JPH08123475A - Method and device for speaker collation - Google Patents

Abstract

PURPOSE: To provide the speaker collation device which considers variance in a speaker and estimates a threshold value adaptive to secular variation of the features of a speech previously in a short time with a small feature sample quantity. CONSTITUTION: Code books by speakers at respective periods are generated on the basis of speech features by the speakers at different periods A-C and stored in a code book file 100. When the threshold value is determined, a code book 101 by the speakers at a period D of a 1st speaker is generated and on the basis of this code book and the past stored code books of the same speaker and other speakers by the speakers, the intra-code-book distance and the inter- code-book distance of the 1st speaker are derived. On the basis of correlative values determined by the speakers at specific time intervals, an intra-period- difference-speaker distance is found from the intra-code-book distance and the inter-period-difference-speaker distance is found from the inter-code-book distance. Then the initial threshold value is adjusted so as to obtain an equal error rate with those intra-speaker distance and inter-speaker distance.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speaker verification technique used in a speaker recognition system or a voice recognition system, and more particularly to a feature sample having a small threshold value for each speaker adapted to a change with time of a voice feature of a speaker. The present invention relates to a method for estimating a quantity in a short time and an apparatus for realizing the method.

[0002]

2. Description of the Related Art A speaker verification device is a device for determining whether an identification name expressed by a speaker matches a true identification name of the speaker itself. Normally, when performing speaker verification, a speaker identification name to be verified and a codebook corresponding to this identification name are registered in advance, and at the time of speaker verification, the actual voice of the speaker and the identification name are registered. It is input and the codebook designated by this identification name is compared with the actual voice of the speaker to detect the feature difference. When the feature difference is less than or equal to a predetermined threshold set for each speaker, the asserted identification name is a true identification name, and the speaker is determined to be the person. In other cases, the asserted identification name is a false identification name and the speaker is determined to be an impostor. As described above, in speaker verification, it is important to determine what value the speaker-specific threshold value is, and whether or not this value is an appropriate value greatly affects the speaker identification rate.

There are roughly two causes for erroneous recognition during speaker verification. One is a case in which the speaker recognizes that the identification name is false, even though the speaker expresses the true identification name. This false recognition rate (FRR: False)
Rejection Rate). The other is a case where the speaker recognizes a false name even though he / she has expressed a false name, and this false recognition rate is the false acceptance rate (FAR). Called. By the way, if the threshold value for each speaker is increased, the probability that the speaker is the true person increases even if the feature difference is large. Therefore, FRR is low, but FAR is high. On the contrary, if the threshold value for each speaker is lowered, the FAR becomes lower but the FRR becomes higher. In this way, the FRR and the FAR have a relationship such that when one becomes lower, the other becomes higher. Since the false recognition rate is represented by the average value of both, it is preferable to adjust the threshold for each speaker so that the average value of both is as small as possible.

Conventionally, various methods for determining this threshold have been proposed. The first method is FRR and FAR
There is an equal error rate setting method that sets speaker-specific thresholds so that and become equal, and is introduced in Chapter 9 of "Digital Speech Processing" (Author: Sadahiro Furui; Publisher: Tokai University Press). FIG. 4 is a block diagram for realizing the equal error rate setting method. The principal learning voice and the impostor learning voice are input to the voice input terminal 400, and the preprocessing unit 401 outputs each voice for a predetermined time length. Store for each frame. The feature amount extraction unit 402 extracts the feature amount of each voice. The vector quantizer 403 vector-quantizes the respective feature amounts extracted from the voice based on the speaker codebook 404 corresponding to the identification name, and the same-speaker distortion distance (hereinafter referred to as “distortion distance within the same speaker” of the code vector thus obtained. , Intra-speaker distance) and the inter-speaker distortion distance (hereinafter, inter-speaker distance) are stored in the intra-speaker / inter-speaker distance storage unit 405. The FRR / FAR calculation unit 406
The FRR using the intra-speaker distance and a predetermined initial threshold value.
And FAR are calculated using the inter-speaker distance and the initial threshold. The FRR / FAR comparison unit 408 compares the values of FRR and FAR, and if they are not equal, the threshold adjustment unit 407 adjusts the initial threshold, and F
Return to the RR / FAR calculation unit 406. And FRR and FA
When R becomes equal, the adjustment is finished, and the value is output as the threshold of the speaker.

As a second method, a method of setting a threshold in consideration of the distribution of inter-speaker distances (S. Furui, "Cepstral
Analysis Technique for Automatic Speaker Verifica
tion, "IEEE Trans. Acoustics, Speech, and Signal Pro
cessing, vol.ASSP-29, No.2, pp.254-272, April 1981
(See) is known. FIG. 5 is a block diagram for realizing this method. After the learning voice is input to the voice input terminal 500, the preprocessing unit 501, the feature amount extraction unit 502,
The configuration up to the vector quantizer 503 and the speaker codebook 504 is the same as that shown in FIG. The feature of this method is that the inter-speaker distances obtained by the vector quantization unit 503 are stored in the inter-speaker distance storage unit 505, and the inter-speaker standard deviation calculation unit 506 and the inter-speaker average value calculation unit are respectively stored. In 507, the average value and standard deviation of the inter-speaker distance are obtained, and the threshold value calculation unit 508 derives the threshold value based on the statistical parameter obtained as a result.

As a third method, the "speaker verification method and apparatus" disclosed by the present inventors (Japanese Patent Application No. 6-4).
1615 specification). This method selects the principal codebook corresponding to the identification name asserted by the speaker from the plurality of codebooks and the other-speaker codebooks other than that, and selects a predetermined amount of code vector that appears from the other-speaker codebook. The statistical value of the feature difference from the person's codebook is calculated, and the threshold value is obtained from this. In other words, the feature is that the distance between codebooks is converted into the distance between speakers.
This method is realized by the blocks 600 to 612 in FIG.

[0007]

Each of the above-mentioned methods is a method of determining a threshold value based on a characteristic sample collected at a specific time, and considers that the characteristic of human voice changes with time. Absent. Therefore, the identification rate of the speaker may decrease with time. In order to estimate the threshold value that adapts to changes in the characteristics of human voice over time, it is possible to deal with it by creating a standard pattern of speech characteristics by using a feature sample for each speaker as long as possible. However, regardless of whether the long-term feature sample is stored as it is or the voice feature is extracted and stored, a huge memory capacity must be secured in the speaker verification device. There is a problem in that it takes a long time to calculate the threshold for each speaker because the number of etc. becomes enormous.

Further, since the first and second methods are methods of setting the threshold value ex post facto, they cannot be fully utilized in applications where the threshold value is set in advance by estimation or the like. Since the method of 3 does not consider the variation in the inter-speaker distance, there is a possibility of rejecting the person with a high probability at the time of speaker verification.

An object of the present invention is to solve the above problems and to estimate in advance a threshold value that takes into account variations in the distance between speakers and that adapts to changes over time in voice characteristics with a small amount of feature samples and in a short time. And a device for performing this method.

[0010]

According to the present invention, if the codebook size is equal to or larger than a certain value, the distance between codebooks and the distance between speakers, the distance between codebooks and the distance between speakers are
As shown in FIG. 2 and FIG. 3, respectively, there is a strong correlation, and these correlations are robust to the time difference. The feature is that the optimum threshold for each speaker is determined in consideration of the difference in the characteristics of the voice of the speaker.

FIG. 2 shows a feature quantity obtained by extracting a voice of a unit size sentence having a predetermined number of words with a predetermined frame length, for example.
FIG. 9 is a correlation measurement diagram between the intra-codebook distance and the intra-speaker distance based on eight sets of male voices when a set is used. For example, the codebook size is 512 and the time difference is 9 months. As is clear from this figure, if the speakers are the same, the two have a linear correlation, and if the intra-speaker distance is y and the intra-codebook distance is x, there is a relationship of y = ax + b according to the present invention. It became clear as a result of verification by the person. Although there are some variations in the coefficients a and b in this equation, there is a robust tendency in the time difference as a whole. In the example of FIG. 2, a was 0.944 and b was 0.101.

Further, FIG. 3 is an actual measurement diagram of the correlation between the inter-codebook distance and the inter-timing talker distance based on a set of female voices. As in the case of FIG. 2, the codebook size is 512, The time difference is 9 months. Referring to FIG.
Despite the set, it is clear that the linear correlation is the same as in FIG. 2, and this tendency is the same even when the number of sets is increased. Although not shown, the relationship between the inter-codebook distance based on the male voice and the inter-speaker distance was similar. That is, if the intra-codebook distance and the inter-codebook distance are known, the inter-speaker distance and the inter-speaker distance can be derived based on these correlations.

According to the speaker verification method of the present invention utilizing such a property, the first codebook is created based on the speech feature of the first speaker whose threshold is to be determined at any time, and A first correlation between the intra-codebook distance of the speaker and the inter-codebook distortion distance is derived, and further, the intra-codebook distance prepared for the first speaker and the time difference intra-speaker distance at the time. Based on the value and the second correlation value between the inter-codebook distance and the inter-timing talker distance, a process for deriving the inter-timing talker distance and the inter-timer talker distance at the time is performed. After deriving the inter-speaker distance and the inter-speaker distance, calculate the false rejection rate and the impostor acceptance rate based on each of these distances and an arbitrarily set initial threshold value, as in the first method of the related art. At the same time, the initial threshold value may be adjusted so that the false rejection rate and the impostor acceptance rate are equal.

As is apparent from FIGS. 2 and 3, the first and second correlation values are values obtained in advance at a predetermined time interval for each speaker, and the first correlation value is the distortion within the codebook. The linear correlation value between the distance and the time difference distortion distance within the same speaker, and the second correlation value is the linear correlation value between the inter-codebook distortion distance and the inter-speaker distortion distance.

Further, the speaker verification apparatus of the present invention utilizing the above characteristics creates a speaker-specific codebook for each period based on the speaker-specific voice characteristics of different periods, and prepares a speaker-specific codebook. It has means for storing the number of appearances of the code vector appearing in the process in the memory together with the speaker-specific codebook of the time. This means can be realized by using a known technique. Also, a codebook creating means for each target speaker that creates a first codebook of the first speaker for which a threshold value is to be determined, the created first codebook, and the first speaker that has been saved, Code vectors having the above-mentioned number of appearances are made to appear from the past codebooks of other speakers, and these code vectors are quantized by the code vector of the first codebook, and the distortion distance in the codebook of the first speaker, and A means for deriving the inter-codebook distortion distance between one speaker and another speaker, and the intra-codebook distortion distance measured in advance at a predetermined time interval for each speaker and the time difference distortion distance within the same speaker. Corresponding to the first speaker, and a correlation value storage means for storing a first correlation value between the codebook and a second correlation value between the inter-codebook distortion distance and the time difference distortion distance between other speakers. Read the first and second correlation values Then, a time difference distortion distance deriving means for deriving a time difference distortion distance within the first speaker and a time difference distortion distance between other speakers at the time concerned, and each time difference distortion distance and an arbitrarily set initial threshold value. And a threshold adjusting means for adjusting the initial threshold so that the false rejection rate and the impostor acceptance rate are calculated on the basis of the false rejection rate and the impostor acceptance rate.

[0016]

In the present invention, a plurality of speaker-specific codebooks at different times are prepared for each speaker instead of the long-term feature sample, and the number of times the code vector appears in each speaker-specific codebook is calculated. Save it. Further, preferably, the correlation shown in FIGS. 2 and 3, that is, the first correlation value and the second correlation value are stored in the correlation value storage means before the threshold value is determined. When determining the threshold of the first speaker, the code vector is represented according to the code string representing the code vector and the number of appearances of the code string from the code books of the first speaker and other speakers at each past time, and the threshold is set. The distance between codebooks and the distance between codebooks corresponding to the determination time are calculated. Next, the intra-codebook distance, the inter-codebook distance, the first correlation value and the second correlation value read from the correlation value storage means.
Based on the correlation value, the inter-time-speaker distance and the inter-time-speaker distance are derived.

As described above, it is possible to approximate the distance between speakers in the first speaker and the distance between speakers in advance by simply using the codebook created for each speaker in the past. The amount of memory used in the person verification device is significantly saved as compared with the conventional one. Incidentally, when the voice waveform is stored as it is by each conventional method (short type), the sampling frequency (1 / sec) × the voice duration (se
c) × 2 (bytes), when the feature vector is extracted from the speech waveform and stored (double type), a memory capacity of total number of frames × feature vector size × 16 (bytes) is required. In the case of the method and apparatus of the present invention, the memory usage is speaker codebook size × feature vector size × 16 (byte: double type) + speaker codebook size × 4 (byte: int type). Therefore, assuming that there are 10 word voices with a sampling frequency of 16 KHz, an average length of 3 seconds, an analysis frame length of 30 msec, and a frame period of 16 msec, feature samples are recorded every month for one year. If the codebook size is 256, the memory capacity required to store only the feature amounts of the feature samples of the 100 speakers recorded in 1. will be 1/11 times the conventional value. That is, according to the present invention, a memory capacity of about 92% can be saved. The same applies to the amount of calculation, and according to the present invention, the amount of calculation can be reduced by about 92% of the conventional amount. After the inter-speaker distance and the inter-speaker distance are derived, the first
The initial threshold value is adjusted so as to obtain an equal error rate by the same procedure or means as the method of (1), and this is set as the threshold value of the first speaker.

[0018]

Next, an embodiment of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram of a speaker-specific threshold value determining unit in a speaker verification apparatus according to an embodiment of the present invention.
Is a codebook file, 101 is a speaker-specific codebook (D-period speaker-specific codebook), 102 is a vector quantizer, 103 is an intra-codebook distance storage unit, 104 is an inter-codebook distance storage unit, and 105 is correlation. Converted value storage,
Reference numeral 106 denotes a distance conversion speaker distance conversion unit, 107 denotes a time difference speaker distance conversion unit, 108 denotes an FRR / FAR calculation unit, and 109.
Is an FRR / FAR comparison unit, and 110 is a threshold adjustment unit.

In the codebook file 100, a speaker-specific codebook storage unit for storing codebooks created at different times for each speaker, in the illustrated example, A to C, and a code vector of each speaker-specific codebook. And a code data storage unit that stores the number of appearances of each code are stored, and can be reused at any time as a past voice feature for each speaker. It is preferable that the intervals A to C are separated to some extent. This is because there may be no difference in the voice characteristics of the speaker in a short period. Further, the code or code string that has not appeared once is not stored in the code data storage unit. This can further be expected to reduce the size (memory usage) of the codebook file 100 and the amount of calculation thereafter. The code string is a set of codes given to each code vector, for example, corresponding to the centroid of each cluster, and the number of appearances of the code is the process until the vector quantization process is completed. In, the term refers to the appearance frequency data of codes assigned to the same cluster.

The D-speaker-specific codebook 101 extracts an analysis frame from the learning speech of the first speaker (A) whose threshold is to be set in the D period to obtain a feature quantity, and the feature quantity is a vector quantum. It is a codebook created by converting. The D period is an arbitrary period, but is later than the A period, B period, and C period described above.

The vector quantizer 102 uses, for example, the above D.
Based on the time-speaker-specific codebook 101, the same speaker (A) already stored in the codebook file 100
A code vector is made to appear from the code book of the periods A to C according to the number of times of appearance of each code, and this is output to the code book 101 for each speaker of the D period, for example. As a result, the characteristic difference of the codebook at the D period by the same speaker (A), that is, the codebook distance is obtained. This is stored in the intra-codebook distance storage unit 103.

Further, a code vector is made to appear according to the number of appearances of each code from the code book for each speaker of the other speakers (B) in the periods A to C, which is already stored in the code book file 100, and this is made D, for example. The codebook 101 for each speaker of the period is output. As a result, the codebook feature difference between the speakers and the other speakers at the time D, that is, the inter-codebook distance is obtained. This is the inter-codebook distance storage unit 10
Store in 4.

The correlation conversion value storage unit 105 stores, for each speaker,
Linear correlation equation and distance in codebook based on FIGS. 2 and 3
A correlation value uniquely derived from the inter-codebook distance is stored (correlation value storage means). The correlation value here is
It is a 1st correlation value showing the relation between the codebook inside distance and the time difference talker's distance, and the 2nd correlation value showing the relation between the codebook distance and the time difference talker's distance. These correlation values are read with each speaker as a key, and the first correlation value is input to the time difference talker distance conversion unit 106, and the second correlation value is input to the time difference talker distance conversion unit 107. In the case of this embodiment, the first and second correlation values relating to the speaker (A) and the other speaker (B) are read out.

The intra-timing talker conversion unit 106 and the inter-timer talker distance conversion unit 107 respectively talk at the time D based on the intra-codebook distance and the inter-codebook distance and the first and second correlation values. The inter-speaker distance and the inter-speaker distance of the person (A), and use them for FRR / FAR
Input to the calculation unit 108.

The FRR / FAR calculation unit 108 uses the input time difference speaker distance and an arbitrary initial threshold value to perform FRR.
And the FAR is calculated using the input inter-speaker distance and the initial threshold. The FRR / FAR comparison unit 109 compares the values of FRR and FAR, and if they are not equal, the threshold adjustment unit 110 adjusts the initial threshold value, and the process returns to the FRR / FAR calculation unit 108 again. Then, when FRR and FAR become equal, adjustment is completed,
The threshold value at that time is output as the threshold value of the speaker (A).

As described above, in this embodiment, the correlation conversion is performed in advance by obtaining the correlation between the intra-codebook distance and the inter-timing talker distance, and the correlation between the inter-codebook distance and the inter-timing talker distance. The codebooks are stored in the value storage unit 104, and a plurality of speaker-specific codebooks created at different times for each speaker are stored, and these speaker-specific codebooks are used to determine a threshold value. The distance between speakers and the distance between codebooks are derived, and the distance between speakers with different timing and the distance between speakers are calculated based on the above correlation. Therefore, except for the first period (when creating the codebook), The speaker-specific threshold can be determined at any time in each of the above. Further, the memory usage of the codebook file 100 is largely saved as compared with the conventional method, and the time required for threshold calculation is also shortened. As a result, the conventional problems can be solved. In the present embodiment, the first and second correlation values are stored in advance in the correlation conversion value storage unit 105 and the speaker whose threshold value is to be determined is read as a key. The present invention is not limited to this, and may be calculated at any time and input to the inter-timing talker distance conversion unit 106 and the inter-timer talker distance conversion unit 107.

[0027]

As is apparent from the above description, according to the speaker verification method of the present invention, based on the inter-codebook distance and the inter-codebook distance and the first and second correlation values acquired in advance. Since the inter-speaker distance and the inter-speaker distance are derived, it is not necessary to store a long-term feature sample when determining a speaker-specific threshold corresponding to a change in voice features over time. The memory usage of the speaker verification device can be saved. Furthermore, since the amount of feature samples is small, the calculation time is significantly shortened as compared with the conventional method, and there is an effect that the threshold for each speaker can be estimated in a short time.

Further, according to the speaker verification apparatus of the present invention, the codebook created for each speaker at different times in the past is reused, and the first speaker's inter-speaker distance and inter-speaker difference are preliminarily used. Since the inter-person distance can be approximately represented by the correlation value, it is not necessary to use an extra memory even when determining the speaker-specific threshold value that adapts to changes over time in the characteristics of the voice. In addition, since the feature sample amount is small and the calculation time for each distance is shortened, there is an effect that the entire configuration of the speaker-specific threshold value determining means can be simplified.

[Brief description of drawings]

FIG. 1 is a block diagram of a main part of a speaker verification device according to an embodiment of the present invention.

FIG. 2 is an actual measurement diagram showing a correlation between an intra-codebook distance and an intra-speaker distance.

FIG. 3 is an actual measurement diagram showing a correlation between an inter-codebook distance and an inter-speaker distance.

FIG. 4 is a block diagram for realizing an equal error threshold setting method that is a first conventional method.

FIG. 5 is a block diagram for realizing a threshold value setting method that takes into consideration the distribution of inter-speaker distances, which is the second conventional method.

FIG. 6 is a block diagram for realizing a method of converting a distance between codebooks into a distance between speakers, which is a third conventional method.

[Explanation of symbols]

100 codebook file 101 speaker-specific codebook created at an arbitrary time, for example, D time 102 vector quantization unit 103 intra-codebook distance storage unit 104 inter-codebook distance storage unit 105 correlation conversion value storage unit (correlation value storage means ) 106 time difference talker distance conversion unit 107 time difference talker distance conversion unit 108 FRR / FAR calculation unit 109 FRR / FAR comparison unit 110 threshold adjustment unit

Claims (5)

[Claims] 1. A speaker verification method having a threshold value determining process for determining a speaker-specific threshold value used for distinguishing from another speaker, wherein the threshold value determining process is performed by any one of the first speakers whose threshold value is to be determined. While creating the first codebook based on the voice features of the period,
This first codebook and the first created before that period
In-codebook distortion distance from the speaker's codebook,
And the inter-codebook distortion distance between the first codebook and the codebook of another speaker created before the period, respectively, and further, the intra-codebook distortion distance prepared for the first speaker and Based on the first correlation value between the first speaker's time difference distortion distance and the second correlation value between the inter-codebook distortion distance and the other speaker's time difference distortion distance, A speaker verification method comprising a step of deriving a time difference distortion distance within one speaker and a time difference speaker distortion distance between a first speaker and another speaker. 2. The threshold value determining step further calculates a false rejection rate and an imposter acceptance rate based on each time difference distortion distance at the relevant time and an arbitrarily set initial threshold value, 2. The speaker verification method according to claim 1, further comprising the step of adjusting the initial threshold value so that the speaker acceptance rates become equal. 3. The first and second correlation values are values obtained in advance at a predetermined time interval for each speaker, and the first correlation value is the difference in distortion within the codebook and the time difference within the same speaker. The linear correlation value with a distortion distance, and the second correlation value is a linear correlation value between the inter-codebook distortion distance and the inter-speaker distortion distance. Person verification method. 4. A speaker verification apparatus having a speaker-specific threshold value determining means used for distinguishing from another speaker, wherein a speaker-specific codebook for each period is created based on speaker-specific voice characteristics of different periods. Means for storing the number of appearances of the code vector appearing in the process of creating the codebook for each speaker in the memory together with the codebook for each speaker at the time, and the first speaker of the first speaker for which the threshold value is determined. Codebook creating means for creating a codebook for each target speaker, the created first codebook, and the first stored
A code vector having the above-mentioned number of appearances is made to appear from the past codebooks of the speaker and another speaker, and these code vectors are quantized by the code vector of the first codebook to distort the distance within the codebook of the first speaker. , And a means for deriving an inter-codebook distortion distance between the first speaker and another speaker, and a time difference within the same speaker and the intra-codebook distortion distance measured in advance at a predetermined time interval for each speaker. Correlation value storage means for storing a first correlation value between the distortion distance and a second correlation value between the inter-codebook distortion distance and the time difference distortion distance between other speakers, and the first speaker. And a time difference distortion distance deriving means for deriving the time difference distortion distance within the first speaker and the time difference distortion distance between other speakers at the time concerned by reading the first and second correlation values corresponding to A speaker verification device characterized by the above. 5. The speaker verification apparatus according to claim 1, wherein the false rejection rate and the impostor acceptance rate are calculated based on each time difference distortion distance and an arbitrarily set initial threshold, and the false rejection rate and spoofing are calculated. And a threshold adjusting unit that adjusts the initial threshold so that the acceptance rates of the speakers are equal to each other.