication result output unit performs the authentication based on the number of times the similarity is equal to or greater than a threshold in the similarity calculation process executed for each n. Authentication device. As the feature amount of each sound obtained from one or a plurality of users, N rows and N columns each having a correlation value of an envelope of sound components between N (N is a plurality) bands arranged on the frequency axis as elements. A database that stores the interband correlation matrix in association with the user ID;
A voice input unit that receives the voice of the user to be identified;
A feature amount extraction unit that generates an N-by-N interband correlation matrix having the correlation value of an envelope of a speech component between N bands arranged on the frequency axis as elements, from the speech received by the speech input unit; ,
Each of N types of n satisfying 1 ≦ n ≦ N , or each of M (N> M) types of n determined in advance among the N types, and each interband correlation matrix stored in the database, For each combination, the one of the n-th row and the n-th column of the inter-band correlation matrix generated by the feature amount extraction unit and the inter-band correlation matrix stored in the database is changed to the other n-th row and n-th column. A feature amount comparison unit that executes a similarity calculation process for calculating the similarity between the correlation matrices between both bands after the substitution,
Based on the execution result of the similarity calculation process, the user estimated to be the generation source of the voice received by the voice input unit is selected from among the users who are generation sources of the inter-band correlation matrices stored in the database. And a discrimination result output unit for outputting the ID of the selected user. As the feature amount of each sound obtained from one or a plurality of users, N rows and N columns each having a correlation value of an envelope of sound components between N (N is a plurality) bands arranged on the frequency axis as elements. A database that stores the interband correlation matrix in association with the user ID;
A voice input unit that receives the voice of the user to be identified;
A feature amount extraction unit that generates an N-by-N interband correlation matrix having the correlation value of an envelope of a speech component between N bands arranged on the frequency axis as elements, from the speech received by the speech input unit; ,
Each of N types of n satisfying 1 ≦ n ≦ N , or each of M (N> M) types of n determined in advance among the N types, and each interband correlation matrix stored in the database, For each combination, the n-th row and the n-th column are thinned out from both the inter-band correlation matrix generated by the feature amount extraction unit and the inter-band correlation matrix stored in the database. A feature amount comparison unit that executes similarity calculation processing for calculating the similarity between correlation matrices;
Based on the execution result of the similarity calculation process, the user estimated to be the generation source of the voice received by the voice input unit is selected from among the users who are generation sources of the inter-band correlation matrices stored in the database. And a discrimination result output unit for outputting the ID of the selected user.   The identification result output unit obtains, for each n, the one in which the maximum similarity is obtained by the similarity calculation process among the inter-band correlation matrices stored in the database, and stores each band stored in the database. The user who is the generation source of the inter-band correlation matrix having the highest number of times of obtaining the maximum similarity among the inter-correlation matrices is selected as the user who is estimated to be the generation source of the voice received by the voice input unit. The speaker identification device according to claim 8 or 9, characterized by the above.

Images