JP5151103B2 - Voice authentication apparatus, voice authentication method and program - Google Patents

Description

  The present invention relates to an authentication technique using voice.

Features of the voice (hereinafter referred to as “registered voice”) collected in advance from a legitimate user and the voice (hereinafter referred to as “authentication voice”) collected from the subject of authentication (hereinafter referred to as “authenticated person”) Conventionally, a voice authentication technique has been proposed in which the authenticity of the person to be authenticated is determined by comparing the distance to the feature amount with a threshold value. Patent Document 1 discloses a configuration in which a threshold value is changed according to the purpose of authentication and required accuracy.
JP 2003-248661 A

  FIG. 8 is a graph used for evaluation of voice authentication. The FRR (False Rejection Rate) in the figure means the probability (authentication rejection rate) that the authenticity is denied even if the person to be authenticated is a valid user, and the FAR (False Acceptance Rate) is This means the probability (acceptance rate of others) that the authenticity is affirmed even though the person to be authenticated is not a valid user. As can be seen from the figure, if the threshold value used for authentication is set to the numerical value a in FIG. 8, a legitimate user is maintained while maintaining a sufficiently high probability that an unauthorized person will be rejected. The possibility of being rejected is sufficiently reduced.

  However, since the characteristics of the authentication voice are affected by the voice generated around at the time of authentication (hereinafter referred to as “noise at the time of authentication”), the distance between the authentication voice and the registered voice varies according to the noise at the time of authentication. Therefore, each curve of FRR and FAR moves in parallel along the horizontal axis (distance) according to the characteristics of the noise during authentication. For example, when the FAR indicated by the solid line in FIG. 8 changes to the broken line L1, the probability that the validity of another person is misidentified increases (that is, the accuracy of authentication decreases), and the FRR in FIG. In the case of the fluctuation, the probability that legitimate user authentication is rejected increases (that is, convenience decreases). That is, in the conventional voice authentication, there is a problem that the balance between the accuracy and convenience of authentication is lost in accordance with the characteristics of noise during authentication.

  Even if the threshold value is changed according to the purpose of authentication and the required accuracy as in Patent Document 1, the above problems are not solved at all. In addition, when authentication is performed with a portable electronic device typified by a mobile phone, the characteristics of noise during authentication vary depending on the environment in which the electronic device is used. To do. In view of such circumstances, an object of the present invention is to solve the problem of maintaining a balance between the accuracy and convenience of authentication regardless of noise during authentication.

In order to solve the above-described problems, a voice authentication device according to one aspect of the present invention includes a characteristic analysis unit that analyzes characteristics of noise at the time of authentication that occurs around a person to be authenticated at the time of authentication, and a registration that is registered in advance. The threshold value is set so as to satisfy the relationship according to the characteristic of the noise during authentication analyzed by the characteristic analysis means with respect to the volume ratio between the noise during registration and the volume of the registered voice (eg, the volume ratio EN_SN in FIG. 4). Depending on the result of comparison between the threshold value setting means to be set and the index value indicating the similarity between the registered voice and the authentication voice collected from the authenticated person and the threshold value set by the threshold setting means And authentication means for authenticating. According to the above aspect, since the threshold value is variably set according to the noise at the time of authentication, it is possible to maintain the balance between the accuracy and convenience of authentication regardless of the noise at the time of authentication.

In a preferred aspect of the present invention , the threshold value setting means sets the threshold value so that the threshold value changes along a straight line or a curve corresponding to the noise at the time of authentication with respect to the volume ratio between the noise during registration and the registered voice. More specifically, the threshold value setting means compares the volume ratio between the noise during registration and the volume of the registered voice (for example, the volume ratio V_SN) and the volume ratio between the noise during registration and the volume of the registered voice. The threshold value is set so that the threshold value changes along a straight line or a curve according to the difference (for example, DIF_SN1 to DIF_SN3 in FIG. 4). According to the above aspect, it is possible to maintain the balance between the accuracy and convenience of authentication by a simple process regardless of the characteristics of registration noise and authentication noise. The threshold setting means includes a registration noise and a volume ratio of the registered voice and an authentication noise (more specifically, a difference between the authentication noise and the volume ratio of the authentication voice and the registration noise and the volume ratio of the registration voice) and the threshold value. The threshold value may be set based on a table that defines the relationship between the threshold value and the threshold value may be calculated by calculation using a mathematical expression that expresses the relationship between these numerical values. The contents of the table and the mathematical formula can be appropriately changed according to an instruction from the user, for example.

  In one aspect of the present invention, the threshold value setting means corrects the threshold value according to the difference between the noise at the time of authentication and the noise at the time of registration (for example, the correlation value NOISE_DIF in FIG. 6) (for example, the correction unit 54 in FIG. 1). including. According to the above aspect, the correlation between the noise at the time of authentication and the noise at the time of registration at the time of actual authentication is the noise at the time of authentication assumed when determining the relationship between the noise at the time of registration and the volume ratio of the registered voice and the threshold value. Even when the correlation with the time noise is different, the balance between the accuracy of authentication and convenience can be effectively maintained by the correction means correcting the threshold value. The correction means may set the correction value for the threshold based on a table that defines the relationship between the difference between the noise at authentication and the noise at registration and the correction value (for example, the correction value A1 in FIG. 6). The threshold value may be calculated by a calculation using a mathematical expression expressing the relationship. The contents of the table and the mathematical formula can be appropriately changed according to an instruction from the user, for example.

  In one aspect of the present invention, the threshold setting means includes a correction means (for example, a correction unit in FIG. 1) that corrects the threshold according to the time length of the authentication voice or the registered voice (for example, the utterance length EN_SPEEECH_LEN and the utterance length V_SPEECH_LEN in FIG. 7). 54). According to the above aspect, the actual utterance length of the registered voice or authentication voice is different from the utterance length assumed when determining the relationship between the noise during registration and the volume ratio of the registered voice and the threshold value. In addition, it is possible to effectively maintain the balance between the accuracy of authentication and convenience by correcting the threshold value by the correcting means. The correction means may set a correction value for the threshold based on a table that defines the relationship between the time length of the authentication voice or the registered voice and the correction value (for example, the correction value A2 in FIG. 7). The threshold value may be calculated by a calculation using a mathematical expression that expresses. The contents of the table and the mathematical formula can be appropriately changed according to an instruction from the user, for example.

  Note that the variable serving as a reference for determining the correction value for the threshold value is not limited to the difference between the noise at the time of authentication and the noise at the time of registration, or the time length of the authentication voice or the registered voice. For example, correction means for correcting the threshold according to the ratio of the time length of voiced sound and unvoiced sound among the registered voice and authentication voice, or correction means for correcting the threshold according to the number of syllables of the registration voice and authentication voice , It may be included in the threshold setting means. In any aspect, similarly to the above, a configuration in which the correction value is determined according to the table or the mathematical formula or a configuration in which the contents of the table or the mathematical formula are variable is adopted.

The present invention is also specified as an operation method (voice authentication method) of the voice authentication device according to each of the above aspects. A voice authentication method according to one aspect of the present invention analyzes a characteristic of noise at the time of authentication that occurs around an authenticated person at the time of authentication . relative volume ratio, so as to satisfy the analysis relationship in accordance with the characteristics of the authentication noise, set the threshold value, the index value indicating the feature value classes whether the collected authentication voice from the registration voice and the person to be authenticated And the person to be authenticated is authenticated according to the result of comparison with the set threshold value. According to the above method, the same operation and effect as the voice authentication device according to the present invention are exhibited.

The voice authentication device according to each aspect described above is realized by hardware (electronic circuit) such as a DSP (Digital Signal Processor) dedicated to each process, and a general-purpose arithmetic processing device such as a CPU (Central Processing Unit). This is also realized through collaboration with programs. The program according to the present invention includes a characteristic analysis process for analyzing a characteristic of noise at the time of authentication generated around a person to be authenticated, and a volume ratio between the noise at the time of registration of the registered voice registered in advance and the volume of the registered voice. respect, so as to satisfy the relationship according to the characteristics of the authentication noise analyzed by the characteristics analysis process, and threshold setting process for setting a threshold value, feature quantity of the authentication voice taken from the registration voice and the person to be authenticated This is the content for executing the authentication process for authenticating the person to be authenticated in accordance with the comparison result between the index value indicating similarity and the threshold value set in the threshold value setting process. Even with the above program, the same operations and effects as the voice authentication device according to each of the above aspects can be obtained. The program of the present invention is provided to a user in a form stored in a portable recording medium such as a CD-ROM and installed in a computer, or provided from a server device in a form of distribution via a network. Installed on the computer.

<A: Configuration of voice authentication device>
FIG. 1 is a block diagram showing a configuration of a voice authentication apparatus according to one embodiment of the present invention. The voice authentication device 100 is a device that determines the legitimacy of the person to be authenticated (whether or not the user is an authorized user registered in advance) based on the voice when the person to be authenticated speaks a specific word. It is installed in various electronic devices such as mobile phones and information processing devices. The characteristic analysis unit 20, the authentication unit 40, and the threshold setting unit 50 illustrated in FIG. 1 may be realized by an arithmetic processing device such as a CPU executing a program, or may be realized by a hardware circuit such as a DSP. May be.

  The operation of the voice authentication device 100 is divided into initial registration and authentication. The initial registration is an operation for registering voice (registered voice) uttered by a legitimate user prior to authentication. The authentication is an operation for authenticating the authenticity of the person to be authenticated by comparing the registered voice and the voice uttered by the person to be authenticated (authentication voice). The operation unit 10 includes a plurality of operators that are operated by a user. The user can instruct the voice authentication apparatus 100 to start initial registration or authentication by appropriately operating the operation unit 10.

  The input unit 15 and the characteristic analysis unit 20 in FIG. 1 are used for detecting characteristics of authentication voice and noise around the voice authentication device 100 (authentication noise) at the time of authentication. As shown by R, it is used to detect the characteristics of registered voice and noise around the voice authentication device 100 (registration noise).

  The input unit 15 is a sound collection device that generates an acoustic signal S corresponding to surrounding sounds (voice and noise). As illustrated in FIG. 2, the acoustic signal S is divided into a non-vocal segment P1 and a vocal segment P2. The utterance section P2 is a section in which a legitimate user utters a registered voice during initial registration or a section in which an authenticated person utters an authentication voice during authentication. On the other hand, the non-voicing section P1 is a section in which the registered voice and the authentication voice are not uttered. Since various kinds of noise can be generated in the environment where the voice authentication device 100 is installed, it is not completely silent (the amplitude of the acoustic signal S is zero) even in the non-speech section P1, and as shown in FIG. Or noise during authentication is picked up by the input unit 15.

  The acoustic signal S generated by the input unit 15 is supplied to the characteristic analysis unit 20 of FIG. The characteristic analysis unit 20 is a means for analyzing the sound collected by the input unit 15, and includes a section detection unit 22, a switching unit 23, a noise analysis unit 25, a voice analysis unit 26, and a feature analysis unit 28. The section detection unit 22 classifies the non-speech section P1 and the utterance section P2. For example, the section detection unit 22 detects a time point when the amplitude of the acoustic signal S increases or decreases discontinuously as a boundary between the non-speaking section P1 and the speaking section P2. It should be noted that various known techniques are employed for the division between the non-speaking section P1 and the speaking section P2.

  The switching unit 23 is means for selectively switching the supply destination of the acoustic signal S generated by the input unit 15. Of the acoustic signal S, the section that the section detection unit 22 recognizes as the non-voice section P1 is supplied to the noise analysis section 25, and the section that the section detection section 22 recognizes as the utterance section P2 includes the voice analysis section 26 and the feature analysis section 28. To be supplied.

  The noise analysis unit 25 is a means for analyzing the characteristics of the noise during registration and the noise during authentication based on the acoustic signal S in the non-voice section P1. The noise analysis unit 25 of the present embodiment periodically analyzes the characteristics of the acoustic signal S within the non-vocal period P1. When the initial registration or the start of authentication is instructed according to the operation on the operation unit 10, the noise analysis unit 25, as shown in FIG. 2, from the time of the instruction to a time point before the predetermined time length. The result of analysis in the section (hereinafter referred to as “detection section”) P is determined as the characteristics of noise during registration and noise during authentication. In the following explanation, items related to registration noise and registration voice are indicated by a code including “EN (enroll)”, and items related to authentication noise and authentication voice include “V (verify)”. Indicated by the sign.

  As shown in FIG. 1, the noise analysis unit 25 according to the present embodiment calculates a frequency characteristic EN_NOISE_FC and a noise level EN_NOISE_LEVEL with respect to noise at the time of initial registration, and a frequency characteristic V_NOISE_FC and a noise level V_NOISE_LEVEL with respect to noise at the time of authentication. And calculate. The noise level (EN_NOISE_LEVEL, V_NOISE_LEVEL) is an average value of the intensities (sound pressures) of components belonging to a predetermined frequency band in the acoustic signal S in the detection section P in the non-voice section P1. The frequency characteristics (EN_NOISE_FC, V_NOISE_FC) are information indicating the intensity of each component when the acoustic signal S in the detection section P is divided into a plurality of frequency bands. Therefore, the noise analysis unit 25 includes, for example, a plurality of band pass filters (filter banks) having different pass bands. However, the noise analysis unit 25 may be a means for calculating the frequency spectrum as frequency characteristics (EN_NOISE_FC, V_NOISE_FC) by frequency analysis such as FFT (Fast Fourier Transform) processing.

  The voice analysis unit 26 analyzes the characteristics of the registered voice and the authentication voice based on the acoustic signal S in the utterance section P2. The voice analysis unit 26 according to the present embodiment calculates the utterance level EN_SPEECH_LEVEL and the utterance length EN_SPEECH_LEN for the registered voice during initial registration, and calculates the utterance level V_SPEECH_LEVEL and the utterance length V_SPEECH_LEN for the authentication voice during authentication. The utterance level (EN_SPEECH_LEVEL, V_SPEECH_LEVEL) is an average value of the intensities of components belonging to a predetermined frequency band in the acoustic signal S in the utterance section P2. The utterance length (EN_SPEECH_LEN, V_SPEECH_LEN) indicates the time length of the utterance section P2 (that is, the time length during which the utterance is continued). The length of time from when the amplitude of the acoustic signal S sharply increases (start point of the utterance interval P2) to when the amplitude of the acoustic signal S decreases sharply (end point of the utterance interval P2) is the utterance length (EN_SPEECH_LEN, V_SPEECH_LEN) Detected.

  The feature analysis unit 28 is a means for analyzing the features of registered speech and authentication speech. The feature analysis unit 28 of the present embodiment calculates the feature amount EN_DATA of the registered voice at the time of initial registration, and calculates the feature amount V_DATA of the authentication voice at the time of authentication. The feature amount (EN_DATA, V_DATA) is a time-series vector sequence of cepstrum calculated from the acoustic signal S in the utterance section P2. Therefore, means for executing various operations including frequency analysis (for example, FFT processing) is preferably employed as the feature analysis unit 28.

  The storage device 32 is a means for storing various information used for authentication. For example, as illustrated in FIG. 1, the storage device 32 stores, as an authentication dictionary, various types of information specified by the characteristic analysis unit 20 regarding registered speech and registration noise. That is, the frequency characteristic EN_NOISE_FC and noise level EN_NOISE_LEVEL specified by the noise analyzer 25, the utterance level EN_SPEECH_LEVEL and utterance length EN_SPEECH_LEN specified by the speech analyzer 26, and the feature quantity EN_DATA specified by the feature analyzer 28 are prior to authentication. And stored in the storage device 32. The storage device 32 may be a device fixedly installed in the voice authentication device 100 or a portable device (memory) that can be freely attached to and detached from the voice authentication device 100.

  The authentication unit 40 is means for authenticating the authenticity of the person to be authenticated by comparing the registered voice and the authentication voice, and includes a distance calculation unit 42 and a determination unit 44. The distance calculation unit 42 calculates a distance DIST between the feature value V_DATA generated by the feature analysis unit 28 for the authentication speech and the feature value EN_DATA stored in the storage device 32. For the calculation of the distance DIST, various pattern matching techniques such as DP matching for calculating a normalized distance between the vector strings of the feature quantities EN_DATA and V_DATA are used. As the distance DIST is smaller, the authentication voice is more similar to the registration voice (that is, the person to be authenticated is more likely to be a valid user).

  The determination unit 44 determines the authenticity of the person to be authenticated by comparing the distance DIST calculated by the distance calculation unit 42 with a threshold value TH. That is, the determination unit 44 affirms the authenticity of the person to be authenticated when the distance DIST is less than the threshold TH (that is, when the registered voice and the authentication voice are similar), and when the distance DIST exceeds the threshold TH (that is, If the registration voice and the authentication voice are different, the authenticity of the person to be authenticated is denied. The result of determination by the determination unit 44 is output from the output unit 60. For example, a display device that outputs the authentication result as an image or a sound emitting device that outputs the authentication result by sound is suitably employed as the output unit 60.

  The threshold value setting unit 50 is a means for variably setting the threshold value TH used for determination by the determination unit 44 in accordance with noise at the time of authentication or noise at the time of registration, and includes an initial value setting unit 52 and correction units 54 and 56. . The initial value setting unit 52 sets an initial value of the threshold value TH based on information generated by the characteristic analysis unit 20 at the time of initial registration and authentication. The threshold value TH set by the initial value setting unit 52 is output to the determination unit 44 through correction by the correction units 54 and 56. The correcting unit 54 corrects the threshold value TH according to the difference between the authentication noise and the registration noise. The correcting unit 56 corrects the threshold value TH according to the utterance length EN_SPEECH_LEN at the time of initial registration. The storage unit 35 in FIG. 1 stores a table used by the threshold setting unit 50 for setting and correcting the threshold TH. A specific storage area in the storage device 32 may be used as the storage unit 35.

<B: Operation of voice authentication device>
Next, the operation of the voice authentication device 100 will be described focusing on the process in which the threshold setting unit 50 sets the threshold TH during authentication. The threshold value setting unit 50 executes the process of FIG. 3 every time the time when authentication is necessary comes. For example, when an electronic device equipped with the voice authentication device 100 starts operating when power is turned on, or when the electronic device starts a predetermined operation (for example, access to specific information). There is a time. When authentication starts, the person to be authenticated instructs the start of utterance by operating the operation unit 10 and then utters a predetermined word to the input unit 15. The noise analysis unit 25 specifies the frequency characteristic V_NOISE_FC and the noise level V_NOISE_LEVEL from the acoustic signal S in the detection section P whose end point is the operation timing of the operation unit 10, and utters from the acoustic signal S in the subsequent speech section P2. Specify level V_SPEECH_LEVEL and utterance length V_SPEECH_LEN.

As shown in FIG. 3, the threshold setting unit 50 calculates a volume ratio EN_SN between the noise at the time of registration and the registered voice (step S10). The volume ratio EN_SN is a relative ratio between the utterance level EN_SPEECH_LEVEL and the noise level EN_NOISE_LEVEL stored in the storage device 32, and is calculated by the following equation (1), for example.
EN_SN = log (EN_SPEECH_LEVEL / EN_NOISE_LEVEL) ...... (1)

Next, the threshold setting unit 50 calculates the volume ratio V_SN between the noise at the time of authentication and the authentication voice (step S11). The volume ratio V_SN is a relative ratio between the noise level V_NOISE_LEVEL supplied from the noise analysis unit 25 and the utterance level V_SPEECH_LEVEL supplied from the voice analysis unit 26, and is calculated by the following equation (2) in the same manner as the volume ratio EN_SN. The
V_SN = log (V_SPEECH_LEVEL / V_NOISE_LEVEL) ...... (2)

式(3)における数値EN_MAG(i)は、複数の周波数帯域のうち変数ｉで指定される周波数帯域における登録時雑音の強度であり、数値EN_MAG(i)から減算される数値EN_MAG_AVEは、変数ｉで指定される周波数帯域における登録時雑音の強度の平均値である。同様に、数値V_MAG(i)は、変数ｉで指定される周波数帯域における認証時雑音の強度であり、数値V_MAG_AVEは当該周波数帯域における認証時雑音の強度の平均値である。したがって、登録時雑音と認証時雑音とが完全に合致する場合には相関値NOISE_DIFが「１」となり、両雑音の特性の相違が拡大するほど相関値NOISE_DIFは減少していく（−１≦NOISE_DIF≦１）。 In the next step S12, the threshold setting unit 50 determines the difference between the initial registration volume ratio EN_SN calculated in step S10 and the current authentication volume ratio V_SN calculated in step S11 (hereinafter referred to as “volume ratio difference value”). DIF_SN is calculated (DIF_SN = V_SN−EN_SN). Further, the threshold setting unit 50 calculates a correlation value NOISE_DIF indicating a correlation between characteristics of registration noise and authentication noise (for example, similarity of spectrum shape) (step S13). The correlation value NOISE_DIF is calculated by the following equation (3), for example.

The numerical value EN_MAG (i) in Expression (3) is the noise intensity at the time of registration in the frequency band specified by the variable i among the plurality of frequency bands, and the numerical value EN_MAG_AVE subtracted from the numerical value EN_MAG (i) is the variable i This is the average value of the noise intensity during registration in the frequency band specified by. Similarly, the numerical value V_MAG (i) is the intensity of authentication noise in the frequency band specified by the variable i, and the numerical value V_MAG_AVE is an average value of the intensity of authentication noise in the frequency band. Therefore, when the registration noise and the authentication noise completely match, the correlation value NOISE_DIF becomes “1”, and the correlation value NOISE_DIF decreases as the difference between the characteristics of both noises increases (−1 ≦ NOISE_DIF). ≦ 1).

  Next, the initial value setting unit 52 specifies an initial value of the threshold TH based on the volume ratio EN_SN calculated in step S10 and the volume ratio difference value DIF_SN calculated in step S12 (step S14). According to the test by the inventors of the present application, the threshold TH for maintaining the authentication accuracy at a high level has led to the finding that the sound volume ratio EN_SN and the sound volume ratio difference value DIF_SN satisfy a predetermined relationship. . That is, as shown in FIG. 4, in each case where the numerical value of the volume ratio EN_SN (volume ratio of noise during registration and registered voice) is changed, the accuracy of authentication satisfies the desired condition (for example, FRR or FAR is When the threshold TH is set to be lower than the initial value, the horizontal axis is the volume ratio EN_SN, and the threshold TH corresponding to each volume ratio EN_SN is plotted on the vertical axis and statistically processed, each point is the volume ratio. There is a tendency to be distributed on a straight line according to the difference value DIF_SN. FIG. 4 shows three straight lines corresponding to three types of volume ratio difference values DIF_SN (DIF_SN1 to DIF_SN3). Now, assuming that the volume ratio EN_SN is the numerical value SNa and the volume ratio difference value DIF_SN is the numerical value DIF_SN1, the authentication with the expected accuracy is realized by selecting the threshold value TH as the numerical value THa. As shown in the figure, as the volume ratio EN_SN and the volume ratio difference value DIF_SN increase, the threshold value TH for maintaining the authentication with the expected accuracy increases.

  Based on the above knowledge, the initial value setting unit 52 sets the initial value of the threshold value TH so as to satisfy the linear relationship corresponding to the volume ratio difference value DIF_SN with respect to the volume ratio EN_SN calculated in step S10. . For example, as shown in FIG. 4, when the volume ratio EN_SN is a numerical value SNa, when the numerical value DIF_SN1 is calculated in step S12, the numerical value THa is set as an initial value of the threshold value TH, and in step S12, the numerical value DIF_SN2 Is calculated, the numerical value THb is set as the initial value of the threshold value TH. When the volume ratio difference value DIF_SN does not correspond to a preset numerical value, the initial value of the threshold value TH is calculated by interpolation. For example, when the intermediate value DIF_SN between the numerical value DIF_SN1 and the numerical value DIF_SN2 is calculated in step S12, the intermediate value THc between the numerical value THa corresponding to the numerical value DIF_SN1 and the numerical value THb corresponding to the numerical value DIF_SN2 is the initial value of the threshold value TH. Calculated as

  The initial value setting unit 52 of the present embodiment calculates the initial value of the threshold value TH from the volume ratio EN_SN and the volume ratio difference value DIF_SN based on a table created so as to satisfy the above conditions. FIG. 5 is a conceptual diagram showing the contents of the table used in step S14. As shown in the figure, a plurality of tables corresponding to different volume ratio difference values DIF_SN (DIF_SN1, DIF_SN2, DIF_SN3,...) Are stored in the storage unit 35. In the table corresponding to one volume ratio difference value DIF_SN, each numerical value of the volume ratio EN_SN and the threshold value TH are set so that the volume ratio EN_SN and the threshold value TH satisfy a linear relationship corresponding to the volume ratio difference value DIF_SN. The initial value is associated. In step S14, the initial value setting unit 52 searches for one table corresponding to the volume ratio difference value DIF_SN calculated in step S12, and corresponds to the volume ratio EN_SN calculated in step S10 among the searched tables. The attached threshold value TH is set as an initial value. When there is no table corresponding to the volume ratio difference value DIF_SN calculated in step S12, the initial value setting unit 52 determines the volume ratio EN_SN from each table corresponding to the volume ratio difference value DIF_SN before and after the volume ratio difference value DIF_SN. The threshold value TH corresponding to is calculated, and the initial value is calculated by interpolation of each threshold value TH.

  Incidentally, the relationship among the volume ratio EN_SN, the volume ratio difference value DIF_SN, and the threshold value TH illustrated in FIG. 4 is determined on the assumption that, for example, the characteristics of the noise during registration and the noise during authentication are equivalent. In practice, however, the characteristics of registration noise and authentication noise are often different. Therefore, the correction unit 54 corrects the initial value of the threshold value TH calculated in step S14 according to the correlation value NOISE_DIF between the noise during registration and the noise during authentication (step S15). More specifically, the correction unit 54 calculates the corrected threshold value TH by adding the correction value A1 corresponding to the correlation value NOISE_DIF to the initial value of the threshold value TH.

  FIG. 6 is a graph showing the relationship between the correlation value NOISE_DIF and the correction value A1. The correlation value NOISE_DIF calculated by Equation (3) varies within the range from "-1" to "1" depending on the correlation between the registration noise and the authentication noise, and the characteristics of both noises are completely If they match, it is “1”. If the relationship between the noise at the time of registration and the noise at the time of authentication is equivalent to that at the time of determining the relationship in FIG. 4 (both noises match in this embodiment), it is not necessary to correct the threshold value TH according to the correlation value NOISE_DIF. . On the other hand, the threshold value TH should be corrected so that the relationship between the registration noise and the authentication noise deviates from the determination of the relationship shown in FIG. Therefore, as shown in FIG. 6, when the correlation value NOISE_DIF is “1”, the correction unit 54 sets the correction value A1 to zero (no correction), and as the correlation value NOISE_DIF is smaller than “1”. The correction value A1 is set so as to be a large numerical value. More specifically, a table in which the correlation value NOISE_DIF and the correction value A1 are associated with each other is stored in the storage unit 35, and the correction unit 54 determines the correction value A1 based on this table.

  Further, the relationship between the volume ratio EN_SN and the volume ratio difference value DIF_SN and the threshold value TH shown in FIG. 4 is determined on the assumption that, for example, the registered voice continues for a predetermined time length L0. However, the utterance length EN_SPEECH_LEN of the registered voice in actual initial registration varies. Therefore, the correcting unit 56 corrects the threshold value TH corrected in step S15 according to the utterance length EN_SPEECH_LEN (step S16). More specifically, the correction unit 56 calculates the corrected threshold value TH by adding the correction value A2 corresponding to the utterance length EN_SPEECH_LEN to the threshold value TH. The threshold value TH after correction by the correction unit 56 is used for comparison with the distance DIST in the determination unit 44.

  FIG. 7 is a graph showing the relationship between the utterance length EN_SPEECH_LEN and the correction value A2. If the utterance length EN_SPEECH_LEN is equivalent to the time length L0 at the time of determining the relationship in FIG. 4, it is not necessary to correct the threshold value TH according to the utterance length EN_SPEECH_LEN. Therefore, when the utterance length EN_SPEECH_LEN stored in the storage device 32 is equal to the time length L0, the correction unit 56 sets the correction value A2 to zero. Further, since the utterance of the registered voice becomes more stable as the utterance length EN_SPEECH_LEN is longer, the feature quantity EN_DATA accurately reflects the basic voice quality of the legitimate user, and the accuracy of the distance DIST is improved. Therefore, even if the threshold value TH is lowered from the viewpoint of reducing the possibility of authenticating the validity of another person, the possibility of denying a legitimate user does not unreasonably increase. Therefore, the correction unit 56 selects a negative value corresponding to the utterance length EN_SPEECH_LEN as the correction value A2 when the utterance length EN_SPEECH_LEN is longer than the time length L0, and when the utterance length EN_SPEECH_LEN is shorter than the time length L0, A positive number corresponding to the utterance length EN_SPEECH_LEN is selected as the correction value A2. More specifically, a table in which the utterance length EN_SPEECH_LEN is associated with the correction value A2 is stored in the storage unit 35, and the correction unit 56 determines the correction value A2 based on this table.

  As described above, in this embodiment, the threshold value TH is variably set according to the relationship between the authentication noise and the authentication speech (V_SN) and the relationship between the registration noise and the registration speech (EN_SN). Regardless of the characteristics of noise at the time of authentication and noise at the time of registration, authentication can be maintained at a desired accuracy. In other words, the accuracy of authentication can be maintained at a high level by reducing the FAR, while reducing the FRR and improving the convenience without being affected by the noise at the time of authentication or noise at the time of registration.

  Particularly in the present embodiment, the threshold value TH is set based on the knowledge that the volume ratio EN_SN and the suitable threshold value TH satisfy a straight line relationship according to the volume ratio difference value DIF_SN. Therefore, there is an advantage that the optimum threshold value TH can be specified with high accuracy while sufficiently reducing the number of variables necessary for setting the threshold value TH. Furthermore, since the threshold value TH is corrected according to the correlation value NOISE_DIF and the utterance length EN_SPEECH_LEN, the threshold value TH that faithfully reflects the environment of the voice authentication device 100 at the time of initial registration or authentication can be used for authentication. .

<C: Modification>
Various modifications can be made to the above embodiment. An example of a specific modification is as follows. In addition, you may combine each following aspect suitably.

(1) Modification 1
In the above embodiment, the configuration in which the table stored in the storage unit 35 is used is exemplified. However, the initial value of the threshold value TH and the correction values A1 and A2 are selected by a calculation process using a predetermined calculation formula. Also good. For example, a plurality of mathematical expressions representing straight lines corresponding to the sound volume ratio difference value DIF_SN (each straight line in FIG. 4 defining the relationship between the sound volume ratio EN_SN and the threshold value TH) are stored in the storage unit 35, and the initial value setting unit 52 is The threshold value TH is calculated by substituting the sound volume ratio EN_SN calculated in step S10 into the mathematical formula corresponding to the sound volume ratio difference value DIF_SN calculated in step S12. Further, a configuration in which the correction unit 54 calculates the correction value A1 based on a mathematical expression representing the relationship between the correlation value NOISE_DIF and the correction value A1, and a correction unit 56 based on the mathematical expression representing the relationship between the utterance length EN_SPEECH_LEN and the correction value A2. A configuration for calculating the correction value A2 is also adopted.

(2) Modification 2
The variable that determines the degree of correction of the threshold value TH is not limited to the correlation value NOISE_DIF or the utterance length EN_SPEECH_LEN. For example, instead of the configuration in which the threshold TH is corrected according to the utterance length EN_SPEECH_LEN, or in addition to this configuration, the configuration in which the threshold TH is corrected according to the utterance length V_SPEECH_LEN of the authentication speech, or the average of the utterance length EN_SPEECH_LEN and the utterance length V_SPEECH_LEN A configuration in which the threshold value TH is corrected according to the value may be employed. For example, similarly to the relationship between the utterance length EN_SPEECH_LEN and the correction value A2, the threshold value TH is corrected so that the threshold value TH decreases as the utterance length V_SPEECH_LEN increases.

  Further, the threshold value TH may be corrected based on variables other than those listed above. For example, the threshold value TH may be corrected according to the ratio of the time length of voiced sound to unvoiced sound in the registered voice or authentication voice (speech section P2). The higher the ratio of voiced sound, the more accurate the distance DIST improves because the feature values (EN_DATA, V_DATA) are values that faithfully reflect the voice quality of the speaker. Therefore, even if the threshold value TH is lowered, the FRR does not rise unduly. Therefore, a configuration is adopted in which the threshold value TH is corrected so that the threshold value TH decreases as the ratio of voiced sound in the registered voice and authentication voice increases. In addition, since the accuracy of the distance DIST improves as the number of syllables (mora) in the registered voice or authentication voice increases, for example, the threshold TH is corrected so that the threshold TH decreases as the number of syllables in the registered voice or authentication voice increases. It is good also as composition to do.

(3) Modification 3
A configuration in which the relationship between the threshold value TH and each variable is variable is also employed. For example, the relationship between the volume ratio EN_SN or the volume ratio difference value DIF_SN and the initial value of the threshold TH may be changed by updating the table according to the operation of the operation unit 10. Similarly, a configuration in which the relationship between each variable and the degree of correction with respect to the threshold value TH is variable is also employed. For example, the correction value A1 corresponding to the correlation value NOISE_DIF and the correction value A2 corresponding to the utterance length EN_SPEECH_LEN can be changed according to the operation on the operation unit 10. According to these configurations, it is possible to realize authentication according to a user's request. Moreover, in the configuration in which the threshold value TH and the correction values (A1, A2) are calculated by calculating the mathematical formula as exemplified in the first modification, the content of the mathematical formula (for example, the coefficient of each variable) depends on the operation on the operation unit 10. May be changed.

(4) Modification 4
In the above embodiment, the configuration in which the distance DIST between the registered voice and the authentication voice is used for the authentication is exemplified, but the numerical value that is an index of the similarity between the two voices is not limited to the distance DIST. For example, a configuration in which authentication is performed based on a variable (index value) having a property that the numerical value increases as the characteristics of the registered voice and the authentication voice approach each other is also adopted. In this configuration, the magnitude relationship of the threshold value TH with respect to each variable is reversed from the above form. That is, for example, contrary to FIG. 4, the threshold value TH is set so that the threshold value TH decreases as the volume ratio EN_SN increases.

(5) Modification 5
Although the case where the threshold value TH changes linearly according to the volume ratio EN_SN is illustrated in the above embodiment, the relationship between the volume ratio EN_SN and the threshold value TH is appropriately changed. For example, the threshold value TH changes so that the volume ratio EN_SN and the threshold value TH satisfy the relationship of the curve according to the volume ratio difference value DIF_SN (that is, the volume ratio EN_SN changes along the curve according to the volume ratio difference value DIF_SN). Thus, a configuration in which the threshold TH is specified according to the volume ratio EN_SN and the volume ratio difference value DIF_SN is also employed. As described above, in the preferred embodiment of the present invention, the relationship between the volume ratio EN_SN and the threshold value TH differs according to the noise at the time of authentication (typically a linear or curved relationship according to the volume ratio difference value DIF_SN). It is sufficient that the threshold value TH is set so as to be equal to), and the specific relationship between the volume ratio EN_SN corresponding to one volume ratio difference value DIF_SN and the threshold value TH is, for example, a test for deriving the relationship of each variable. It is appropriately changed depending on the result and statistical processing on the result.

(6) Modification 6
In the above embodiment, the case where the voice authentication device 100 is used for distinguishing between a legitimate user and another person is exemplified. However, the voice authentication apparatus 100 is specified from a password and authentication voice registered in advance based on the voice of the legitimate user. The voice authentication apparatus 100 according to the above embodiment can also be used for voice password authentication for determining a match with a password.

It is a block diagram which shows the structure of the voice authentication apparatus which concerns on one form of this invention. It is a wave form diagram of the acoustic signal which an input part generates. It is a flowchart which shows the flow of the process by a threshold value setting part. It is a graph which shows the relationship between volume ratio EN_SN and volume ratio difference value DIF_SN, and threshold value TH. It is a conceptual diagram which shows the content of the table used for determination of a threshold value. It is a graph which shows the relationship between correlation value NOISE_DIF and correction value A1. It is a graph which shows the relationship between utterance length EN_SPEECH_LEN and correction value A2. It is a graph which shows the relationship between a threshold value, FAR, and FRR.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 ... Voice authentication apparatus, 10 ... Operation part, 15 ... Input part, 20 ... Characteristic analysis part, 22 ... Section detection part, 23 ... Switching part, 25 ... Noise analysis part, 26 ... Voice Analysis unit 28... Feature analysis unit 32... Storage device 35... Storage unit 40... Authentication unit 42. ... initial value setting section, 54 ... correction section, 56 ... correction section, 60 ... output section.

Claims (7)

A characteristic analysis means for analyzing characteristics of noise generated during authentication around the person to be authenticated;
A threshold is set so as to satisfy the relationship according to the characteristic of the noise during authentication analyzed by the characteristic analysis unit , with respect to the volume ratio between the registered noise and the registered voice at the time of registration of the registered voice registered in advance. Threshold setting means;
Authentication means for authenticating the person to be authenticated according to a result of comparison with the threshold value of the registration voice and the index value and the threshold value setting means for indicating the class not feature quantity of the authentication voice taken from a person to be authenticated is set A voice authentication device comprising: The threshold value setting means sets the threshold value so that the threshold value changes along a straight line or a curve corresponding to the authentication noise with respect to a volume ratio between the registration noise and the registered voice.
The voice authentication device according to claim 1 . The threshold setting means is configured to respond to a difference between a volume ratio between the noise during registration and the registered voice, and a volume ratio between the noise during authentication and the authentication voice and a volume ratio between the noise during registration and the registered voice. Set the threshold so that the threshold varies along a straight line or curve
The voice authentication device according to claim 2 . The threshold value setting unit includes a correction unit that corrects a threshold value according to a difference between the authentication noise and the registration noise.
The voice authentication device according to any one of claims 1 to 3 . The threshold setting unit includes a correction unit that corrects the threshold according to a time length of the authentication voice or the registered voice.
The voice authentication device according to any one of claims 1 to 4 . Analyzing the characteristics of noise generated during authentication around the person being authenticated,
A threshold value is set so as to satisfy the relationship according to the characteristics of the analyzed noise at the time of authentication, with respect to the volume ratio between the registered noise and the registered voice at the time of registration of the registered voice registered in advance .
Voice authentication method for authenticating the person to be authenticated according to a result of comparison between the registration voice and the feature amount of the index value and the set threshold value that indicates the kind not the collected authentication voice from the person to be authenticated. On the computer,
Characteristic analysis processing that analyzes the characteristics of noise at the time of authentication that occurs around the person being authenticated during authentication,
A threshold is set so as to satisfy the relationship according to the characteristic of the noise during authentication analyzed by the characteristic analysis process, with respect to the volume ratio between the registered noise and the registered voice at the time of registration of the registered voice registered in advance. Threshold setting processing;
Authentication processing for authenticating the person to be authenticated according to a result of comparison with the threshold value set by the threshold value setting process and the index value indicating the kind not feature quantity of the authentication voice taken from the registration voice and the person to be authenticated A program that executes and.

Images