JPH04293097A - Speaker identification device - Google Patents

Abstract

PURPOSE:To offer the speaker identification device which identifies a speaker corresponding to an inputted voice by vector quantization. CONSTITUTION:This speaker identification device is equipped with code books 12 for plural speakers to be recognized, a vector quantization part 11 which performs the vector quantization of voice parameters of respective frames of the input voice according to the code books 12, a code book selection part 13 which selects a code book minimizing the quantization distortion of the quantized voice parameter among the code books 12, and a decision part 15 which decides that the recognized speaker corresponding to the code book selected most among the code books 12 in a specific time is the speaker of the input voice.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speaker identification device for identifying a speaker based on input voice.

0002

[Prior Art] As a first conventional speaker identification method, a text-independent speaker identification method using vector quantization is known. A speaker identification method that uses the smallest integrated value in the time direction as the recognition result is the one proposed by Rosenberg and Soong.
) et al. (F.K. Soon et al.
Co-author, “Vector quantization method for speaker recognition”, IC
ASSP Proceedings, Institute of Electrical and Electronics Engineers, pp. 387-390
, 1985 (F.K. Soong, et al.
．． , ”A Vector Quantaizatio
n Approach to Speaker Rec
ignition”, Proc. ICASSP,
IEEE, pp. 387-390, 1985)
.

A second conventional text-independent speaker identification method using vector quantization is Rosenberg's
(Rosenberg), Soong et al. (Co-authored by F.K. Soong et al.,
"On the Use of Instantaneous and Transient Spectral Information in Speaker Recognition", Proceedings of ICASSP'86, Institute of Electrical and Electronics Engineers, pp. 877-880, 1986 (F.K. Soo
ng, et al. ,”On the Use of
Instant and Transi
tional Spectral Information
On-in Speaker Recognition
”, Proc. ICASSP´86, TOKYO
, IEEE, pp. 877-880, 1986),
In addition to the audio parameters of each frame, time differentials of the audio parameters are calculated, vector quantized, and the recognition result is the smallest value obtained by integrating the weighted sum of quantization distortion in the time direction.Speaker identification method is known.

[0004] A third conventional text-independent speaker identification method using vector quantization is to vector quantize the audio parameters of each frame and then perform segment quantization to remove quantization distortion in the temporal direction. Sugiyama has researched a speaker identification method in which the recognition result is the smallest integrated value of March 1963, pp. 75-76).

[0005]

[Problems to be Solved by the Invention] However, in the conventional text-independent speaker identification method using the first vector quantization described above, the magnitude of quantization distortion due to vector quantization is In addition to the difference, since it differs depending on the phoneme uttered, there is a problem that when the quantization distortion is integrated in the time direction, the distribution overlaps between speakers, reducing the speaker identification rate.

[0006] Furthermore, in the conventional text-independent speaker identification method using the second and third vector quantization described above, variations in patterns explode when attempting to capture dynamic features of speech. As the data size increases, there is a shortage of audio data to create a codebook.
There is a problem that the speaker identification rate decreases.

In view of the above problems of the conventional text-independent speaker identification method using vector quantization, the present invention provides that the distribution of quantization distortion due to vector quantization does not overlap between speakers, and A speaker identification device having a high speaker identification rate is provided, which is configured to supply sufficient speech data for creating a codebook when capturing dynamic features of speech.

[0008]

[Means for Solving the Problems] The present invention provides codebooks of a plurality of speakers to be recognized, and quantization means for vector quantizing the speech parameters of each frame of input speech based on the plurality of codebooks. , selection means for selecting a codebook from a plurality of codebooks that minimizes quantization distortion of quantized speech parameters, and recognition of the codebook most frequently selected from the plurality of codebooks within a predetermined time. This is achieved by a speaker identification device including a determination means for determining the target speaker as the speaker of the input voice.

[0009]

[Operation] According to the speaker identification device of the present invention, the quantization means vector quantizes the speech parameters of each frame of the input speech based on the codebooks of a plurality of speakers to be recognized, and the selection means vector quantizes the speech parameters of each frame of the input speech. A codebook is selected from multiple codebooks that minimizes the quantization distortion of the voice parameters.
The determining means determines the speaker based on the codebook selected most frequently within a predetermined time from among the plurality of codebooks, and determines the speaker to be recognized in the codebook selected most frequently as the speaker of the input speech. .

The speaker identification device of the present invention may include a modification means for modifying the codebooks of a plurality of speakers to be recognized using learning vector quantization.

[0011]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of a speaker identification device according to the present invention will be described in detail with reference to the drawings.

FIG. 1 shows the configuration of an embodiment of a speaker identification device according to the present invention.

The speaker identification device shown in FIG. 1 includes a speech analysis section 10,
A vector quantization unit 11 connected to the speech analysis unit 10, a codebook 12 connected to the vector quantization unit 11, a codebook selection unit 13 connected to each vector quantization unit 11, and a codebook selection unit 13 and a determining section 15 connected to each frame counter section 14.

The input speech is converted into speech parameters frame by frame by the acoustic analysis section 10. The voice parameter that captures the individuality of voice may be any one of spectrum analysis, vocal tract characteristic analysis, and vocal cord characteristics.

Furthermore, by combining spectrum analysis, vocal tract characteristic analysis, and vocal fold characteristics, even better speech parameters can be constructed.

Next, the operation of speaker identification by the speaker identification device shown in FIG. 1 will be explained.

When identifying a speaker, in order to reduce the influence of overlapping distributions of quantization distortion due to differences in the magnitude of quantization distortion depending on the uttered phoneme, a codebook selection process based on quantization distortion is performed immediately after quantization. 1 codebook for each frame.
2 is selected, and the number of selected frames is counted instead of calculating the sum of quantization distortion.

The extracted audio parameters are quantized by the vector quantizer 11 based on the codebook 12, and the distortion at that time is calculated according to equation (1).

Let the i-th code included in the codebook 12 of speaker S be Cs(i).

Let the input audio parameter time series be X and the vector at frame t be X(t). Letting d(X(t), Cs) be the distortion when the vector X(t) at frame t is quantized using the codebook Cs, it is expressed as follows.

), G(y
).

Next, the calculated distortion is sent to the codebook selection section 1.
3 and select a codebook according to equation (2).

The codebook selection unit 13 receives the quantization distortion d(X(t), Cs) of the input X(t) for each speaker calculated by the vector quantizer, and selects the codebook that minimizes the distortion. For the selected codebook 12 in which speaker s(t) is selected, the frame count unit 14 counts the number of frames according to equation (3).

The frame counting section 14 counts each frame t.
s(t) for each codebook, and count the number of frames N(X,s) in which each codebook was selected.

N(X,s)=N(X,s)+1 (s=s(X
(t))) N(X, s) = N(X, s)
(When s≠s(X(t)))...(3) The number of counted frames is compared in the determination unit 15, and the formula (4)
Output the identification results according to the following.

That is, the determination unit 15 outputs the second half of the predetermined result r(X) of audio input for a predetermined number of frames.

Hereinafter, with reference to FIG. 2, a description will be given of codebook learning using the speaker identification device of FIG. 1.

In FIG. 2, the same reference numerals are given to the parts common to those in FIG. 1. Furthermore, a learning database 16 for input speech, a learning section 17 connected to each codebook 12, and a learning control section 18 connected to the learning database 16 and the learning section 17 are added. has been done.

First, in order to deal with the lack of audio data when creating a codebook, correction is performed by learning vector quantization after the codebook is created.

The initial codebook is created using the k-means method, LBG algorithm, etc., similar to conventional speaker identification methods.

Next, learning vector quantization will be explained.

In general, in learning vector quantization, when performing classification using a codebook, reference vectors in the codebook are repeatedly moved little by little in order to reduce identification errors, and accurate discrimination boundaries are formed using the codebook. do.

First, the extracted audio parameters are quantized by the vector quantizer 11 based on the codebook 12, and the distortion at that time is calculated using equations (5) and (6).

Let Bs be the codebook of each speaker s (s=1, 2, . . . , S), and let Bs(i) be the i-th code included in the codebook Bs. Also, the speaker to be identified r(
r = 1, 2, ..., R) training audio time series as Xr(t)
If (t=1, 2, . . . , Tr), then S=R generally holds.

The distortion when the speech parameters of each frame of each speaker are vector quantized using each codebook Cs is d
(Xr(t), Cs), and the reference vector in the codebook Cs used at that time is Ks(Xr(t)).

The codebook selection unit 13 calculates equations (7) and (8).

If two codebooks with minimum distortion are selected and the speaker of the first selected codebook Cs is s1 and the speaker of the second selected codebook Cs is s2, then the learning control unit 18 specifies the learning data to be inputted from the learning database 16 and at the same time indicates the correct answer to the learning section 17.

At this time, s1(Xr(t))≠r and s2(Xr(
t))=r...(9), then Cs1(
Move Ks1(Xr(t))) away from Xr(t) and move Cs2(Ks2(Xr(t))) to Xr(t)
(10) Perform the above procedure on the learning data of all speakers to be recognized, and end the learning when the learning has converged sufficiently and the movement of the reference vector becomes less than a predetermined amount.

The learning section 17 performs the operation shown in equation (10) under the condition shown in equation (9). Since various modifications other than equations (9) and (10) can be considered for learning vector quantization, there is no need to limit the equations to equations (9) and (10).

In the above embodiment, vector quantization is performed on a static parameter vector for each frame, but the same method can be used for dynamic patterns spanning two or more frames. . In this case, the effect of learning vector quantization becomes even greater.

[0041]

[Effect of the invention] A codebook of multiple recognition target speakers,
A quantization means vector quantizes the audio parameters of each frame of input audio based on multiple codebooks, and a codebook that minimizes quantization distortion of the quantized audio parameters from the multiple codebooks. Since it is equipped with a selection means for selecting, and a determination means for determining the recognition target speaker of the codebook selected most frequently from a plurality of codebooks within a predetermined time as the speaker of the input speech, the uttered phoneme It is possible to reduce the influence of differences in the magnitude of quantization distortion, and as a result, speakers can be identified more accurately.

[Brief explanation of drawings]

FIG. 1 shows the configuration of an embodiment of a speaker identification device of the present invention.

FIG. 2 is an explanatory diagram during codebook learning of the speaker identification device according to the present invention.

[Explanation of symbols]

10 Acoustic analysis department 11 Vector quantization section 12 Codebook 13 Codebook selection section 14 Frame count section 15 Judgment section 16 Learning database 17　Learning Department 18 Learning control section

Claims (2)

[Claims] 1. Codebooks of a plurality of speakers to be recognized, and quantization means for vector quantizing speech parameters of each frame of input speech based on the plurality of codebooks.
selection means for selecting a codebook from the plurality of codebooks that minimizes quantization distortion of the quantized speech parameters; and a codebook selected most often from the plurality of codebooks within a predetermined time. a speaker identification device, comprising determining means for determining a speaker to be recognized as the speaker of the input voice. 2. The speaker identification device according to claim 1, further comprising a modification means for modifying the codebooks of the plurality of speakers to be recognized using learning vector quantization.