JP2009116278A - Method and device for register and evaluation of speaker authentication - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method and device for register and evaluation of speaker authentication. <P>SOLUTION: The method for register of speaker authentication comprises the steps of: generating a plurality of sound feature vector sequences based on each of a plurality of the speeches of the same content spoken by the speaker; generating a reference template from the plurality of sound feature vector sequences; generating a false impostor feature vector sequence corresponding to each of the plurality of the sound feature vector sequence, based on a plurality of codes and a code book including the feature vector corresponding to the plurality of codes; and selecting an optimal sound feature subset, based on the plurality of sound feature vector sequences, the reference template and the plurality of false imposter feature vector sequence. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to information processing technology, and particularly to speaker authentication technology.

  Different speakers can be recognized based on the utterance characteristics of each speaker when the speaker is speaking, and thus speaker authentication can be performed. K. Yu, J. Mason, J. Oglesby, “Speaker recognition using hidden Markov models, dynamic time warping and vector quantisation” (Vision, Image and Signal Processing, IEE Proceedings, Vol. 142, Oct. 1995, pp. 313- 18) introduces three speaker recognition engine technologies such as HMM, DTW (Dynamic Time Warping) and VQ.

  Typically, the speaker authentication process includes a registration phase and a verification phase. In the registration phase, the speaker template of the speaker (user) is generated according to the utterance including the password spoken by the speaker himself, and in the verification phase, the test utterance includes the same password spoken by the speaker himself. Whether it is a thing is determined based on the speaker template. Therefore, the quality of the speaker template is very important for the entire authentication process.

  In speaker verification systems based on DTW, many features are required at the input of each frame in order to obtain reliable performance. In general, these features are heard from all speakers in the same way, and each speaker's features are ignored. Several schemes have been proposed to customize the optimal feature set for each speaker by choosing an appropriate feature subset from the acoustic feature set. By this method, the verification performance can be improved, and the memory required for the template can be reduced. However, effective criteria for feature selection is a challenge, especially when available information is limited.

  Known optimization methods can be defined in terms of two components: performance criteria and search methods. For the first component, the normal performance criteria are performance in the imposter database, eg “Speaker recognition using discriminative features selection” in B. Sabac (2002) ICSLP-2002, pp.2321-2324 Request the False Accept Rate used as a reference. That is, it is necessary to test the performance of different feature subsets with many client and spoiler trials to find the best one. However, in speaker verification systems that allow password selection, spoofer data is rarely available.

K. Yu, J. Mason, J. Oglesby, “Speaker recognition using hidden Markov models, dynamic time warping and vector quantisation” (Vision, Image and Signal Processing, IEE Proceedings, Vol. 142, Oct. 1995, pp. 313- 18). “Speaker recognition using discriminative features selection” in ICSLP-2002, pp. 2321-2324.

  To solve the problems of the prior art, the present invention provides a method and apparatus for registration of speaker authentication, a method and apparatus for verification of speaker authentication, and a speaker authentication system.

  According to one aspect of the present invention, generating a plurality of acoustic feature vector sequences based on each of a plurality of utterances having the same content spoken by a speaker, and generating a reference template from the plurality of acoustic feature vector sequences And, based on a codebook including a plurality of codes and feature vectors corresponding to the plurality of codes, generating a pseudo-spoofer feature vector sequence corresponding to each of the plurality of acoustic feature vector sequences; Selecting an optimal acoustic feature subset based on a plurality of acoustic feature vector sequences, the reference template and the plurality of pseudo-spoofer feature vector sequences, and a method for enrollment of speaker authentication is provided. .

  According to another aspect of the invention, based on the optimal acoustic feature subset generated during registration to generate a test acoustic feature vector sequence from the test utterance and to obtain an optimal test acoustic feature vector sequence. Optimizing the test acoustic feature vector sequence; and determining whether the test utterance is a registered utterance spoken by the same speaker based on a reference template and the optimal test acoustic feature vector sequence. A method for verification of speaker authentication is provided.

  According to another aspect of the present invention, an acoustic feature extraction unit that generates an acoustic feature vector sequence from an utterance spoken by a speaker, and a plurality of acoustic features corresponding to a plurality of utterances of the same content spoken by the speaker Based on a template generation unit that generates a reference template from a vector sequence, and a code book that includes a plurality of codes and feature vectors corresponding to the plurality of codes, a pseudo-fake person feature corresponding to each of the plurality of acoustic feature vector sequences A pseudo-spoofer data generation unit that generates a vector sequence; and an optimization unit that selects an optimal acoustic feature subset based on the plurality of acoustic feature vector sequences, the reference template, and the plurality of pseudo-spoofer feature vector sequences; , An apparatus for registration of speaker authentication is provided.

  According to another aspect of the present invention, a test acoustic feature extraction unit that generates a test acoustic feature vector sequence from a test utterance, and an optimal acoustic feature generated during registration to obtain an optimal test acoustic feature vector sequence A test optimization unit for optimizing the test acoustic feature vector sequence based on a subset; and whether a test utterance is a registered utterance spoken by the same speaker based on a reference template and the optimal test acoustic feature vector sequence An apparatus for verifying speaker authentication, comprising: a determination unit for determining;

  According to another aspect of the present invention, there is provided a speaker authentication system comprising the above apparatus for registration of speaker authentication and the above apparatus for verification of speaker authentication.

FIG. 6 is a flowchart illustrating a method for enrollment of speaker authentication according to an embodiment of the present invention. FIG. 5 is a flowchart illustrating a method for enrollment of speaker authentication according to another embodiment of the present invention. FIG. 5 is a flowchart illustrating a method for verification of speaker authentication according to an embodiment of the present invention. 1 is a block diagram of an apparatus for registration of speaker authentication according to an embodiment of the present invention. FIG. 1 is a block diagram of an apparatus for verification of speaker authentication according to an embodiment of the present invention. 1 is a block diagram of a speaker authentication system according to an embodiment of the present invention.

  The above features, advantages and objects of the present invention will be better understood from the following description of embodiments with reference to the drawings.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the drawings.
FIG. 1 is a flowchart of a method for enrollment of speaker authentication according to an embodiment of the present invention. As shown in FIG. 1, first, in step 101, a plurality of acoustic feature vector sequences is generated based on each of a plurality of utterances having the same content spoken by a speaker. Each of the acoustic feature vectors may include a plurality of acoustic features indicating an utterance in the form of, for example, MFCC (Mel Frequency Cepstrum Coefficient). However, the present invention is not particularly limited to this, and the acoustic features of the speech include LPCC (Linear Predictive Cepstrum Coefficient), energy, fundamental frequency or wavelet analysis, pitch and duration, and their first derivative on the time axis. And other known and future modes obtained on the basis of information such as the second derivative and the like. All features that may be suitable for speaker recognition can be merged to function as a feature set. Then, using the feature selection method described below, the optimal feature subset can be customized for the speaker by selecting multiple features from the feature set automatically set for each speaker's characteristics during the registration process. Is done.

  Next, in step 105, a reference template is generated from a plurality of acoustic feature vector sequences. For example, a first acoustic feature vector sequence is first selected as an initial template, and then a second acoustic feature vector sequence is generated in time alignment with the first acoustic feature vector by using the DTW method. A new template is generated by using the average of the corresponding feature vectors in the two acoustic feature vector series. A third acoustic feature vector sequence is then generated to be temporally aligned with the new template, and such a cycle is performed until all acoustic feature vector sequences are merged into independent templates (called template merging). Is implemented. For details, see “Cross-words reference template for DTW-based speech recognition systems” by W. H. Abdulla, D. Chow and G. Sin (IEEE TENCON 2003, pp.1576-1579). In the present invention, the mode for generating the reference template is not particularly limited.

  Next, in step 110, corresponding pseudo-impostor feature vector sequences for a plurality of acoustic feature vector sequences are generated based on the codebook. The code book used in the present embodiment is a code book trained in the acoustic space of the entire application. For example, for a Chinese application environment, the codebook needs to cover the acoustic space for Chinese utterances, and for the English application environment, the codebook needs to cover the acoustic space for English utterances. Of course, for some special application environments, the acoustic space covered by the codebook can be changed accordingly.

  The code book of this embodiment includes a large number of codes and feature vectors corresponding to the large numbers of codes. The number of codes depends on the size of the acoustic space, the required compression ratio, and the required compression quality. The larger the acoustic space, the more codes are needed. Under the same acoustic space conditions, the smaller the number of chords, the higher the compression ratio, and the larger the number of chords, the higher the quality of the compressed template. According to a preferred embodiment of the present invention, the number of codes is preferably 256-512 under the acoustic space of normal Chinese utterances. Of course, it is possible to appropriately adjust the number of codes and the covered acoustic space of the code book according to different needs.

  Specifically, in order to convert the acoustic feature vector sequence into a corresponding code sequence in this step, a code is designated for each feature vector of the acoustic feature vector sequence. For example, a feature vector close to the feature vector of the acoustic feature vector sequence can be found by calculating a distance such as the Euclidean distance between the feature vector of the acoustic feature vector sequence and each feature vector in the codebook. . The code corresponding to the closest feature vector in the code book is designated as the feature vector of the acoustic feature vector series.

  Next, the code sequence is converted into a corresponding feature vector sequence based on codebook codes and feature vectors corresponding to those codes as a pseudo-spoofer feature vector sequence.

  Next, in step 115, an optimal acoustic feature subset is selected based on the plurality of acoustic feature vector sequences, the reference template, and the plurality of pseudo impersonator feature vector sequences. Specifically, individual acoustic feature subset candidates are examined, and the acoustic feature subset that maximizes the identification rate of the reference template for the pseudo-impostor feature vector sequence is selected as the optimal acoustic feature subset.

  According to an embodiment of the present invention, according to individual acoustic feature subset candidates, a plurality of DTW distances de (i) (referred to as speaker distance) between a reference template and a plurality of acoustic feature vector sequences, and a reference template And DTW distance dp (i) (referred to as inter-speaker distance) between each of the pseudo-fake person feature vector series. Minimize the ratio between the DTW distance de (i) between the reference template and the plurality of acoustic feature vector sequences and the DTW distance dp (i) between the reference template and the plurality of pseudo-impersonator feature vector sequences The acoustic feature subset to be selected is selected as the optimal acoustic feature subset.

  According to another embodiment of the present invention, a plurality of DTW distances de (i) between a reference template and a plurality of acoustic feature vector sequences and a reference template and a plurality of pseudo-impersonators according to individual acoustic feature subset candidates. A plurality of DTW distances dp (i) between the feature vector series are respectively calculated. A difference between a plurality of DTW distances de (i) between the reference template and a plurality of acoustic feature vector sequences and a plurality of DTW distances dp (i) between the reference template and a plurality of pseudo-impersonator feature vector sequences, and reference A ratio of a plurality of DTW distances de (i) between the template and the plurality of acoustic feature vector sequences and a sum of a plurality of DTW distances dp (i) between the reference template and the plurality of pseudo-impersonator feature vector sequences. The acoustic feature subset that minimizes is selected as the optimal acoustic feature subset.

  Further in accordance with an embodiment of the present invention, the step of examining individual acoustic feature subset candidates is performed in a specific range. For example, the specific range includes acoustic feature subset candidates in which the number of acoustic features is greater than the specific number.

  As can be seen from the above description, the method for enrollment of speaker authentication according to the present embodiment can select an optimal acoustic feature subset when there is no impersonator database, and is extracted by this. Features are more discriminating. Using this technique, text-dependent speaker verification systems achieve significant improvements.

  FIG. 2 is a flowchart of a method for enrollment of speaker authentication according to another embodiment of the present invention. This embodiment will be described with reference to the drawings. The description of the same parts as those of the previous embodiment will be omitted as appropriate.

  As shown in FIG. 2, steps 101-115 of the method for enrollment of speaker authentication according to this embodiment are the same as the embodiment shown in FIG. 1, and will not be repeated again here.

  Compared to the previously described embodiment, this embodiment further includes a step 220 for compression of the reference template. Specifically, it may include compressing the dimensions of the acoustic feature vectors in the reference template based on the optimal acoustic feature subset, or compressing the number of acoustic feature vectors in the reference template based on the codebook. You may include that.

  For a mode for compressing the number of acoustic feature vectors in a reference template based on a codebook, please refer to Chinese Patent Application No. 2005110115300.5 (title “Speaker”) filed by the present applicant on November 11, 2005. See "Apparatus and method for compressing and synthesizing templates and speaker authentication"). A plurality of adjacent features having the same specific code in the reference template are replaced with feature vectors. For example, an average vector of the set of certain adjacent feature vectors having the same code is first calculated. The calculated average vector is then used to replace the set of adjacent feature vectors with the same code.

  If there are multiple sets of such multiple adjacent features with the same code in the reference template, they can be exchanged one by one in the above mode. Therefore, the plurality of feature vectors are replaced one by one with one feature vector, and the number of feature vectors in the reference template is reduced. Thus, the template is compressed.

  As can be seen from the above description, the method for enrollment of speaker authentication according to the present embodiment can not only select an optimal acoustic feature subset when there is no impersonator database, but also a reference template. Can be compressed. Therefore, the required memory for the template is reduced. Also, the calculation load is reduced. This technique is used to achieve significant improvements in text-dependent speaker verification systems.

  FIG. 3 is a flowchart showing a method for verifying speaker authentication according to an embodiment of the present invention under the same inventive concept. Hereinafter, the present embodiment will be described with reference to FIG. The description of the same parts as those of the previous embodiment will be omitted as appropriate.

  As shown in FIG. 3, first, in step 301, a test acoustic feature vector series is generated according to a test utterance. Since it is similar to step 101 of FIG. 1 described above, the acoustic feature vector can include a plurality of acoustic features each indicating an utterance in the form of, for example, MFCC (Mel Frequency Cepstrum Coefficient). However, the present invention is not particularly limited to this, and the acoustic features of the speech include LPCC (Linear Predictive Cepstrum Coefficient), energy, fundamental frequency or wavelet analysis, pitch and duration, and their first derivative on the time axis. And other known and future modes obtained on the basis of information such as the second derivative and the like. All features that may be suitable for speaker recognition can be merged to function as a feature set.

  Next, in step 305, the test acoustic feature vector sequence is optimized based on the optimal acoustic feature subset generated during registration to obtain an optimal test acoustic feature vector sequence. The method of selecting the optimal acoustic feature subset has been described in the previous embodiment and will not be repeated here again.

  Next, in step 310, it is determined whether the test utterance is a registered utterance spoken by the same speaker based on the reference template and the optimal test acoustic feature vector sequence. Specifically, for example, a DTW matching score is first calculated between the reference template and the optimal test acoustic feature vector sequence, and then the DTW is determined to determine whether the input utterance is a registered utterance spoken by the same speaker. The matching score is compared with a threshold value.

  In the present invention, all known and future modes for calculating the DTW matching score between the reference template and the optimal test acoustic feature vector sequence are applicable. In the present invention, all known and future modes for setting a threshold value for identification can be applied.

  As can be seen from the above description, the method for confirmation of speaker authentication according to the present embodiment can use the optimal acoustic feature set selected in the registration stage, and is extracted by the method. The features are more discriminating and the system achieves significant improvements.

  FIG. 4 is a block diagram of an apparatus for enrollment of speaker authentication according to an embodiment of the present invention under the same inventive concept. Hereinafter, this embodiment will be described with reference to FIG. The description of the same parts as those of the previous embodiment will be omitted as appropriate.

  As shown in FIG. 4, the apparatus 400 for registration of speaker authentication includes an acoustic feature extraction unit 401 that generates an acoustic feature vector sequence from an utterance spoken by the speaker, and the same content spoken by the speaker. Based on a template generation unit 402 that generates a reference template from a plurality of acoustic feature vector sequences corresponding to a plurality of utterances, and a code book 704 including a plurality of codes and feature vectors corresponding to the plurality of codes. A pseudo-spoofer data generation unit 403 that generates a pseudo-spoofer feature vector sequence corresponding to each of the acoustic feature vector sequences of the plurality of acoustic feature vector sequences, Based on this, an optimization unit 405 for selecting an optimal acoustic feature subset is provided.

  The pseudo-impersonator data generation unit 403 includes a conversion unit vector-code conversion unit 4031 that specifies a code for each feature vector in the acoustic feature vector sequence in order to convert the acoustic feature vector sequence into a corresponding code sequence; Based on the codes and the feature vectors corresponding to those codes, a conversion unit code-vector conversion unit 4032 for converting the code sequence into the corresponding feature vector sequence as the pseudo-impersonator feature vector sequence is included.

  According to the embodiment of the present invention, the conversion unit vector-code conversion unit 4031 searches a code book for a feature vector closest to the feature vector of the acoustic feature vector sequence, and adds the feature vector of the acoustic feature vector sequence to the feature vector. A code corresponding to the closest feature vector in the code book is designated.

  According to an exemplary embodiment of the present invention, the optimization unit 405 examines individual acoustic feature subset candidates, and determines the acoustic feature subset that maximizes the identification rate of the reference template for the pseudo-fake person feature vector sequence as the optimal feature set. It is configured to select as the acoustic feature subset.

  According to another embodiment of the present invention, the optimization unit 405 examines individual acoustic feature subset candidates, and includes a plurality of DTW distances de (i) () between the reference template and the plurality of acoustic feature vector sequences. And a plurality of DTW distances dp (i) (referred to as inter-speaker distances) between the reference template and the plurality of pseudo impersonator feature vector sequences, and calculating the plurality of DTWs. The acoustic feature set that minimizes the ratio of the plurality of DTW distances de (i) to the distance dp (i) is selected as the optimal acoustic feature subset.

  According to another embodiment of the present invention, the optimization unit 405 examines individual acoustic feature subset candidates, and a plurality of DTW distances de (i) between the reference template and the plurality of acoustic feature vector sequences, And calculating a plurality of DTW distances dp (i) between the reference template and the plurality of pseudo impersonator feature vector sequences, and the plurality of DTW distances de (i) and the plurality of DTW distances dp (i) And the acoustic feature set that minimizes the ratio of the sum of the difference between the plurality of DTW distances de (i) and the plurality of DTW distances dp (i) as the optimal acoustic feature subset. Yes.

  According to another embodiment of the invention, the optimizer 405 is configured to examine the individual acoustic feature subset candidates in a specific range. For example, the specific range includes acoustic feature subset candidates in which the number of acoustic features is larger than the specific number.

  As shown in FIG. 4, the apparatus 400 for enrollment of speaker authentication further includes a compression unit 406 that compresses the dimension of the acoustic feature vector in the reference template based on the optimal acoustic feature subset.

  According to another embodiment of the present invention, the compression unit 406 is further configured to compress the number of acoustic feature vectors in the reference template based on a codebook.

  As can be seen from the above description, the apparatus for registration of speaker authentication according to the present embodiment can implement the method for registration of speaker authentication described in the previous embodiment in implementation. In addition, it is possible to select an optimal acoustic feature subset when there is no impersonator database, and the extracted features are more discriminating. Using this technique, text-dependent speaker verification systems achieve significant improvements.

  Furthermore, the reference template is correspondingly compressed by reducing the necessary memory for the template. Also, the calculation load is reduced.

  FIG. 5 is a block diagram of an apparatus for verification of speaker authentication according to an embodiment of the present invention under the same inventive concept. Hereinafter, this embodiment will be described with reference to FIG. The description of the same parts as those of the previous embodiment will be omitted as appropriate.

  As shown in FIG. 5, an apparatus 500 for verifying speaker authentication according to the present embodiment includes a test acoustic feature extraction unit 501 that generates a test acoustic feature vector sequence from a test utterance, and an optimal test acoustic feature vector. A test optimization unit 502 for optimizing the test acoustic feature vector sequence based on an optimal acoustic feature subset generated during registration to obtain a sequence, and a reference template and the optimal test acoustic feature vector sequence The determination unit 503 determines whether the test utterance is a registered utterance spoken by the same speaker.

  The determination unit includes a DTW calculation unit 5031 that calculates a DTW matching score between the reference template and the optimal test acoustic feature vector sequence. The determination unit 503 compares the DTW matching score with a threshold value to determine whether the test utterance is a registered utterance spoken by the same speaker.

  As can be seen from the above description, the apparatus for verification of speaker authentication according to the present embodiment can implement the method for verification of speaker authentication described in the previous embodiment in implementation. Also, by using the optimal acoustic feature subset selected during the registration stage, it is possible to select the optimal feature set that gives the characteristics of each speaker. Thus, the extracted features are more discriminatory and the system achieves significant improvements. Furthermore, the reference template is correspondingly compressed by reducing the necessary memory for the template. Also, the calculation load is reduced.

  FIG. 6 is a block diagram of a speaker authentication system according to an embodiment of the present invention under the same inventive concept. Hereinafter, the present embodiment will be described with reference to the drawings. The description of the same parts as those of the previous embodiment will be omitted as appropriate.

  As shown in FIG. 6, the system for speaker authentication according to the present embodiment includes a registration device 400 that can be a device for registration of speaker authentication described in the previous embodiment, and the previous embodiment. It includes a verification device 500 that can be a device for verification of the described speaker authentication. The reference template and the optimal feature subset generated by the registration device 400 are transferred to the verification device 500 by any communication mode such as a network, an internal channel, or a recording medium such as a magnetic disk.

  As can be seen from the above description, the system for speaker authentication according to the present embodiment selects the optimal acoustic feature subset when there is no impersonator database at the registration stage, and is optimal according to the characteristics of each speaker. It is possible to select an optimal feature set according to the characteristics of each speaker by using the optimal feature set selected at the registration stage in the verification stage. Thus, the extracted features are more discriminatory and the system achieves significant improvements. Furthermore, the reference template is correspondingly compressed by reducing the required memory for the template. Also, the calculation load is reduced.

  The device 400 for enrollment of speaker authentication, the device 500 for verification of speaker authentication and their various components can be composed of dedicated lines or chips, and can also be a computer (processor). It can also be implemented by executing the corresponding program.

  On the other hand, referring to some exemplary embodiments of the method and apparatus for enrollment of speaker authentication, the method and apparatus for verification of speaker authentication, and the system for speaker authentication according to the present invention. However, these embodiments are not comprehensive, and those skilled in the art can make various changes and modifications within the spirit and scope of the present invention. Accordingly, the invention is not limited to these embodiments, and the scope of the invention is defined by the appended claims.

Claims (25)

Generating a plurality of acoustic feature vector sequences based on each of a plurality of utterances of the same content spoken by a speaker;
Generating a reference template from the plurality of acoustic feature vector sequences;
Generating a pseudo-impersonator feature vector sequence corresponding to each of the plurality of acoustic feature vector sequences based on a codebook including a plurality of codes and feature vectors corresponding to the plurality of codes;
Selecting an optimal acoustic feature subset based on the plurality of acoustic feature vector sequences, the reference template, and the plurality of pseudo impersonator feature vector sequences;
A method for enrollment of speaker authentication, comprising: Generating a pseudo-spoofer feature vector sequence corresponding to each of the plurality of acoustic feature vector sequences,
Designating a code for each feature vector in the acoustic feature vector sequence to convert the acoustic feature vector sequence to a corresponding code sequence;
Converting the code sequence into a corresponding feature vector sequence as the pseudo-spoofer feature vector sequence based on the code and the feature vector corresponding to the code;
The method of claim 1, comprising: Specifying a code for each feature vector in the acoustic feature vector sequence includes:
Searching the codebook for a feature vector closest to the feature vector of the acoustic feature vector sequence;
Designating a code corresponding to the closest feature vector in the codebook to the feature vector of the acoustic feature vector sequence;
The method of claim 2 comprising: Selecting the optimal acoustic feature subset comprises:
Examining individual acoustic feature subset candidates;
Selecting the acoustic feature subset that maximizes the identification rate of the reference template for the pseudo-impersonator feature vector sequence as the optimal acoustic feature subset;
The method of claim 1, comprising: Selecting the optimal acoustic feature subset comprises:
Examining individual acoustic feature subset candidates;
A plurality of DTW distances de (i) between the reference template and the plurality of acoustic feature vector sequences, and a plurality of DTW distances dp (i) between the reference template and the plurality of pseudo impersonator feature vector sequences. A step of calculating
Selecting the acoustic feature set that minimizes the ratio of the plurality of DTW distances de (i) to the plurality of DTW distances dp (i) as the optimal acoustic feature subset;
The method of claim 1, comprising: Selecting the optimal acoustic feature subset comprises:
Examining individual acoustic feature subset candidates;
A plurality of DTW distances de (i) between the reference template and the plurality of acoustic feature vector sequences, and a plurality of DTW distances dp (i) between the reference template and the plurality of pseudo impersonator feature vector sequences. A step of calculating
Minimizing the ratio of the difference between the plurality of DTW distances de (i) and the plurality of DTW distances dp (i) to the sum of the plurality of DTW distances de (i) and the plurality of DTW distances dp (i) Selecting the acoustic feature set to be the optimal acoustic feature subset;
The method of claim 1, comprising:   The method according to any one of claims 4 to 6, wherein the step of examining the candidates of the individual acoustic feature subsets is performed in a specific range.   8. The method of claim 7, wherein the specific range includes acoustic feature subset candidates in which the number of acoustic features is greater than the specific number.   The method of claim 1, further comprising compressing dimensions of acoustic feature vectors in the reference template based on the optimal acoustic feature subset.   The method of claim 1, further comprising compressing a number of acoustic feature vectors in the reference template based on the codebook. Generating a test acoustic feature vector sequence from the test utterance;
Optimizing the test acoustic feature vector sequence based on the optimal acoustic feature subset generated during registration to obtain an optimal test acoustic feature vector sequence;
Determining whether the test utterance is a registered utterance spoken by the same speaker based on a reference template and the optimal test acoustic feature vector sequence;
A method for verification of speaker authentication, comprising: The step of determining includes
Calculating a DTW matching score between the reference template and the optimal test acoustic feature vector sequence;
Comparing the DTW matching score to a threshold value to determine if the test utterance is a registered utterance spoken by the same speaker;
The method of claim 11, comprising: An acoustic feature extraction unit that generates an acoustic feature vector sequence from an utterance spoken by a speaker;
A template generation unit that generates a reference template from a plurality of acoustic feature vector sequences corresponding to a plurality of utterances of the same content spoken by a speaker;
Based on a code book including a plurality of codes and feature vectors corresponding to the plurality of codes, a pseudo-imposed person data generation unit that generates a pseudo-imposed person feature vector series corresponding to each of the plurality of acoustic feature vector series;
An optimization unit that selects an optimal acoustic feature subset based on the plurality of acoustic feature vector sequences, the reference template, and the plurality of pseudo impersonator feature vector sequences;
An apparatus for registration of speaker authentication, comprising: The pseudo impersonator data generation unit
A conversion unit vector-code conversion unit that designates a code for each feature vector in the acoustic feature vector sequence in order to convert the acoustic feature vector sequence into a corresponding code sequence;
Based on the code and a feature vector corresponding to the code, a conversion unit code-vector conversion unit that converts the code sequence into a corresponding feature vector sequence as the pseudo-impersonator feature vector sequence;
14. The apparatus of claim 13, comprising:   The conversion unit vector-code conversion unit searches a feature vector closest to the feature vector of the acoustic feature vector sequence from a code book, and the feature closest to the feature vector of the acoustic feature vector sequence in the code book. The apparatus of claim 14, wherein the apparatus is configured to specify a code corresponding to a vector.   The optimization unit examines individual acoustic feature subset candidates, and selects the acoustic feature subset that maximizes the identification rate of the reference template with respect to the pseudo-impersonator feature vector sequence as the optimal acoustic feature subset. 14. The apparatus of claim 13, wherein the apparatus is configured.   The optimization unit examines individual acoustic feature subset candidates, a plurality of DTW distances de (i) between the reference template and the plurality of acoustic feature vector sequences, and the reference template and the plurality of pseudo-falsifications. The acoustic feature set that calculates a plurality of DTW distances dp (i) between the person feature vector series and minimizes a ratio of the plurality of DTW distances de (i) and the plurality of DTW distances dp (i) 14. The apparatus of claim 13, wherein the apparatus is configured to select as the optimal acoustic feature subset.   The optimization unit examines individual acoustic feature subset candidates, a plurality of DTW distances de (i) between the reference template and the plurality of acoustic feature vector sequences, and the reference template and the plurality of pseudo-falsifications. Calculating a plurality of DTW distances dp (i) between the plurality of DTW distances de (i) and the plurality of DTW distances dp (i). The configuration of claim 13, wherein the acoustic feature set that minimizes the ratio of the difference between i) and the plurality of DTW distances dp (i) is selected as the optimal acoustic feature subset. The device described.   The apparatus according to any one of claims 16 to 18, wherein the optimization unit is configured to examine candidates of the individual acoustic feature subset in a specific range.   The apparatus of claim 19, wherein the specific range includes acoustic feature subset candidates in which the number of acoustic features is greater than the specific number.   The apparatus of claim 13, further comprising a compression unit that compresses dimensions of acoustic feature vectors in the reference template based on the optimal acoustic feature subset.   The apparatus of claim 13, further comprising a compression unit that compresses the number of acoustic feature vectors in the reference template based on the codebook. A test acoustic feature extraction unit that generates a test acoustic feature vector sequence from the test utterance;
A test optimization unit for optimizing the test acoustic feature vector sequence based on an optimal acoustic feature subset generated during registration to obtain an optimal test acoustic feature vector sequence;
A determination unit that determines whether the test utterance is a registered utterance spoken by the same speaker based on a reference template and the optimal test acoustic feature vector sequence;
An apparatus for verification of speaker authentication, comprising: The determination unit
A DTW calculator that calculates a DTW matching score between the reference template and the optimal test acoustic feature vector sequence;
A determination unit that compares the DTW matching score with a threshold value to determine whether the test utterance is a registered utterance spoken by the same speaker;
24. The apparatus of claim 23, comprising: An apparatus according to any one of claims 13 to 22,
An apparatus according to claim 23 or claim 24;
A system for speaker authentication, comprising:

Images