JP2005300958A - Talker check system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To suppress noise effects when checking talkers. <P>SOLUTION: A talker check system concerning this invention uses a talker check device 1 which picks up a part of the input voice as a check section and compares it with the registered voice to recognize the identity between this registered voice and the input voice. This talker check system 1 has a voice power calculating means 107 to obtain the voice power S in the picked up check section and a noise section voice power calculating means 108 to compute a part of the voice power N of the input voice preceding the check section, and a talker discriminating means 108 which compares the check section with the registered voice to recognize the identity between those talkers and decides that the talkers of these voices are not the same talker when the value SNR obtained by dividing S by N falls the predetermined threshold TH. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a speaker verification device that determines the identity of a person who utters registered speech and verification speech, and in particular, improves the accuracy of identity determination processing in consideration of the effect of noise on verification processing. Related to technology.

  In mobile communication terminals such as mobile phones, a speaker verification function using a voice processing technique is attracting attention as a method for easily locking and unlocking without performing key operations. Also in automobiles, it has been studied to increase the added value of a product by a speaker verification function that performs locking and unlocking by voice instead of locking and unlocking by keys. In home appliances, a user interface that can easily lock and release an operation so that a child does not perform an erroneous operation is required, and a speaker verification function is considered to be an effective solution for that purpose.

  These speaker verification functions must all be premised on being used in an environment with a high noise level. For example, a mobile phone is used in an environment where the noise level is extremely high, such as in a station yard or on the street, and in the case of an automobile, road noise must be taken into consideration. For home appliances, it is necessary to configure a speaker verification function so that it does not react to daily noise such as the sound of a TV, air conditioner, or vacuum cleaner.

  As described above, as a speech recognition technique or a speaker verification technique in consideration of the influence of noise, an S / N ratio between a test voice on which noise is superimposed and noise is obtained, and a plurality of frequency characteristics are obtained. Therefore, a technique for selecting a preferable frequency characteristic based on the S / N is known (Patent Document 1).

Japanese Patent Application Laid-Open No. 5-197387 “Speech Recognition Method”

  When the speaker verification device is left in a noise for a very long time, the registration data happens to match the noise data without the registrant speaking, and there is a possibility that an incorrect verification may be output. . More specifically, the speaker verification function calculates a distance value between a registered voice prepared in advance and an input voice uttered at the time of verification, and determines whether the distance value is below a predetermined threshold value. However, it is determined that the collation is performed when the threshold value is below the threshold value. However, there is a problem that the collation threshold value that is optimal in a quiet environment is not necessarily optimal under noise.

  The speaker verification apparatus according to the present invention obtains a ratio S / N between the acoustic power S of a section detected and verified as speech and the acoustic power N of a section immediately before and immediately after the verified section, and S / N If the input signal is small, the input speech is rejected to prevent erroneous verification due to noise.

The speaker verification device according to the present invention selects a part of the input speech as a verification interval, compares the verification interval with the registered speech, and determines the speaker identity between the registered speech and the input speech. In person verification device,
A matching section acoustic power calculating means for calculating the acoustic power S of the selected matching section;
A noise section sound power calculating means for calculating a sound power N of a part of the input speech preceding or following the matching section;
The collation interval is collated with the registered voice to determine the speaker identity between the registered voice and the input voice, and when the value SNR obtained by dividing S by N falls below a predetermined threshold TH, these voices Speaker determination means for determining that the speakers are non-identical,
It is equipped with.

  Thus, the speaker verification apparatus according to the present invention acquires the noise level based on the acoustic power of the section preceding the verification section, and the ratio S / of the acoustic power S of the verification section and the acoustic power N of the noise section. When N is small, erroneous verification can be prevented by failing speaker verification.

Next, embodiments of the present invention will be described with reference to the drawings.
Embodiment 1 FIG.
1 is a block diagram showing the configuration of a speaker verification apparatus according to Embodiment 1 of the present invention. In the figure, a speaker verification device 1 includes a voice input unit 101, an acoustic analysis unit 102, a voice segment detection unit 103, a registration unit 104, a registration data storage unit 105, a continuous DP matching unit 106, and an average acoustic power calculation unit 107 in the verification segment. , A pre-matching section average sound power calculation unit 108, an SNR calculation unit 109, a determination unit 110, and a notification unit 111 are provided.

  In the speaker verification device 1, a voice input unit 101 is a device that converts a voice signal collected by a microphone into a digital signal. The acoustic analysis unit 102 calculates, from the speech data converted into digital data, a part (circuit or circuit) that calculates an acoustic feature amount that is a feature suitable for speaker verification (for example, a well-known feature such as a mel cepstrum or acoustic power). Element). The voice section detection unit 103 is a part that detects an actual utterance section by detecting the rise and fall of the sound power from features such as the sound power. The registration unit 104 is a part for registering the acoustic feature amount of the detected section in the registration data storage unit 105 described later. The registered data storage unit 105 is a storage element or circuit for storing acoustic feature quantities, or a storage medium.

  The continuous DP matching unit 106 is a part that calculates the distance between the acoustic feature quantity of the input speech and the acoustic feature quantity of the registered data registered in the registered data storage unit 105 by the continuous DP method. The average acoustic power calculation unit 107 in the verification section is a part that calculates the average acoustic power in the verification section from the acoustic power in the verification section detected by the continuous DP matching unit 106. The pre-matching section acoustic power calculation unit 108 is a part that calculates the average sound power from the sound power in the predetermined section immediately before the matching section. The SNR calculation unit 109 is a part that calculates the ratio between the average acoustic power in the verification section and the acoustic power in the predetermined section immediately before the verification section. The determination unit 110 is a part that determines the success or failure of collation by comparing the calculation results of the continuous DP matching unit 106 and the SNR calculation unit 109 with predetermined conditions.

  The notification unit 111 is a unit for notifying a speaker who is a user visually, auditorily, or tactilely about the determination result of the determination unit 110. In order to visually notify the speaker, the notification means 111 may be configured by using a device such as an LED (light emitting diode), a lamp, or a display, and a speaker is used to audibly notify the speaker. It may be used. In addition, in order to notify the speaker of the determination result tactilely, the notification means 111 is configured by a vibrator function using an eccentric motor.

  The average acoustic power calculation unit 107 in the matching section is an example of a matching section acoustic power calculation unit. The pre-matching section acoustic power calculation unit 108 is an example of a noise section acoustic power calculation unit. The SNR calculator 109 and the determination unit 110 constitute an example of a speaker determination unit. Further, since the average acoustic power calculation unit 107 in the verification interval and the acoustic power calculation unit 108 before the verification calculate the power, it is not necessary to be provided as an independent part, and the process of the continuous DP matching unit 106 is performed. If the power is calculated at, the power calculated there may be regarded as the output of the intra-collation average acoustic power calculation unit 107 and the pre-collation interval acoustic power calculation unit 108.

  In addition to the configuration shown in FIG. 1, control units such as a central processing unit (CPU), a computer, and a DSP (Digital Signal Processor) having a general-purpose control function are provided. You may employ | adopt the structure which controls. Further, the configuration of FIG. 1 may be replaced with a combination of a control program and a computer program that causes a CPU or computer to execute processing corresponding to each part of FIG. However, hereinafter, each part in FIG. 1 will be described as performing the corresponding function.

  FIG. 2 is a front view of the speaker verification device 1 when configured as a mobile phone. In the figure, the part indicated by reference numeral 101a is a microphone provided in the voice input unit 101, and the notification means 111 corresponds to a liquid crystal display. The notification unit 111 may be configured using a speaker or a vibrator.

  Next, the operation of the speaker verification device according to Embodiment 1 of the present invention will be described. FIG. 2 is a flowchart showing the operation of registration processing of registered speech in this speaker verification apparatus. The speaker utters the keyword to be registered, and the voice input unit 101 samples the input voice and converts it into digital data (step S01). Next, the acoustic analysis unit 102 calculates an acoustic feature amount which is a feature suitable for speaker verification (for example, a well-known feature such as a mel cepstrum or acoustic power) from the voice data converted into digital data (step S02). Next, the speech segment detection unit 103 detects the actual speech segment by detecting the rise and fall of the acoustic power from the characteristics such as the acoustic power (step S03). Next, the registration unit 104 registers the acoustic feature amount of the detected section in the registration data storage unit 105 (step S04). Here, in order to confirm the stability of the utterance, two or more utterances may be uttered, and the average value of the data may be registered only when the mutual distance is small.

  Next, the operation at the time of verification in this speaker verification device will be described. FIG. 3 is a flowchart showing the operation of verification processing in this speaker verification device. The audio input unit 101 sequentially acquires audio, samples the acquired audio, and converts it into digital data (step S11). Next, the acoustic analysis unit 102 calculates an acoustic feature amount that is a feature suitable for speaker verification (for example, a well-known feature such as a mel cepstrum or acoustic power) from the speech data converted into digital data (step S12).

  Subsequently, the continuous DP matching unit 106 calculates the distance between the acoustic feature quantity of the input speech and the acoustic feature quantity of the registered data registered in the registered data storage unit 105 by the continuous DP method, and based on this distance value. The collation interval is detected (step S13).

となる。 Here, the continuous DP method will be described. The continuous DP method is a method in which DP matching is continuously performed from the beginning of input voice data. Specifically, an arbitrary restriction in the input voice is obtained by using a tilt restriction as shown in FIG. Is matched with the registered data. The distance between the acoustic feature quantity at the i-th time of the input speech and the j-th acoustic feature quantity of the registered data is expressed as d (i, j) as a local distance. In general, since the acoustic feature is expressed as a vector, if the dimension of the vector is K, the acoustic feature at the i-th time of the input speech is represented by (V (i) ₁ , V (i) ₂ , V ( i) ₃ ,..., V (i) _K ), and the acoustic feature quantities at the j-th time of the registered data are (R (j) ₁ , R (j) ₂ , R (j) ₃ ,..., R (j) _K )It can be expressed as. At this time, when using the Euclidean distance as the local distance,

It becomes.

Next, the cumulative distance g (i, j) between the i-th time of the input voice and the j-th time of the registered data is defined below.

At the same time, the path length (weight) is defined below.

として定義し、これが第ｉ_０時刻で所定値を下回った場合に時刻ｉ_０で照合したとする（以下、累積距離値という場合には正規化した累積距離値を指すものとする）。 Formulas 2 and 3 are calculated from j = 0 to j = J (J is the length of the registered data) for j at each time i. The cumulative distance of Equation 2 depends on the previous cumulative distance and the local distance, and by selecting the minimum value, ij pairs passing through the cumulative distance are sequentially determined. G (i ₀ , J) at a certain time i ₀ represents the minimum cumulative distance when the i ₀ time of the input voice is associated with the end of the registered pattern. At the same time, a route from the starting point of j = 0 to j = J is determined for the i-j pair, and this is called an optimal path. c (i ₀ , J) indicates the length of the optimum path. Therefore, the cumulative distance normalized by the length variation is

Defined as, and this is collated at time i ₀ if below a predetermined value at the i ₀ time (hereinafter, shall refer to the cumulative distance value normalized in the case of cumulative distance values).

In the continuous DP matching matching means, the optimum correspondence (optimum path) between the registered data and the input speech data input up to the current time i ₀ and the local at each point of the optimum path by the procedure as described above. The distance and the cumulative distance value when the optimum path is passed are obtained simultaneously. At this time, the optimum path is determined as shown in FIG. 4, for example. In the figure, M is the length of the collated utterance.

  The continuous DP matching unit 106 sets a section in which the accumulated distance G obtained by Equation 4 is equal to or less than a predetermined value as a matching section.

Next, the average acoustic power calculation unit 107 in the matching section calculates the average acoustic power (referred to as S) in the matching section from the acoustic power in the matching section detected by the continuous DP matching unit 106 (step S14). . Specifically, collating at time i, the collating sections are i−M + 1, i−M + 2,..., I−1, i, and their acoustic powers are P (i−M + 1), P (i−M + 2),. When P (i), S is calculated as follows. M is the length of the collated voice.

Next, the average acoustic power calculation unit before collation section 108 calculates the average acoustic power (referred to as N) from the acoustic power in the predetermined section immediately before the collation section (step S15). Specifically, when the length of the immediately preceding predetermined section is L, N is calculated as follows.

  These acoustic powers are temporarily stored in the collation interval average acoustic power calculation unit 106 and the pre-collation interval average acoustic power calculation unit 107 as the acoustic power at each time sequentially output by the acoustic analysis means, and the collation interval Each of the average sound power is calculated by going back those data.

Next, the SNR calculator 109 calculates the ratio of S and N (S / N) obtained in steps S14 and S15, and sets this as the SNR (step S16). Next, the determination unit 110 compares the accumulated distance value G output by the continuous DP matching unit 106 and the verification section with a predetermined verification threshold value TH _G , G <TH _G , and SNR and the predetermined threshold value TH _SNR. When SNR <TH _SNR, it is determined that the speakers of the registered voice and the input voice are the same (step S17: the same). If either condition of G <TH _G and SNR <TH _SNR is not satisfied, the speakers are determined to be non-identical (step S17: non-identical), and the process proceeds to step S18. Here, the SNR is described as the ratio of S and N, but log (S / N) obtained by taking the logarithm of the ratio of S and N may be used as the SNR.

  In step S18, the determination status is notified to the speaker via the notification means. FIG. 5 is a diagram showing an example of the determination status displayed on the display of the mobile phone by the notification means 111. In the figure, 111a is a character string indicating that the mobile phone is currently in speaker verification, and 111b is a character string indicating the result of speaker verification. These are merely examples of the configuration method of the notification unit 111, and any method may be used as long as the determination status is understood by the user. By providing the notification means 111 in this way, when the collation process fails, the user understands that fact and tries to collate again, or takes measures such as moving to a place with less noise. Will be able to. Then, it returns to step S11 and acquires the audio | voice data of the next time.

  As is clear from the above, according to the speaker verification device according to Embodiment 1 of the present invention, the SNR is obtained when the average acoustic power of the detected verification section is not larger than the average acoustic power immediately before it. Becomes a smaller value, and step S17 is determined to be false (not identical). As a result, it is possible to avoid erroneous collation due to coincidence between the acoustic data of the noise section and the registered data in a noisy environment.

  In the configuration described above, the registration process may be performed by a device other than the speaker verification device 1, and only the registered voice may be transferred to the speaker verification device 1 via flash memory, communication data, or the like. Needless to say, it is good.

  Further, the notification unit 111 may not be an essential component depending on the application of the speaker verification device 1. For example, if the purpose is to lock / unlock a car, informing the unidentified person that it can be unlocked by voice if it is inadvertently notified that the verification has failed, the safety is threatened. Is also possible. Therefore, if it is desirable for safety to not know that the speaker verification function is operating in this way, it is better to display nothing even if the speaker verification fails, so the notification means 111 is omitted. It is good to configure.

  Further, in the first embodiment of the present invention, an explanation will be given by taking an example in which the noise section acoustic power calculation means is configured using the pre-matching section acoustic power calculation unit 108 that calculates the sound power of the section preceding the matching section. did. However, it goes without saying that the acoustic power of the section following the verification section (the section immediately after the verification section) may be calculated instead of the process of calculating the power of the section preceding the verification section. In this case, N (i) may be obtained by calculating the total power of the immediately following section in Equation 6 and dividing this total by the length L.

Embodiment 2. FIG.
The speaker verification device according to Embodiment 1 is configured to constantly reject the input speech by comparing the calculated SNR with a certain threshold value. However, other threshold values may be set depending on the scene. The speaker verification apparatus according to the second embodiment has such a feature.

  FIG. 7 is a block diagram showing a configuration of a speaker verification apparatus according to Embodiment 2 of the present invention. In the figure, a matching threshold value setting unit 201 is a part that recalculates and resets the matching threshold value. The other components having the same reference numerals as those in FIG. 1 are the same as those in the first embodiment, and thus the description thereof is omitted.

  Next, the operation of the speaker verification device (hereinafter referred to as speaker verification device 1) according to Embodiment 2 of the present invention will be described. Since the operation at the time of registration is the same as that of the first embodiment, description thereof is omitted. FIG. 8 is a flowchart showing the operation of the speaker verification device 1 during verification. In the figure, steps S11 to S13 are the same as those in the first embodiment. In step S24, the pre-matching section average sound power calculation unit 108 calculates the average sound power N from the sound power in the predetermined section immediately before the matching section. Specifically, collating at time i, the collating sections are i−M + 1, i−M + 2,..., I−1, i, the length of the immediately preceding predetermined section is L, and the acoustic power at time i is P ( When i), N is calculated by Equation 6.

  The acoustic power at each time sequentially output by the acoustic analysis unit 102 is temporarily stored and held by the pre-matching section average acoustic power calculation unit 108. Then, when the verification section is detected, the average sound power is calculated by going back to those data.

と設定する。すなわち、騒音が小さい場合には照合閾値を小さくするとともに、騒音が大きくなると照合閾値もそれに伴って大きくしていく。ただし、照合閾値に下限（ＴＨ_ａ）と上限（ＴＨ_ａ＋（ＴＨ_Ｍ−ＴＨ_Ｎ）×ｗ）を持たせるようにしている。その他、ＴＨをＮの任意の増加関数の関数値として算出するようにしてもよい。こうすることで、騒音レベルが大きくなった場合に照合閾値が緩くなった結果、本人棄却率が増加してしまうこと防ぐことができるのである。なお、Ｎは対数をとった対数音響パワーとしても実現可能である。 Next, the matching threshold value setting unit 201 determines the matching threshold value TH _G according to a predetermined correspondence relationship from N calculated by the acoustic power calculation unit 108 before the matching section. Specifically, TH _a , TH _N , TH _M , w are set as predetermined values, respectively.

And set. That is, when the noise is low, the verification threshold value is decreased, and when the noise increases, the verification threshold value is increased accordingly. However, the collation threshold value has a lower limit (TH _a ) and an upper limit (TH _a + (TH _M −TH _N ) × w). In addition, TH may be calculated as a function value of an arbitrary increase function of N. By doing so, it is possible to prevent the person rejection rate from increasing as a result of the collation threshold becoming loose when the noise level increases. Note that N can also be realized as a logarithmic acoustic power with a logarithm.

Next, the determination unit 110 determines that the speakers are the same when G <TH _G for the distance value G output from the continuous DP matching unit 106 and the verification threshold value TH _G set by the verification threshold value setting unit 201. (Step S26). If G <TH _G is not satisfied, the voice data at the next time is acquired assuming that the speakers are not identical (step S26).

  With the above processing, an optimum threshold value can be set even when the speaker verification device is used in an environment where the noise level varies, so that it is possible to prevent the person rejection rate from increasing.

と設定する。これは音声と騒音の音響パワーの比が大きい場合には照合閾値を小さくして、比が小さくなると照合閾値もそれに伴って大きくしていくが、照合閾値は上下限をもっている場合である。このようにすると、音声と騒音のパワー比が小さくなった場合に照合閾値が緩くなるので本人棄却率が増加することを防ぐことができる。なお、閾値の算出方法は数８に限定されるものではなく、任意の増加関数を用いて算出してよい。また、音声と騒音のパワー比Ｓ／Ｎの対数をとったｌｏｇ（Ｓ／Ｎ）をＳＮＲの値に用いてもよい。 In the above description, the threshold value TH _G is calculated as an increase function of the acoustic power N calculated by the pre-matching section average acoustic power calculation unit 108, but the threshold value TH is used as an increase function of the SNR instead of the acoustic power N. _G may be calculated. Specifically, for example

And set. This is a case where the collation threshold value is reduced when the ratio of the sound power of the voice and the noise is large, and the collation threshold value is increased with a decrease in the ratio, but the collation threshold value has upper and lower limits. In this way, when the power ratio between voice and noise becomes small, the collation threshold value becomes loose, so that it is possible to prevent the person rejection rate from increasing. Note that the threshold calculation method is not limited to Equation 8, and may be calculated using an arbitrary increase function. Further, log (S / N) obtained by taking the logarithm of the power ratio S / N between voice and noise may be used as the value of SNR.

In the above-described processing, the configuration example is shown so that the collation threshold value THG is recalculated. However, the configuration is the same as in the first embodiment, and the threshold value TH _SNR to be compared with the _SNR is recalculated instead of the collation threshold value THG. You may comprise.

  Needless to say, the speaker verification apparatus according to the second embodiment of the present invention may also be provided with notification means as in the first embodiment.

Embodiment 3 FIG.
The speaker verification device according to Embodiment 2 of the present invention is configured to apply different thresholds depending on acoustic power and SNR. In addition, you may comprise so that a different threshold value may be applied according to a use situation (contents of operation to unlock). The speaker verification apparatus according to Embodiment 3 of the present invention has such a feature.

FIG. 9 is a block diagram showing a configuration of a speaker verification apparatus according to Embodiment 3 of the present invention. 1 differs from FIG. 1 in that the determination unit 110 is configured to refer to the registered data storage unit 105. Further, in the speaker verification device (speaker verification device 1) according to the third embodiment of the present invention, the configuration of the registration data storage unit 105 is different. FIG. 10 is a diagram illustrating a configuration example of data stored in the registration data storage unit 105 when the speaker verification device 1 is configured as a mobile phone. As shown in the figure, the registered data storage unit 105 according to the third embodiment stores a threshold field and a registered voice data field for each command. The threshold field indicates the contents of a command that can be executed by voice by the mobile phone. On the other hand, the registered voice data is a registered voice registered in advance by the user to activate this command. Here, for the sake of simplicity, the registered voice data is written in Roman letters, but since it is actually voice data, features such as a mel cepstrum and acoustic power are recorded instead of such a format. The threshold stores a TH _SNR to be compared with the SNR for the command. It should be noted that if this threshold is small, there will be more phases to be collated even if there is some noise, and if this threshold is large, it will be difficult to perform collation because it is strict against noise generation. is there.

  For example, in the example of FIG. 10, 2.0 is set as the threshold value for receiving and speaking. On the other hand, if this mobile phone is equipped with a digital camera function and the shutter is released when the speaker speaks “high, cheese”, in such applications, noise resistance (due to noise) Since it is thought that the photo opportunity should be more important than the need to keep the rate of occurrence of erroneous verification low, set the threshold to be smaller than other commands and operations, such as threshold 1.5. Yes. By doing so, it is possible to improve usability while maintaining safety by preventing erroneous verification due to noise.

The operation of the speaker verification device 1 is the same as that in the first embodiment. However, the determination unit 110 reads the TH _SNR stored in response to the registered voice command that the continuous DP matching unit 106 has determined to have the shortest distance from the registered data storage unit 105 during the matching process. The point is different.

  As is clear from the above, according to the speaker verification device of the third embodiment of the present invention, different thresholds are applied depending on the situation of use, so that erroneous verification due to noise is prevented and safety is maintained. Usability can be improved

Note that the threshold value may be reset as in the second embodiment using the TH _SNR stored in the registered data storage unit 105 for each command as an initial value. Further, as shown in the second embodiment, the upper limit value and the lower limit value of the threshold value may be stored in the registered data storage unit 105 for each command, and may be referred to in the verification processing by the determination unit 110. Needless to say.

  The present invention can be applied to, for example, a speaker verification device that determines whether or not to start using a device based on voice uttered by an operator.

It is a block diagram which shows the structure of Embodiment 1 of this invention. It is a front view at the time of comprising the speaker collation apparatus of Embodiment 1 of this invention as a mobile telephone. It is a flowchart which shows operation | movement of the speaker collation apparatus of Embodiment 1 of this invention. It is a flowchart which shows operation | movement of the speaker collation apparatus of Embodiment 1 of this invention. It is a figure for demonstrating the principle of operation of DP matching in the speaker collation apparatus of Embodiment 1 of this invention. It is a figure which shows the structural example of the alerting | reporting means of Embodiment 1 of this invention. It is a block diagram which shows the structure of Embodiment 2 of this invention. It is a flowchart which shows operation | movement of the speaker collation apparatus of Embodiment 2 of this invention. It is a block diagram which shows the structure of Embodiment 3 of this invention. It is a figure which shows the structural example of the data memorize | stored in the speaker collation apparatus of Embodiment 3 of this invention.

Explanation of symbols

105 registered data storage unit,
106 continuous DP matching unit,
107 average acoustic power calculation section in the verification section,
108 section average sound power calculation section before verification,
109 SNR calculator,
110 determination unit,
111 notification means,
201 Collation threshold value setting unit.

Claims (7)

In the speaker verification device that selects a part of the input speech as a verification interval, and compares the verification interval with the registered speech to determine speaker identity between the registered speech and the input speech,
A matching section acoustic power calculating means for calculating the acoustic power S of the selected matching section;
A noise section sound power calculating means for calculating a sound power N of a part of the input speech preceding or following the matching section;
The collation interval is collated with the registered voice to determine the speaker identity between the registered voice and the input voice, and when the value SNR obtained by dividing S by N falls below a predetermined threshold TH, these voices Speaker determination means for determining that the speakers are non-identical,
A speaker verification device comprising: The speaker verification device according to claim 1, wherein the speaker determination means changes TH based on N. 3. The speaker verification device according to claim 2, wherein the speaker determination means calculates TH as a function value of an increasing function of N, and compares TH with SNR. The speaker verification apparatus according to claim 1, wherein the speaker determination unit changes TH based on SNR. 5. The speaker verification apparatus according to claim 4, wherein the speaker determination means calculates TH as a function value of an increasing function of SNR and compares TH with SNR. 2. The speaker verification device according to claim 1, wherein the speaker determination means sets a different value to TH according to the type of registered speech to be verified, and compares TH and SNR. The speaker verification device according to claim 1, wherein the speaker determination unit notifies the speaker of the determination status via the notification unit when the SNR is lower than TH.

Images