JP2006078654A - Voice authenticating system, method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice authenticating system for markedly improving noise resistance and reliability, and provide a method and a program. <P>SOLUTION: The voice authenticating system comprises outputting to successively generate prescribed characteristic parameters that represent the individuality of a speaker on the basis of an input voice signal, detecting a sound-existing section and a voice-existing section of the voice signal respectively, and extracting the characteristic parameters of the sound-existing section and the voice-existing section in the voice signal from the characteristic parameters generated. Speaker authentication is conducted on the basis of the extracted characteristic parameters. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a voice authentication apparatus, method, and program, and is suitable for application to a voice authentication apparatus that performs speaker authentication based on parameters obtained by encoding processing using a CELP (Code Excited Linear Prediction) encoding method, for example. Is.

  In recent years, with the expansion and penetration of services using networks such as electronic commerce and online banking, the need for personal authentication to confirm the legitimacy of users is increasing from the viewpoint of information security and privacy protection. Under such circumstances, conventionally, as a simple and inexpensive personal authentication method, personal authentication based on the possession and knowledge of the user such as an IC card and a password has been performed.

  However, according to such a conventional method, there is a problem accompanied by danger such as theft, loss or forgetting. In recent years, so-called biometric authentication using personal physical or behavioral features such as speech, handwriting, fingerprints or facial expressions has attracted attention.

For example, it is known that phonological information and personality information coexist in speech, phonological information can be used for speech recognition, and personality information can be used for speaker recognition. In digital voice communication using a mobile phone, voice is encoded by the CELP encoding method for transmission, but phonetic information and personality information are stored in the encoded voice. . Therefore, it has been conventionally proposed to perform personal authentication using a parameter called LSP (Line Spectrum Pair) obtained by such encoding (see Non-Patent Document 1, for example).
Takefumi Mogaki and Naohisa Komatsu, “Text-speaking speaker verification method using CELP method”, Computer Security Symposium '98, October 1998, p201-206

  However, personal authentication using personality information contained in speech (hereinafter referred to as speech authentication or speaker authentication) is easily affected by environmental noise, and the accuracy of authentication is significantly degraded under conditions of high environmental noise. To do.

  In practice, ideal speech that does not contain noise recorded in the research ATR Japanese speech database is CELP encoded, and speaker authentication is performed using the resulting LSP. It is obtained by adding CELP coding processing to the speech by adding noise of “car”, “crowd”, “station” or “street” recorded in the database. An experiment for evaluating the reliability of speaker authentication under a noisy environment by performing speaker authentication using the obtained LSP, an experimental result as shown in FIG. 15 was obtained.

  As is apparent from FIG. 15, the error rate when no noise is added is 5.13 [%], whereas the error rate when the “car” noise is added is 30.80. [%], Error rate when adding “crowd” noise is 15.87 [%], error rate when adding “station” noise is 14.52 [%], “intersection ( The error rate is 9.47 [%] when the noise of “street” is added, and the reliability of speaker authentication is low in a noisy environment.

  In addition, as shown in FIG.

FRR (False Rejection Rate) defined by, and the following formula

This is performed using EER (Equal Error Rate), which is an error rate at a point where the value of FAR (False Acceptance Rate) is the same. The specifications of each noise data are shown in FIG.

  However, there may be various environments in which speaker authentication is actually used. Therefore, when constructing such a voice authentication device, it is desired to improve noise resistance so that practically sufficient authentication accuracy can be obtained even under various noise environments.

  The present invention has been made in view of the above points, and an object of the present invention is to propose a voice authentication apparatus, method, and program capable of significantly improving noise resistance and reliability.

  In order to solve such a problem, in the present invention, in the voice authentication apparatus, a feature parameter generating means for sequentially generating and outputting a predetermined feature parameter representing the personality of the speaker based on the input voice signal, and the voice signal Based on the detection results of the voiced section detecting means for detecting the voiced section of the voice signal, the voiced section detecting means for detecting the voiced section of the voice signal, and the voiced section detecting means and the voiced sound detecting means Feature parameter extracting means for extracting feature parameters of voiced and voiced sections in the speech signal from the feature parameters sequentially output from the means, and speaker authentication based on the feature parameters extracted by the feature parameter extracting means And a speaker authentication means.

  As a result, this voice authentication device can perform speaker registration and speaker verification using only feature parameters that exhibit stable characteristics against environmental noise, so environmental noise during speaker registration and speaker verification Can be made less susceptible to

  According to the present invention, in the voice authentication method, predetermined feature parameters representing the individuality of the speaker are sequentially generated and output based on the input voice signal, and the voiced and voiced sections of the voice signal are determined. Speaker authentication is performed based on the first step for detecting each, the second step for extracting the feature parameters of the voiced and voiced sections of the speech signal from the generated feature parameters, and the extracted feature parameters. And a third step to be performed.

  As a result, according to this voice authentication method, speaker registration and speaker verification can be performed using only feature parameters that show stable characteristics against environmental noise. Therefore, at the time of speaker registration or speaker verification It can be made less susceptible to environmental noise.

  Furthermore, in the present invention, the program sequentially generates and outputs predetermined characteristic parameters representing the personality of the speaker based on the voice signal input to the computer, and also includes the voiced and voiced sections of the voice signal. A second step of extracting the feature parameters of the voiced and voiced sections of the speech signal from the generated feature parameters, and speaker authentication based on the extracted feature parameters And a third step of performing the process.

  As a result, according to this program, speaker registration and speaker verification can be performed using only feature parameters that show stable characteristics against environmental noise. Can be made less susceptible to

  According to the present invention, in the voice authentication apparatus, method, and program, the predetermined feature parameters representing the personality of the speaker are sequentially generated and output based on the input voice signal, By detecting each voiced sound section, extracting the feature parameters of the voiced sound section and the voiced sound section from the generated feature parameters, and performing speaker authentication based on the extracted feature parameters, Speaker registration and speaker verification using only feature parameters that show stable characteristics against environmental noise can be performed, making it less susceptible to environmental noise during speaker registration and speaker verification. Thus, it is possible to realize a voice authentication apparatus, method, and program capable of significantly improving noise resistance and reliability.

  Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

(1) First Embodiment (1-1) Principle As one of the causes of deterioration in authentication accuracy of speaker authentication under a noisy environment, there is no voice, that is, a section that does not include speaker personality information Is used for speaker authentication.

  Therefore, for example, in a voice authentication apparatus that performs speaker authentication using voice information encoded by the CELP encoding method, only LSPs of frames that show stable characteristics against environmental noise are selectively extracted. It is considered that registration processing and verification processing using LSP is effective as noise-proof processing.

  In this case, in ITU / T (International Telecommunication Union / Telecommunication Standardization Sector), a CS-ACELP (Conjugate Structure-Algebraic CELP) encoding method which is a kind of CELP encoding method is standardized as G.729. In order to apply the CS-ACELP coding method to VoIP and the like, a method for compressing a silent section and reducing the bit rate is standardized as G.729 Annex B.

  In G.729 Annex B, a VAD (Voice Activity Decision) algorithm is adopted as an algorithm for detecting a silent section. Therefore, by extracting only the LSP of the frame of the voice section using this VAD algorithm and performing speaker registration and speaker verification using this LSP, it is considered that it can be made less susceptible to environmental noise. It is done.

  On the other hand, the pitch (Pitch delay), which is one of the parameters obtained by the encoding process using the CS-ACELP encoding method, represents the fundamental period of the excitation signal, and is a parameter corresponding to the level of the sound in the voiced sound. is there. Voiced sound has stronger power than unvoiced sound, and LSP of voiced sound can stably express the characteristics of the vocal tract of the speaker. Therefore, it can be expected that a stable LSP with little change can be obtained even in a noisy environment by extracting the LSP of the frame of the voiced sound section using the pitch.

Here, the applicant of the present invention investigated the behavior of voiced / silent frames by the VAD algorithm and the relationship between the pitch and voiced / unvoiced frames.
(1) A frame with a stable pitch has a small change in LSP due to noise.
(2) The pitch is hardly stable in the unvoiced sound frame, but stable in the voiced sound frame.
(3) The fact that the pitch may be stable in a part of the silent frame was revealed.

  Therefore, from these facts, it is considered that a frame of voiced sound can be extracted with high accuracy by combining the VAD algorithm and a technique for extracting a stable section of pitch. That is, by selecting a frame within a section determined as a voiced section by the VAD algorithm and determined as a voiced section based on the pitch, only the LSP of a frame showing a stable characteristic against environmental noise is extracted. By performing speaker registration and speaker verification using this LSP, it is considered that the influence of environmental noise can be reduced and speaker authentication with high reliability can be performed. Hereinafter, a voice authentication device to which such a principle is applied will be described.

(1-2) Configuration of Voice Authentication Device According to First Embodiment In FIG. 1, reference numeral 10 denotes a voice authentication device according to this embodiment as a whole, and a speaker's voice is collected by a microphone (not shown). The speech signal S1 obtained by doing so is input to the CELP encoding unit 2.

  The CELP encoding unit 2 performs encoding processing by the CS-ACELP encoding method on the supplied speech signal S1, and performs frame selection processing on the quantized LSP for each obtained frame (10 [ms]) Send to part 3. The CELP encoding unit 2 sends the pitch (pitch delay) obtained for each subframe (5 [ms]) by the CS-ACELP encoding process to the voiced / unvoiced sound determination unit 4 as pitch information.

  Based on the supplied pitch information, the voiced / unvoiced sound determination unit 4 determines, for each subframe, whether the target frame is a voiced sound section or an unvoiced sound section.

  Specifically, the voiced / unvoiced sound determination unit 4 first extracts pitches of N (for example, 10) points centered on the target frame in accordance with the voiced / unvoiced sound determination processing procedure RT1 shown in FIG. SP1) Thereafter, the extracted 10 pitch values are approximated by a straight line by the least square method, and the mean square error D at this time is calculated (step SP2).

Next, the voiced / unvoiced sound determination unit 4 determines that the mean square error D obtained in this way is less than a first threshold value _Dth that is predetermined for the mean square error D, and the pitch at which attention is paid at that time. Is determined to be less than a predetermined second threshold _{Pth for} the value P (step SP3).

  When the voiced / unvoiced sound determination unit 4 obtains an affirmative result as a result of this determination, it determines that the frame is in a voiced sound section with a stable pitch (step SP4) and obtains a negative result. If this is the case, it is determined that the frame is in an unvoiced sound section whose pitch is not stable (step SP5), and this determination result is sent to the frame selection processing unit as a voiced / unvoiced sound determination signal S2 (step SP5). SP6).

  Furthermore, the voiced / unvoiced sound determination unit 4 thereafter repeats the same processing for each subframe (step SP1 to step SP6). In this manner, the voiced / unvoiced sound determination unit 4 determines, as the voiced / unvoiced sound determination signal S2, the determination result as to whether the target frame is a voiced sound section or an unvoiced sound section for each subframe. Sequentially sent to the frame selection processing unit 3.

  On the other hand, the sound signal S <b> 1 from the microphone is also given to the sound / silence determination unit 5. The voice / silence determination unit 5 determines whether the frame is a voiced segment or a voiceless segment for each frame using the VAD algorithm for the supplied audio signal S1.

  Actually, the voice / silence determination unit 5 firstly performs full-band energy, low-band energy, and zero crossing every frame (10 [ms]) according to the voice / silence determination processing procedure RT2 by the VAD algorithm shown in FIG. Find the rate. In addition, the sound / silence determination unit 5 calculates an LSP in the same manner as in the CELP encoding unit 2 described later, and calculates the spectrum of this LSP (step SP11).

  Then, the sound / silence determination unit 5 obtains the difference between the spectrum, the entire band energy, the low band energy, and the zero crossing rate obtained from these LSPs and the average values of these four parameters in the environmental noise described later. Thus, a difference parameter between the four parameters is generated (step SP12).

  Further, the sound / silence determination unit 5 determines whether or not the entire band energy obtained in step SP11 is 15 [dB] or less (step SP13). While the sound within is determined to be “noise” (step SP14), if a negative result is obtained, the sound within the frame is determined from the position in the four-dimensional vector space created by the four difference parameters. It is initially determined whether it is “speech” or “noise” (step SP15).

  Subsequently, the voice / silence determination unit 5 performs smoothing processing on the determination result based on the initial determination results for the previous several frames, and thereby determines whether the frame is “voice” or “noise”. Is finally determined (step SP16). As a result, for example, when the second frame and the previous frame are both “speech” and the initial determination of the target frame is “noise”, the energy difference from the previous frame is smaller than a certain threshold value. In the target frame, it is determined that the sound is continuous, and in the final determination, “noise” is changed to “speech”.

  Then, the sound / silence determination unit 5 determines “sound section” and “noise” when it determines “speech” based on the final determination result thus obtained. The determination result of “silent section” is sent to the frame selection processing unit 3 as a sound / silence determination signal S3 (step SP7).

  The sound / silence determination unit 5 determines whether or not the frame is finally determined to be “noise” (step SP18), and if a negative result is obtained, the process proceeds to the next frame ( Step SP11), if a positive result is obtained, the spectrum, full-band energy, low-band energy and zero-crossing rate obtained from the LSP extracted for the frame are used to obtain these four in the background noise. Each average value of the parameters is updated (step SP19).

  The sound / silence determination unit 5 thereafter repeats the same processing for each frame (step SP11 to step SP19). In this way, the sound / silence determination unit 5 sequentially sends the determination result of “sound period” or “silence period” to the frame selection processing unit 3 as the sound / silence determination signal S3 for each frame.

  Based on the supplied voiced / unvoiced sound determination signal S2 and voiced / silent sound determination signal S3, the frame selection processing unit 3 determines whether or not the frame is a sounded section and a voiced sound section for each frame. Judging. At this time, the frame selection processing unit 3 determines whether or not it is a voiced section obtained based on the voiced / silent judgment signal S3 in units of frames (10 [ms]), whereas the voiced / voiceless voice Since the determination result of whether or not it is a voiced sound section obtained based on the determination signal S2 is a subframe (5 [ms]) unit, as shown in FIG. 4, based on the determination result of the odd-numbered subframe. It is determined whether or not the frame is in a voiced sound section.

  Then, based on the determination result, the frame selection processing unit 3 selects only the LSPs of the frames determined to be the voiced and voiced segments among the LSPs of the frames sequentially supplied from the CELP encoding unit 2. Extract. In the registration mode, the frame selection unit 3 sends the LSP of the frames in the voiced and voiced intervals obtained in this way to the weighting processing unit 6.

  The weighting processing unit 6 adds a weight obtained by multiplying each LSP supplied from the frame selection processing unit 3 by an average value of the LSP (hereinafter referred to as an average LSP) by a predetermined weighting coefficient w. Apply processing. In this case, it is known that the average LSP of the speaker represents the individuality. By applying such weighting processing, the individuality is emphasized, and as shown in FIG. 5, the 10-dimensional Euclidean created by the LSP The distribution for each speaker in the space can be separated. Then, the weighting processing unit 6 sends the weighted LSP obtained in this way to the clustering unit 7.

  As shown in FIG. 6, the clustering unit 7 performs a clustering process based on the LBG + splitting algorithm on the LSP supplied from the frame selection processing unit 3.

Here, the LBG algorithm is a design algorithm that starts from an appropriate initial codebook C _N ⁽⁰⁾ and converges to a good codebook by repeatedly applying a division condition and a representative point condition to the learning sequence T. The clustering process by the LBG algorithm is performed according to the clustering process procedure RT3 shown in FIG.

That is, in the clustering process by the LBG algorithm, first, the number of dimensions is K, the number of levels is N, and N initial quantization representative vectors y ₁ ⁽⁰⁾ , y ₂ ⁽⁰⁾ ,..., Y _N ^{(0) are included.} A learning sequence T composed of an initial codebook C _N ⁽⁰⁾ , L K-dimensional learning vectors x ₁ ⁽⁰⁾ , x ₂ ⁽⁰⁾ ,..., X _N ⁽⁰⁾ , a convergence determination threshold ε, Are set in advance, m is set to “0”, and initial strain D ⁽⁻¹⁾ is set to infinity (m = 0, D ⁽⁻¹⁾ = ∞) (step SP1).

Subsequently, the average distortion D ^(m) is minimized under the codebook C _N ^(m) composed of the quantized representative vectors y ₁ ^(m) , y ₂ ^(m) ,..., Y _N ^(m). N regions _p ¹ to ^{_{(m), p 2 (m}} ), ......, defined division of the learning sequence T to ^{p N (m)} _P ^{N (m)} is after applying the divided condition. That is, the region corresponding to the quantization representative vector _{^{y i (m) P i (}} m) is given by a set of learning vector strain and y _{i ^(m)} among the N quantization representative vector is minimized It is done. In this way, L learning vectors are divided into N regions. Further, the average distortion D ^(m) that occurs when the learning vector belonging to each region is replaced with the quantized representative vector in that region is calculated (step SP2).

Next, the average strain D ^(m) thus determined is

(Step SP3), if not, the learning sequence T is divided into _N regions p ₁ ^(m) , p ₂ ^(m) ,..., P _N ^(m) . _N quantization representative vectors y ₁ ^(m) , y ₂ ^(m) ,..., Y _N that minimize the average distortion D ^(m) with respect to the learning sequence T under P _N ^(m). the codebook C _N consisting of ^(m) determined by applying the representative point condition. The center of gravity given as the average vector of the learning vectors T belonging to the region P _i ^(m) is set as a quantized representative vector y _i (step SP24). Further, m is incremented, C _N is set as the code book C _N ^(m) (step SP25), and the same processing is repeated.

Then, when the average strain D ^(m) eventually satisfies the equation (3), the processing is stopped, and the code book C _N ^(m) at this time is finally determined as the designed N-level code book (step SP26). ). As a result, an N-level codebook finally converged to an appropriate state can be obtained.

On the other hand, the quality of the code book designed by the LBG algorithm strongly depends on the selection of the initial code book C _N ⁽⁰⁾ and the learning sequence T. The initial codebook C _N ⁽⁰⁾ preferably covers the assumed input vector distribution range. A splitting algorithm is known as a method for generating an initial codebook that satisfies this condition to some extent.

The splitting algorithm, the quantization representative vector _y 1 of N-level codebook _{C N,} y 2, ......, following equation _{y N}

By dividing into two vectors in proximity with the minute vector δ as, quantization representative vector y _1, y _2, ......, y _2N consists 2N level of initial codebook C _{2N ⁽⁰⁾} is Is to be generated. By combining this splitting algorithm with the LBC algorithm, it is possible to design codebooks of 2, 4, 8,.

  Thus, the clustering unit 7 performs clustering processing on the weighted LSP supplied from the weighting processing unit 6 by the LBG + splitting algorithm that combines the LBC algorithm and the splitting algorithm, thereby giving the speaker The same 10-dimensional N (for example, 16) quantized representative vectors as the unique LSP are obtained, and these are stored in the storage unit 8 such as a flash memory or a hard disk as the feature code book CB (FIG. 6) of the speaker. And save.

  On the other hand, in the collation mode, the weighted LSP output from the weighting processing unit 6 is given to the vector quantization unit 10 of the collation unit 9. In the collation mode, the feature code book CB of the registered speaker as a target is read from the storage unit 8 and given to the vector quantization unit 10.

  Thus, as shown in FIG. 8, the vector quantization unit 10 vector-quantizes each LSP of the frames of the voiced and voiced segments sequentially given from the weighting processing unit 6 with the feature codebook CB, thereby The quantization error for each frame is calculated. That is, the vector quantization unit 10 sequentially calculates each distance between each LSP (10-dimensional vector) and 16 quantized representative vectors (10-dimensional vector) constituting the feature code block CB.

  Then, the vector quantization unit 10 determines the smallest one of the 16 distances obtained for the frame in this way (hereinafter referred to as the minimum quantization error) as the quantization error detection signal S4. 11 to send.

  The determination unit 11 compares the value of the minimum quantization error of each frame obtained based on the minimum quantization error detection signal S4 with a predetermined third threshold set in advance for the quantization error. This comparison is performed for the specified number of frames set in accordance with a predetermined time as the voice length of the verification voice. Then, the determination unit 11 counts the number of frames whose minimum quantization error is smaller than the third threshold, and when the number of frames is equal to or greater than a preset fourth threshold, the speaker at that time is determined. It is determined that the user is the person, and the determination result is output as a determination signal S5.

  A specific configuration of the CELP encoding unit 2 is shown in FIG. The CELP encoding unit 2 inputs the supplied audio signal S1 to the preprocessing unit 20 composed of a high-pass filter. Then, the preprocessing unit 20 performs a filtering process on the supplied audio signal S1 with a cut-off frequency of 140 [Hz], and the audio signal S10 obtained by removing the noise component thus obtained is converted into an LPC (Linear Predictor Coefficients). The data is input to the analysis unit 21 and the subtraction unit 22.

  The LPC analysis unit 21 determines the filter coefficient (LPC) of the synthesis filter for each frame by performing linear prediction analysis for the supplied speech signal S10 for each frame (10 [ms]), Set to the synthesis filter 23. The LPC analysis unit 21 thereafter converts each filter coefficient (LPC) thus obtained into LSPs, quantizes these LSPs, and then converts them into the frame selection processing unit 3 (see FIG. Send to 1).

  On the other hand, the synthesis filter 23 performs a filtering process with a filter coefficient (LPC) set at that time on the sound source signal S11 supplied from the addition unit 29 as described later for each subframe, thereby generating a speech signal of the synthesized speech. Is generated and sent to the subtractor 22.

  The subtractor 22 subtracts the synthesized speech signal S12 from the supplied speech signal S10 to obtain a distortion detection signal S13 representing the waveform distortion components of the input speech and synthesized speech, and sends this to the perceptual weighting unit 24. To do.

  At this time, the perceptual weighting unit 24 is provided with the filter coefficient (LPC) of the synthesis filter 23 at that time from the LPC analysis unit 21. Thus, the audibility weighting unit 24 performs a predetermined filtering process of the filter coefficient determined based on the filter coefficient (LPC) of the synthesis filter 23 so as to minimize the audible distortion with respect to the distortion detection signal S13. The perceptual weighting distortion detection signal S14 is sent to the pitch analysis unit 25 and the fixed codebook search unit 26.

The pitch analysis unit 25 determines the pitch L of the synthesized speech and the gain _GP for the pitch component of the synthesized speech by AbS based on the supplied auditory weighting distortion detection signal S14. Specifically, the pitch analysis unit 25 calculates the perceptual weighting distortion detection signal S14 based on the perceptual weighting distortion detection signal S14 provided from the perceptual weighting unit 24 and the sound source signal S11 provided from the addition unit 29 as described later. The pitch L that minimizes the signal level is selected, and this is determined as the pitch L of the current subframe. Then, the pitch analysis unit 25 sends the pitch L thus determined to the voiced / unvoiced sound section determination unit 4 (FIG. 1) as described above.

Note the pitch L is determined, the gain G ₁ is also determined accordingly. A code vector corresponding to the determination result is output from the adaptive codebook 27, and is amplified by the amplification unit 28 with the above-described gain _GP and supplied to the addition unit 29.

At this time, the fixed codebook search unit 26 determines the sound source signal S11 for the differential sound obtained by removing the synthesized sound from the adaptive codebook 27 from the input sound by AbS. Specifically, the fixed codebook search unit 26 outputs one code vector among the code vectors prepared in advance from the fixed codebook 30. Note one gain G _C is determined for a single code vector. Then the code vector outputted from the fixed codebook 30, the amplifying unit 31 after this is amplified processed by the above gain G _C, given to the adder 29.

Addition unit 29, by adding the code vector from the application code book 27 which is amplified treated with these gain G _P, and a code vector from the fixed codebook 30 is amplified processed by the gain G _C, speech production model A sound source signal S11 corresponding to the human vocal fold vibration is generated and transmitted to the synthesis filter 23 corresponding to the human articulatory motion, that is, the transfer function of the vocal tract in the speech generation model, and the pitch analysis unit 25 described above. To do.

  Thus, the sound source signal S11 is subsequently filtered in the synthesis filter 23 as described above to generate a synthesized voice, and distortion between the synthesized voice and the input voice is filtered in the perceptual weighting unit 24. The fixed code book search unit 26 is given. Then, the fixed codebook search unit 26 performs the same operation for all the code vectors, and determines the code vector with the smallest distortion as the code vector at that time.

  In this way, the CELP encoding unit 2 can obtain the LSP and pitch based on the input audio signal S1.

(1-3) Operation and effect of the present embodiment In the above configuration, the voice authentication apparatus 1 detects a voiced sound section based on pitch information obtained by encoding processing using the CS-ACELP encoding method. In addition, the voiced section is detected by the VAD algorithm, and only the LSP of the frame of the voiced section and the voiced section is extracted from the LSP for each frame obtained by the encoding process by the CS-ACELP encoding method, Speaker registration and speaker verification are performed using this LSP.

  Therefore, since this voice authentication device 1 can perform speaker registration and speaker verification using only the LSP of a frame that exhibits stable characteristics against environmental noise, it can be used during speaker registration or speaker verification. It can be made less susceptible to environmental noise.

  In fact, the above-mentioned “car”, “crowd” and “station” recorded in the electronic cooperative noise database are added to the ideal speech without noise included in the ATR Japanese speech database for research. (Station) or “street” noise is added by changing the addition ratio sequentially so that the SN ratio is 10 [dB], 15 [dB], and 20 [dB], and the above frame selection Experiments were performed to evaluate the reliability of speaker authentication in the case where noise proofing processing was performed and in the case where noise proofing processing was not performed. As a result, experimental results as shown in FIGS. 10 and 11 were obtained. Here, the evaluation is also performed using EER.

  As is clear from this experimental result, although there is a slight difference in accuracy depending on the type of noise, the noise proofing process is performed when the noise proofing process by the frame selection process as a whole (FIG. 10) is performed. The error rate is lower than when not performed (FIG. 11). Therefore, it can be seen that highly reliable speaker authentication can be performed by performing the noise proof processing.

  Separately from this, when the voice length (speech time) was changed to 5 seconds, 10 seconds, 20 seconds, and 30 seconds, an experiment for evaluating the relationship between the voice length and the collation accuracy was performed. Experimental results as shown in Fig. 1 were obtained.

  As can be seen from the results of this experiment, the voice length in both cases where the noise immunity processing by the above-described frame selection processing is performed (FIG. 12) and when the noise immunity processing is not performed (FIG. 13). The longer the length is, the higher the collation accuracy is. However, by performing the noise proofing process, it is possible to obtain the same matching precision as when the noise proofing process is not performed in a shorter time. Therefore, it was found that reliable speaker authentication can be performed in a short time by performing such noise proof processing.

  According to the above configuration, since the speaker registration and the speaker verification are performed using only the LSP of the voiced and voiced frame, the environmental noise at the time of speaker registration and speaker verification is obtained. Thus, it is possible to realize a voice collation apparatus that can perform speaker authentication with high noise resistance and high reliability.

(2) Second Embodiment (2-1) Principle In the voice authentication device 1 according to the first embodiment described above with reference to FIG. 1, importance is placed on the individuality that appears in the average LSP. A weighting process is performed in which the average LSP multiplied by a predetermined weighting coefficient w is added to the LSP of the voiced and voiced frame output from the frame selection processing unit 3.

  However, according to such a method, there is a problem that it takes a considerable time until the average LSP is stabilized, and there is also a problem that the average LSP is easily affected by additional noise because it includes a noise component.

  On the other hand, the quantization error calculated in the vector quantization unit 10 of the matching unit 9 described above with reference to FIG. 1 is the LSP based on the matching speech and the quantized representative vector that is a parameter specific to the registered speaker based on the registered speech. The smaller the value, the higher the possibility that the collation speaker is a registered speaker.

  Therefore, even if it is determined whether or not the collation speaker is a registered speaker based on the average of the minimum quantization errors, which are the minimum values of these quantization errors, output from the vector quantization unit 10, It seems that sufficient verification accuracy can be obtained.

  According to this method, when the verification speaker is a registered speaker, it is considered that this value is always a small value. Therefore, it takes much time until the average value of the minimum quantization error is stabilized. It is not necessary, and since the average of the additional noise is not added unlike the weighting processing in the weighting processing unit 6, it is considered that the influence of the additional noise is small.

  Therefore, in the second embodiment, a speaker authentication algorithm based on the average of minimum quantization errors is proposed. Hereinafter, a voice authentication device according to a second embodiment to which such an algorithm is applied will be described.

(2-2) Configuration of Voice Authentication Device According to Second Embodiment FIG. 14 showing the same parts as those in FIG. 1 with the same reference numerals shows a voice authentication device 40 according to the second embodiment. There is a configuration similar to that of the voice authentication device 1 (FIG. 1) according to the first embodiment except that the average distance calculation unit 42 is provided in the matching unit 41 instead of the weighting processing unit 6 (FIG. 1). Has been.

  In practice, in the case of the voice authentication device 40, the LSP of each frame in the voiced and voiced intervals output from the frame selection processing unit 3 is given to the clustering unit 7. Based on the LSP, the clustering unit 7 executes a clustering process based on the LBG + splitting algorithm that combines the above-described LBC algorithm and splitting algorithm, thereby obtaining the same 10-dimensional N (for example, 16) as the LSP unique to the speaker. The quantized representative vectors are obtained and stored in the storage unit 8 as the speaker's feature code book CB (FIG. 6) in the registration mode.

  On the other hand, in the verification mode, the LSP of each frame in the voiced and voiced intervals obtained in the same manner as in the registration mode is given from the frame selection processing unit 3 to the vector quantization unit 10 of the verification unit 41. In the collation mode, the vector quantization unit 10 is also provided with the feature code book CB of the registered speaker as a target among the feature code books CB stored in the storage unit 8.

  In this way, the vector quantization unit 10 determines the 16 quantized representative vectors constituting the feature code block CB for each LSP sequentially given from the frame selection processing unit 3 in the same manner as in the first embodiment. Each of the distances is calculated in order, and the minimum quantization error among the obtained 16 distances is sent to the average distance calculation unit 42 as the quantization error detection signal S4.

  The average distance calculation unit 42, based on the minimum quantization error for each frame of the voiced and voiced sections obtained based on the supplied quantization error detection signal S4, and each quantized representative vector of the feature code block CB An average distance from the LSP based on the collation voice is calculated.

  Actually, the average distance calculation unit 42 calculates the average value of the minimum quantization error for each frame of the voiced and voiced intervals obtained based on the quantization error detection signal S4 given from the vector quantization unit 10. At this time, the average distance calculation unit 42 uses the quantization minimum error for the past predetermined number of frames as the average value of the minimum quantization error for each frame, and averages the quantization minimum errors for the past predetermined number of frames. Is calculated by the moving average method to calculate the above, and the obtained calculation result is sent to the determination unit 43 as the quantization error average value signal S20.

  Thus, the determination unit 43 determines the average value of the minimum quantization error of the corresponding frame obtained based on the quantization error average value signal S20 and the average value in the same manner as in the first embodiment. The average value of the minimum quantization error is compared with the fifth threshold value, and when the average value of the minimum quantization error is smaller than the fifth threshold value, it is determined that the speaker at that time is the person in question. When it is larger than the threshold value of 5, the speaker at that time is determined to be another person, and the determination result is output as a determination signal S21.

(2-3) Operation and effect of the present embodiment In the above configuration, the voice authentication device 40 performs speaker verification based on the average value of the minimum quantization error of the frames in the voiced and voiced sections. .

  Therefore, this voice authentication device 40 performs speaker verification with the same accuracy in a much shorter time than the voice authentication device 1 (FIG. 1) according to the first embodiment that performs the weighting process using the average LSP. It is possible to perform speaker authentication in almost real time.

  Actually, when a simulation was performed, the voice authentication apparatus 1 according to the first embodiment required about 15 seconds to obtain authentication accuracy of an error rate of about 30%. According to this voice authentication device 40, it was confirmed that the same authentication accuracy could be obtained in about 1-2 seconds.

  Further, the voice authentication device 40 is less susceptible to environmental noise than the voice authentication device 1 according to the first embodiment that performs weighting processing using an average LSP, and thus provides highly reliable speaker authentication. It can be carried out.

  Further, in the voice authentication device 40, when calculating the average value of the minimum quantization error for each frame of the voiced section and the voiced section, the voice authentication apparatus 40 is calculated by the moving average method using the quantization minimum errors for the past predetermined number of frames. Therefore, the average value of the minimum quantization error can be calculated using only the latest information, and the speaker authentication with higher reliability can be performed accordingly.

  According to the above configuration, the speaker verification is performed based on the average value of the minimum quantization error of the frames in the voiced and voiced sections, so that the speaker authentication can be performed more quickly and more reliably. A voice authentication device capable of performing

(3) Other Embodiments In the first and second embodiments described above, pitch information obtained by CS-ACELP coding processing is used as a method for detecting a voiced sound section in the speech signal S1. Although the present invention is not limited to this, the present invention is not limited to this, and as a method for detecting the voiced sound section in the speech signal S1, various methods according to the encoding method applied in the speech authentication apparatus can be used. Can be widely applied.

  In the first and second embodiments described above, the case where the VAD algorithm is applied as a method for detecting a voiced section in the audio signal S1 has been described, but the present invention is not limited to this, Various other methods can be widely applied.

  However, as a method for detecting the voiced sound section in the speech signal S1 as in the first and second embodiments, the pitch information obtained by the CS-ACELP encoding process is used, while the speech in the speech signal S1 is used. By applying the VAD algorithm as a method for detecting a section, it is possible to detect a voiced section and a voiced section using parameters obtained in an existing mobile phone. There is an advantage that the voice authentication function according to the present invention can be installed.

  Furthermore, in the above-described second embodiment, only the minimum quantization error among the 16 distances calculated by the vector quantization unit 10 is used to collate with each quantized representative vector of the feature code block CB. Although the case where the average distance to the LSP based on speech is calculated has been described, the present invention is not limited to this, and all 16 distances calculated in the vector quantization unit 10 are used, An average value may be calculated.

  The present invention is widely applied to a voice authentication device that performs speaker authentication based on parameters obtained by encoding processing using a CELP encoding method and a voice authentication device that performs speaker authentication based on various types of voice information. Can be applied.

It is a block diagram which shows the structure of the voice authentication apparatus by 1st Embodiment. It is a flowchart which shows a voiced / unvoiced sound area determination processing procedure. It is a flowchart which shows a sound / silence determination processing procedure. It is a figure with which it uses for description of determination of the voiced sound / unvoiced sound area based on the voiced sound / unvoiced sound signal in a frame selection process part. It is a conceptual diagram with which it uses for description of the weighting process in a weighting process part. It is a conceptual diagram with which it uses for description of the registration process in a voice authentication apparatus. It is a flowchart which shows a clustering process procedure. It is a conceptual diagram with which it uses for description of the collation process in a voice authentication apparatus. It is a block diagram which shows the structure of a CELP encoding part. It is a table | surface which shows the measurement experiment result of EER in various noise environment at the time of performing the noise-proof process by this Embodiment. It is a graph which shows the measurement experiment result of EER in various noise environment at the time of not performing the noise-proof process by this Embodiment. It is a graph which shows the relationship between the audio | voice length at the time of performing the noise-proof process by this Embodiment, and an error rate. It is a graph which shows the relationship between the audio | voice length when not carrying out the noise-proof process by this Embodiment, and an error rate. It is a block diagram which shows the structure of the voice authentication apparatus by 2nd Embodiment. It is a graph which shows the measurement experiment result of EER under various noise environment. It is a characteristic curve figure with which it uses for description of FRR, FAR, and EER. It is a chart which shows the specifications of various noise data.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1, 40 ... Voice authentication apparatus, 2 ... CELP encoding part, 3 ... Frame selection part, 4 ... Voiced / unvoiced sound determination part, 5 ... Sound / silence determination part, 6 ... Weighting process part , 7... Clustering unit, 8... Storage unit, 9 and 41... Verification unit, 10... Vector quantization unit, 11 and 43. , S2... Voiced / unvoiced sound determination signal, S3... Voiced / silent sound determination signal, S4... Quantization error detection signal, S5, S21.

Claims (9)

Feature parameter generating means for sequentially generating and outputting predetermined feature parameters representing the personality of the speaker based on the input audio signal;
A voiced section detecting means for detecting a voiced section of the audio signal;
Voiced sound section detecting means for detecting a voiced sound section of the voice signal;
Based on the detection results of the voiced section detecting means and the voiced sound detecting means, the voiced section of the voice signal and the voiced voice section of the voice parameter are sequentially output from the feature parameters generated from the feature parameter generating means. Feature parameter extraction means for extracting feature parameters;
A voice authentication apparatus comprising: speaker authentication means for performing speaker authentication based on the feature parameter extracted by the feature parameter extraction means. The feature parameter generation means includes:
The audio signal is encoded by a CS-ACELP (Conjugate Structure-Algebraic Code Excited Linear Prediction) encoding method to generate an LSP (Line Spectrum Pair) as the feature parameter in units of frames,
The voiced sound section detecting means is
Detecting a voiced sound section of the voice signal based on pitch information obtained by encoding the voice signal in the feature parameter generation means;
The voiced section detecting means is
The voice authentication device according to claim 1, wherein the voiced section of the voice signal is detected using a VAD (Voice Activity Decision) algorithm. The above speaker authentication means
A feature that generates a feature codebook including a predetermined number of representative vectors unique to the speaker by performing a predetermined clustering process on the extracted feature parameters supplied from the feature parameter extraction means in the registration mode. Codebook generation means;
Storage means for storing and holding the feature code book generated by the feature code book generation means;
Distance calculating means for calculating a distance between the extracted feature parameter supplied from the feature parameter extracting means in the collation mode and each representative vector constituting the feature codebook read from the storage means;
Average distance calculating means for calculating an average value of the distance between the feature parameter and the representative vector calculated by the distance calculating means;
The voice authentication device according to claim 1, further comprising: a determination unit that performs speaker collation determination based on a calculation result of the average distance calculation unit. The average distance calculation means is
The average value of the distance between the characteristic parameter and the representative vector calculated by the distance calculation means is calculated by a moving average method using a predetermined number of the distances in the past. Voice authentication device. A first step of sequentially generating and outputting predetermined feature parameters representing the individuality of the speaker based on the input voice signal, and detecting each of the voiced and voiced sections of the voice signal;
A second step of extracting, from the generated feature parameters, the feature parameters of the voiced section and the voiced section of the voice signal;
A voice authentication method comprising: a third step of performing speaker authentication based on the extracted feature parameter. In the first step,
The audio signal is encoded by a CS-ACELP (Conjugate Structure-Algebraic Code Excited Linear Prediction) encoding method to generate an LSP (Line Spectrum Pair) as the feature parameter in units of frames. 6. The voiced sound section of the voice signal is detected based on the obtained pitch information, and the voiced section of the voice signal is detected using a VAD (Voice Activity Decision) algorithm. The voice authentication method described. The third step is
Feature codebook generation for generating a feature codebook composed of a predetermined number of representative vectors unique to the speaker by performing a predetermined clustering process on the feature parameters extracted in the second step in the registration mode Steps,
A storing step for storing and storing the generated feature codebook;
A distance calculating step for calculating a distance between the feature parameter of the voiced voice section and the voiced voice period in the extracted voice signal and each representative vector constituting the feature code book stored and held in the collation mode;
An average distance calculating step of calculating an average value of the distance between the feature parameter and the representative vector calculated in the distance calculating step;
The voice authentication method according to claim 5, further comprising: a determination step of performing speaker verification determination based on a calculation result in the average distance calculation step. In the above average distance calculation step,
The voice according to claim 7, wherein the average value of the feature parameter calculated in the distance calculation step and the distance between the representative vectors is calculated by a moving average method using a predetermined number of the distances in the past. Authentication method. On the computer,
A first step of sequentially generating and outputting predetermined feature parameters representing the individuality of the speaker based on the input voice signal, and detecting each of the voiced and voiced sections of the voice signal;
A second step of extracting, from the generated feature parameters, the feature parameters of the voiced section and the voiced section of the voice signal;
A program for executing a process comprising: a third step of performing speaker authentication based on the extracted feature parameter.

Images