JP2000330590A - Method and system for collating speaker - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and a system which can surely collate a speaker by a method performing the collation in a time domain. SOLUTION: This speaker collation system comprises a frequency band dividing part 1 for dividing an input speech (sampled input speech) into frequency bands, a non-linear compressing part 2 for operating non-linear compression of a speech signal in a prescribed frequency band of the speech signal divided into the frequency bands, a string-vectorizing part 3 for forming vector string from the non-linear compressed speech signal, a dictionary creating part 4 for creating a dictionary vector string as a dictionary, and a speaker judging part 5 for collating (judging) the speaker by operating the pattern matching between the vector string of the speech to be collated and the dictionary vector string of the dictionary creating part 4.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a speaker verification method and a speaker verification system.

[0002]

2. Description of the Related Art Conventionally, an HMM using a probability model is well known as a method of speaker verification, but a definitive method has not yet been established. In particular, the method of performing the matching in the time domain is very difficult even for the same speaker due to the nature of the voice that varies for each utterance.

[0003]

SUMMARY OF THE INVENTION It is an object of the present invention to provide a speaker verification method and a speaker verification system which can surely perform speaker verification by a method of performing verification in a time domain. .

[0004]

In order to achieve the above object, the invention according to claim 1 comprises a non-linear compression step of non-linearly compressing a sampled input voice and converting it into a low-resolution voice signal. A vector sequence forming step of forming a vector sequence from the linearly-compressed audio signal, and pattern matching between a dictionary vector sequence created in advance in the dictionary and a vector sequence of the voice to be verified are performed, and speaker verification and determination are performed. And a collation determining step.

According to a second aspect of the present invention, in the speaker verification method according to the first aspect, in the non-linear compression step, an appropriate number of sampled values of the sampled input speech are set to one.
After normalizing all sample values in a group with the maximum absolute value of the sample values included in the group, and then converting the normalized sample values to a low-resolution audio signal using a nonlinear quantization function It is characterized by.

According to a third aspect of the present invention, in the speaker verification method according to the first or second aspect, in the vector-sequence conversion step, the speech signal is converted into a vector sequence by associating a group with one vector. In the matching determination step, since the same voice of the same speaker includes similar voice patterns, the vector sequence also includes many similar vectors, and the similar vector is a vector having a high correlation value. Yes, a vector sequence including a large number of vectors having a high correlation value is determined to be a similar voice, and the determination is used for speaker verification determination.

According to a fourth aspect of the present invention, in the speaker verification method according to any one of the first to third aspects, when the vector sequence of the voice to be verified is constituted by one bit of the input voice, The method is characterized in that a speaker sequence is determined by sequentially collating a vector sequence shifted one bit at a time with a dictionary vector sequence by creating a corresponding vector sequence.

According to a fifth aspect of the present invention, there is provided a band dividing section for band-dividing a sampled input sound, a non-linear compressing section for non-linearly compressing a sound signal of a predetermined band among the band-divided sound signals, A vector-sequencing unit that forms a vector sequence from the non-linearly-compressed audio signal, a dictionary creating unit that creates a dictionary vector sequence as a dictionary, a vector sequence of a voice to be verified, and a dictionary vector sequence created by the dictionary creating unit 4 And a speaker determination unit for performing speaker matching and determination.

According to a sixth aspect of the present invention, in the speaker verification system according to the fifth aspect, the dictionary creation unit is configured to determine, for each vector, a vector sequence corresponding to each identical voice uttered by the same speaker. Is calculated for all vector columns, and the sum of the number of vectors whose correlation value is equal to or greater than a predetermined threshold is the largest vector column. Is a dictionary vector sequence of the voice, and a dictionary is registered with a dictionary vector sequence based on the voice of each speaker created by the dictionary creating unit.

According to a seventh aspect of the present invention, in the speaker verification system according to the fifth aspect, the non-linear compression section includes a sampled number of sampled input voices which is 1 in an appropriate number.
After normalizing all sample values in a group with the maximum absolute value of the sample values included in the group, and then converting the normalized sample values to a low-resolution audio signal using a nonlinear quantization function It is characterized by.

According to an eighth aspect of the present invention, in the speaker verification system according to the fifth aspect, the speaker determination unit corresponds to a vector sequence of the voice to be verified and a plurality of speakers registered in the dictionary, respectively. Calculate the correlation value for each vector with each dictionary vector string to be calculated, calculate the total number of vectors whose correlation value is equal to or greater than a predetermined threshold, and correspond to the dictionary vector string where the total number of vectors whose correlation value is equal to or greater than a predetermined threshold is maximum Is determined as the speaker of the voice to be verified.

According to a ninth aspect of the present invention, in the speaker verification system according to the fifth aspect, when the vector sequence of the voice to be verified is formed, the input voice is shifted one bit at a time, and the vector sequence corresponding to the input voice is shifted. It is characterized in that a speaker sequence is determined by sequentially comparing a vector sequence created and shifted one bit at a time with a dictionary vector sequence.

[0013]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a configuration example of a speaker verification system according to the present invention. Referring to FIG. 1, the speaker verification system includes a band dividing unit 1 for band-dividing an input voice (sampled input voice), and non-linearly compressing a voice signal of a predetermined band among the band-divided voice signals. A non-linear compression unit 2, a vector sequence conversion unit 3 that forms a vector sequence from the non-linearly compressed audio signal, a dictionary generation unit 4 that generates a dictionary vector sequence as a dictionary, and a vector sequence and a dictionary generation unit 4 of a voice to be verified. And a speaker determination unit 5 that performs pattern matching with the dictionary vector sequence of (1) and performs speaker verification (determination).

In the present invention, the non-linear compression section 2 non-linearly compresses an M-bit quantized audio signal value and converts it into a lower-bit (m-bit) audio signal, thereby providing the same speaker with the same signal. In the case of audio, fluctuations in the amplitude direction of the audio signal are suppressed, and a time series with high similarity is configured.

Assuming that the i-th n time series composed of m bits is one block (Bi), the block (Bi) can be represented by the following equation (Equation 1). That is, the block (Bi) is composed of n m-bit signals.

[0016]

Bi = {m bits, m bits,..., M bits}

Here, if the block (Bi) is considered as one vector, the converted digital audio signal can be regarded as an n-dimensional vector sequence. That is, the audio signal can be represented as an n-dimensional vector sequence.

The dictionary creating section 4 creates a dictionary vector sequence used for speaker verification by the above-described method.
It is stored in advance as a dictionary. And the speaker determination unit 5
When a voice to be verified is input, a vector sequence corresponding to the voice is constructed in the same procedure, and the most common dictionary vector sequence from a plurality of dictionary vector sequences corresponding to each of a plurality of speakers stored in the dictionary. By finding similar ones, speaker verification can be performed.

In the speaker verification system shown in FIG. 1, for example, a voice (input voice) sampled at 16 kHz is equally divided into eight channels by a band dividing unit 1, and a component of the lowest frequency band [0 to 1 kHz] is divided into eight channels. Speaker verification is performed using characteristic band components as needed.

Hereinafter, in the speaker verification system of FIG.
The method of compressing the voice amplitude in the non-linear compression unit 2, the method of constructing the vector sequence in the vector sequence conversion unit 3, the method of creating a dictionary (dictionary vector sequence) in the dictionary creation unit 4, the speaker verification in the speaker determination unit 5, Each of the determination methods will be described.

First, the processing operation in the nonlinear compression section 2 will be described. In the non-linear compression unit 2, for example, FIG.
As shown in the above, the audio signal (sample value of the input audio) quantized by M bits is grouped into n temporally continuous signals,
Let it be one block (group). Therefore, 1
A block (one group) is composed of n consecutive M-bit audio signals Ai. As described above, when the audio signal quantized by M bits (sample value of the input audio) is grouped into n blocks each, the non-linear compression unit 2 outputs n signals in one block (one group). Maximum value of absolute value of (sample value), block (1 group)
Are normalized (n sample values). Thereafter, the non-linear compression unit 2 calculates the normalized n sample values by using a predetermined non-linear quantization function as shown in FIG.
As shown in (b), the signal is converted into n m-bit signal trains (low-resolution signal trains).

The following equation (Expression 2) represents the converted n m-bit signal strings Bi.

[0023]

## EQU2 ## Bi = {bi1, bi2,..., Bin} =
f {Ai}

Here, f represents a non-linear quantization function.

As described above, according to the present invention, the nonlinear compression section 2
, The sampled values of the input voice sampled are grouped into an appropriate number, and all the sample values in the group are normalized by the maximum value of the absolute values of the sample values included in the group. By converting the audio signal into a low-resolution signal, fluctuation in the amplitude direction of the audio signal is suppressed.

Next, the processing operation in the vector sequencer 3 will be described. Assuming that an arbitrary audio signal S sampled by M bits has a length L, the number BN of blocks becomes as shown in the following equation (Equation 3).

[0027]

BN = Integer part of (L / n) +1

In the equation (3), +1 is used to add 0 when the length L of the audio signal S is divided when the length L is divided for each n to make one block. FIG. 3 is a diagram for explaining this fraction processing. As shown in FIG. 3, when the length L is divided for each n and it is desired to divide the length L into, for example, five blocks as in the example of FIG. 3, a fraction is generated in the last and fifth block. In this case, “5” can be added to the rear side of the fifth halfway block (filled with “0”) to make five blocks. In this way, in Equation 3, the value +1 is used to add 0 when the length L of the audio signal S is divided when the length L is divided for each n to make one block. Since the number of blocks is the number of vectors, an arbitrary audio signal S is represented by a BN n-dimensional vector sequence as shown in the following equation (Equation 4).

[0029]

S = {B ₁ , B ₂ ,..., B _BN }

As described above, the vector-sequencing unit 3 is configured to construct an arbitrary audio signal S sampled by M bits as a vector sequence as shown in Expression 4.

In the dictionary creating section 4, dictionary data is created by the same speaker generating the same voice several times, and generating a vector sequence corresponding to each voice. If the vector sequence corresponding to each voice is T1, T2,..., T1, T2
2,... Are as follows.

[0032]

T1 = ｛B1 _(T1) , B2 _(T1) ,... _{BBN (T1)} ｝ T2 = ｛B1 _(T2) , B2 _{(T2) ,. )} ｝ ・ ・ ・ ・

In general, the lengths of T1, T2,... Are different.

Next, a correlation value for each vector is calculated between all different vector sequences, and the total number of vectors having a correlation value of 0.9 or more is obtained. First, for all the vectors of T1, the correlation values with the vectors of the other vector columns are examined in order, and the sum SUM1 having the number of correlation values of 0.9 or more is obtained as in the following equation.

[0035]

SUM1 = INT (<B1 _(T1) , B2 _(T2) > / 0.9) + INT (<B2 _(T1) , B1 _(T2) > / 0.9) +. ... + INT (<B1 _(T1) , B2 _(T2) > / 0.9) + INT (<B2 _(T1) , B2 _(T2) > / 0.9 ) + ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ + INT (<B _{BN (T1)} , B _{BN (T2)} > /0.9) + ・ ・ ・ ・ ・ ・ ・ ・ ・

Here, <x, y> represents a correlation between x and y, and <*> represents an integer of *. INT (*) represents an integer of *.

Similarly, for all vectors in T2,
The correlation values with the vectors of the other vector columns are examined in order, and the sum of the numbers whose correlation values are 0.9 or more is obtained. The same operation is performed on all the vector columns, and the vector column having the maximum sum of the correlation values of 0.9 or more is set as the dictionary vector column of the voice of the speaker.

In particular, in the low frequency band, since a stationary characteristic of a vowel appears, the same voice of the same speaker contains similar components. Have a similar pattern.

As described above, the dictionary creating unit 4 calculates the correlation value for each vector between vector sequences corresponding to individual identical voices uttered by the same speaker (between all different vector sequences). , The value of which is a predetermined threshold (in the above example,
0.9) is calculated for all the vector columns, and the vector column having the maximum sum of the numbers whose correlation values are equal to or greater than a predetermined threshold value is defined as the dictionary of the speaker's voice. Vector strings, that is, dictionary data. In the dictionary, a dictionary vector sequence based on the voices of the individual (plural) speakers created as described above is registered.

Next, the speaker verification and determination processing in the speaker determination section 5 will be described. In the speaker verification and determination process, the correlation value for each vector between the input voice S (vector sequence of the voice to be verified) and each dictionary vector sequence corresponding to each of a plurality of speakers registered in the dictionary is described above. Similarly, a process is performed to obtain the total number of vectors having a correlation value equal to or greater than a predetermined threshold (for example, 0.9). Is determined as the speaker of the voice to be verified. This is because the similarity with the speech obtained as described above by other speakers is low, and the appearance of similar patterns is low.

That is, in the present invention, the speech signal is treated as a vector sequence by associating the group with one vector, and the same voice of the same speaker contains similar voice patterns. Also includes a large number of similar vectors, the similar vector is a vector having a high correlation value, and a vector sequence including a large number of the vectors having a high correlation value is determined to be similar voices. Used to determine

FIG. 4 shows that when six speakers utter the voice of the word "digital", the correlation value obtained from the voice of the same speaker and the voices of the other five speakers is 0.9 or more. It is a figure (graph) which shows a total number. In FIG. 4, as will be described later, since the speech to be verified is shifted one point at a time to create a vector sequence, and the correlation is calculated for the vector sequence, the horizontal axis of the graph in FIG. 4 indicates the number of shifts. The reason why the to-be-verified voice is shifted by one point in this way is to avoid a problem based on the fact that the to-be-verified voice is generally asynchronous with the dictionary data.

The operation of shifting the to-be-verified voice performed to alleviate the asynchronous problem and the troublesome voice segmentation will be described below with reference to the flowchart of FIG. First, after initially setting the number of shifts N to “1”, the number of correlation values between the (voice vector to be verified) S and the dictionary vector sequence is determined to be 0.9 or more (step S).
2). Next, the to-be-verified voice S is shifted by one sample (step S3), and the process returns to step S2, where the number of correlation values between the (to-be-verified) speech vector S and the dictionary vector sequence is 0.9 or more. In this way, the processes of steps S2 and S3 are repeated n times until the shift of n samples is completed.
(Steps S4 and S5). Then, the point at which the sum of the respective band components exceeds the predetermined threshold value is set as the start time of the sound (step S6). On the other hand, if each band component is equal to or less than a preset value with respect to the sum calculated in step S6, it is determined that there is no audio component in that band portion.

As described above, when constructing a speech vector sequence, the accuracy of speaker verification can be improved by shifting the input speech one bit at a time and creating a corresponding vector sequence.

FIGS. 6 to 10 are diagrams showing specific processing examples.

First, referring to FIG. 6, the original audio signal is an audio signal quantized to 8 bits by sampling at 16 kHz, and the audio signal (16
(kHz sampling 8-bit data) is band-limited to 1 kHz or less, and is converted to 2-kHz sampling 4-bit data by a non-linear quantization function.

Next, as shown in FIG. 7, a 2-bit sampling 4-bit data string is composed of 20 samples in one block. That is, a 20-dimensional vector sequence is formed. Since one block is formed by 20 points, a block, that is, a sound having n vectors is a sound signal of 20n points.

FIGS. 8 to 10 are diagrams for explaining what kind of pattern matching is performed for speaker verification. FIG. 8 shows a dictionary vector sequence {1B, 2B, 3B,... Of voices of a plurality of speakers (n speakers) registered in the dictionary.
, NB}. One dictionary vector string (one block), for example, 1B has a time length of 10 ms, and 2 of {1B1, 1B2,..., 1B20}.
Assume that it contains 0 points.

FIGS. 9 and 10 show voices to be verified # 1A,
2A, 3A,... JA} and a dictionary vector string {1B, 2
It is a figure which shows the pattern matching method of B, 3B, ..., nB}. Referring to FIG. 9, first, the first vector 1A of the speech to be verified is correlated with all the vectors {1B, 2B, 3B,..., NB} of the dictionary vector sequence, and the correlation value is set to 0. Find the number of 9 or more. In the example of FIG. 9, the first vector 1A of the voice to be verified and all the vectors {1B, 2B, 3B,.
, NB}, indicating that the number of the correlation values of 0.9 or more is four. Similarly, the correlation between the second vector 2A of the voice to be verified and all the vectors {1B, 2B, 3B,..., NB} of the dictionary vector sequence is obtained, and the number of the correlation values is 0.9 or more. Is found, the number of correlation values of 0.9 or more is one. Such processing is sequentially performed up to the last vector jA of the voice to be verified, and from the first vector 1A to the last vector jA of the voice to be verified, the correlation value 0.9 or more obtained as described above is obtained. Find the sum of the numbers. In FIG. 9, this sum S1 is shown as S1 = 12.

The vector sequence obtained from the voice to be verified is asynchronous with the dictionary vector sequence. Therefore, the original speech {1A, 2A, 3A,..., JA} is shifted by one point to recreate a new vector sequence as shown in FIG. 10, and in the same manner as described above (as shown in FIG. 9). Then, the sum of the number of vectors (blocks) having a correlation value of 0.9 or more is obtained. In this way, the total sum when shifted by one point is S2, and the total sum when shifted by 19 points is S20.

FIG. 4 shows the speech of the word “digital”
This is an example in which the sums S1 to S20 are plotted. As is apparent from FIG. 4, the number of correlation values of the same speaker of 0.9 or more is compared with the number of correlation values of other speakers of 0.9 or more. High values. Using this result, the following equation (Equation 7) is calculated.

[0052]

(Equation 7)

Here, S (h) _i is the number of correlation values 0.9 or more obtained by shifting the h-th dictionary by i−1 points, and Th is a predetermined threshold value. The number 7 calculates, speaker corresponding to h-th dictionary vector sequence D _S is maximum is obtained as a result matching.

If the voice to be verified is a time series longer than the dictionary voice, the total number of vectors having a correlation value of 0.9 or more may increase. Therefore, the voice to be verified is terminated at the length of the dictionary voice.

[0055]

As described above, according to the first to seventh aspects of the present invention, the sampled input voice is nonlinearly compressed and converted into a low-resolution voice signal, and the input voice is nonlinearly compressed. A vector sequence is formed from the voice signal, and pattern matching is performed between a dictionary vector sequence created in advance in the dictionary and a vector sequence of the voice to be verified, and speaker verification,
Since the determination is performed, speaker verification can be reliably performed by a method of performing verification in the time domain.

In particular, according to the second and seventh aspects of the present invention, an appropriate number of sampled sample values of the input speech are grouped into one group, and the maximum value of the absolute values of the sample values included in the group is calculated. After normalizing all the sample values in the group, the normalized sample values are converted into a low-resolution signal by a non-linear quantization function, so that the fluctuation of the audio signal in the amplitude direction can be suppressed.

According to the fourth and ninth aspects of the present invention, when forming a vector sequence of the voice to be verified, the input voice is shifted one bit at a time, and a vector sequence corresponding to the input voice is created. Since the speaker collation is determined by sequentially collating the vector sequence shifted by one with the speaker vector sequence, the accuracy of the speaker collation can be improved.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration example of a speaker verification system according to the present invention.

FIG. 2 is a diagram for explaining a processing operation in a non-linear compression unit.

FIG. 3 is a diagram for explaining fraction processing.

FIG. 4 is a diagram showing the total number of correlation values obtained from the voices of the same speaker and the voices of other speakers when the voice of the word “digital” is uttered by five speakers; is there.

FIG. 5 is a flowchart for explaining an operation of shifting a voice to be verified.

FIG. 6 is a diagram showing a specific processing example of the speaker verification method of the present invention.

FIG. 7 is a diagram showing a specific processing example of the speaker verification method of the present invention.

FIG. 8 is a diagram for explaining what pattern matching is performed for speaker verification;

FIG. 9 is a diagram for explaining what kind of pattern matching is performed for speaker verification.

FIG. 10 is a diagram for explaining what kind of pattern matching is performed for speaker verification.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Band division part 2 Non-linear compression part 3 Vector sequence formation part 4 Dictionary creation part 5 Speaker judgment part

Claims (9)

[Claims] A nonlinear compression step of nonlinearly compressing a sampled input voice into a low-resolution voice signal; a vector conversion step of forming a vector row from the nonlinearly compressed voice signal; A speaker matching method comprising: performing a pattern matching between a dictionary vector sequence created in advance and a vector sequence of a voice to be verified, and performing a speaker matching and determination process. 2. The speaker verification method according to claim 1, wherein
In the non-linear compression step, an appropriate number of sample values of the sampled input audio are grouped into one group, and all sample values in the group are normalized by the maximum absolute value of the sample values included in the group. Converting the normalized sample value into a low-resolution audio signal using a non-linear quantization function. 3. The speaker collating method according to claim 1, wherein, in the vector-sequencing step, the group is made to correspond to one vector, so that the audio signal is treated as a vector sequence, and the collation-determining step is performed. Then, since the same voice of the same speaker includes a similar voice pattern, the vector sequence also includes many similar vectors, and the similar vector is a vector having a high correlation value. A speaker verification method characterized in that a vector sequence including many high vectors is determined to be a similar voice, and the determination is used for speaker verification determination. 4. In the speaker verification method according to any one of claims 1 to 3, when forming a vector sequence of a voice to be verified, the input voice is shifted by one bit.
A speaker verification method comprising: generating a vector sequence corresponding thereto; and sequentially comparing the vector sequence shifted one bit at a time with a dictionary vector sequence to determine speaker verification. 5. A band division unit for dividing a sampled input sound into a band, a non-linear compression unit for non-linearly compressing a sound signal of a predetermined band among the band-divided sound signals, and a vector from the non-linearly compressed sound signal. A vector-sequencing unit that forms a column, a dictionary creating unit that creates a dictionary vector sequence as a dictionary, and performs pattern matching between a vector sequence of a voice to be matched and a dictionary vector sequence created by the dictionary creating unit 4. A speaker verification system comprising: a speaker determination unit that performs speaker verification and determination. 6. The speaker verification system according to claim 5, wherein the dictionary creation unit calculates a correlation value for each vector between vector sequences corresponding to individual identical voices uttered by the same speaker, A process of calculating the total number of vectors whose values are equal to or greater than a predetermined threshold value is performed for all the vector columns, and the vector sequence having the maximum sum of the number of correlation values equal to or greater than the predetermined threshold value is determined as the dictionary vector of the speech. A speaker collation system, wherein a dictionary is registered with a dictionary vector sequence based on the voice of each speaker created by the dictionary creating unit. 7. The speaker verification system according to claim 5, wherein the non-linear compression section groups a suitable number of sampled values of the sampled input speech into one group, and includes an absolute value of the sample value included in the group. A speaker verification system characterized in that after normalizing all the sample values in a group with the maximum value of, a normalized sample value is converted into a low-resolution audio signal by a non-linear quantization function. 8. The speaker verification system according to claim 5, wherein the speaker determination unit includes a vector of the vector sequence of the voice to be verified and a dictionary vector sequence corresponding to each of a plurality of speakers registered in the dictionary. The correlation value is calculated for each of the vectors, the total number of vectors whose correlation value is equal to or greater than a predetermined threshold is calculated, and the speaker corresponding to the dictionary vector sequence in which the total number of vectors whose correlation value is equal to or more than a predetermined threshold is the maximum, A speaker verification system characterized in that the speaker is determined as a speaker. 9. In the speaker verification system according to claim 5, when forming a vector sequence of the voice to be verified, the input voice is shifted one bit at a time, a vector sequence corresponding to the input voice is created, and shifted one bit at a time. A speaker verification system, wherein the speaker verification is performed by sequentially verifying a vector sequence and a dictionary vector sequence.