CN1932974A - Speaker identifying equipment, speaker identifying program and speaker identifying method - Google Patents

Abstract

In the speaker distance calculation section of the speaker recognition apparatus, a quantized distance between a speech feature vector of a speech feature vector sequence generated from a speech of a speaker to be recognized and a representative vector in a codebook is obtained. The speech feature vector is quantized based on the quantization distance. And the quantization distortion is obtained by using a higher order speech feature vector group in the speech feature vector sequence. In the recognition section of the speaker recognition device, speaker recognition is performed based on quantization distortion, which is, for example, an average value of a plurality of quantization distortions.

Description

Speaker recognition device, speaker recognition program, and speaker recognition method

technical field

The present invention relates to a speaker recognition device, a computer program for speaker recognition, and a speaker recognition method for recognizing a speaker by using personal information included in sound waves.

Background technique

A text-dependent speaker recognition device that recognizes a speaker based on a voice speaking a predetermined content, and a text-independent speaker recognition device that identifies a speaker based on a voice speaking any content have been proposed as speaker recognition devices.

A speaker recognition device generally converts an input sound wave into an analog signal, converts the converted analog signal into a digital signal, performs dispersion analysis of the digital signal, and then generates a speech feature vector sequence including personal information. Here, cepstrum coefficients are used as speech feature vectors. In the registration mode, the speaker recognition device clusters the sequence of speech feature vectors into a predetermined number of clusters, for example, thirty-two clusters, and generates a representative vector that is the centroid of each cluster (see Furui et al. "Speech Information Processing", Morikita Shuppan Co.Ltd, Japan, first edition, pp. 56-57). In addition, in the identification mode, the speaker recognition device calculates the distance between the sequence of speech feature vectors generated from the input sound waves in the registration mode and the pre-registered codebook based on each speech feature vector, and calculates an average value (mean distance) , and identify speakers based on this average distance.

In the case where the speaker recognition device is used as the speaker verification device, the distance between the speech feature vector sequence generated from the speaker to be recognized and the codebook about the speaker is calculated, and the distance and the threshold Compare to perform speaker verification. In the case where the speaker recognition device is used as the speaker identification device, the distance between the speech feature vector sequence generated from the speaker to be identified and the codebooks of all registered speakers is calculated, and from the codebook corresponding to The shortest distance is selected among the plurality of distances of the registered speakers to perform speaker identification.

Currently, cepstral coefficients reflecting the shape of the vocal tract, or pitch indicating the vibration frequency of the vocal cords are generally used as speech feature quantities. Its information includes phonological information indicating the content of the voice, and personal information depending on the speaker. When calculating the difference of the speaker's voice as a distance, since the dispersion of the phonological information is larger than that of the personal information, it is not desirable to compare the dispersion of the phonological information with that of the personal information. Instead, it is desirable to compare the same phonological information. Therefore, according to the existing speaker recognition apparatus, approximate normalization of phonemes is performed by clustering vector dispersion in the observation space, and the speaker distance reflecting personality obtained by comparing approximately identical phonemes is calculated as Amount of distortion.

However, when clustering a sequence of speech feature vectors, there is a problem to which order the speech feature vectors should be set. Usually, a large amount of phonological information exists in the low order, while a large amount of personal information exists in the high order. Therefore, if the speech feature vector order is set to a low order in order to improve the phonetic discrimination performance when clustering, the speaker discrimination performance may be degraded. Conversely, if the speech feature vector is set to a high order in order to improve the speaker discrimination performance, the phoneme discrimination performance may be degraded. This leads to a problem of trade-off relationships. Because of this issue, the speech feature vector order is currently set to the most appropriate order determined by experimental methods.

Contents of the invention

Therefore, it is an object of the present invention to eliminate the trade-off relationship between phonetic discrimination performance and speaker discrimination performance, and to achieve accurate speaker recognition.

According to an aspect of the present invention, there is provided a speaker recognition apparatus, wherein a first speech feature is obtained based on a group of low-order speech feature vectors in a first speech feature vector sequence generated from a speech of a speaker to be registered The distance between the phonetic feature vectors of the sequence of vectors. The first speech feature vector sequence is clustered based on the obtained distance, and a codebook including a plurality of representative vectors is generated and stored. Each speech feature vector in (a) the second speech feature vector sequence and (b) A quantization distance between corresponding ones of the plurality of representative vectors stored in the codebook. Each speech feature vector in the second sequence of speech feature vectors is quantized based on the obtained quantization distance. And based on the high-order speech feature vector group of the second feature vector sequence, the quantization distortion between each speech feature vector of the second speech feature vector sequence and a corresponding one of the plurality of representative vectors stored in the codebook is obtained . Speaker identification is performed based on the obtained quantization distortion.

According to another aspect of the present invention, there is provided a speaker recognition apparatus in which a weighting based on a first weight is obtained between speech feature vectors of a first speech feature vector sequence generated from a speech of a speaker to be registered Vector distance. The first speech feature vector sequence is clustered based on the obtained weighted vector distances, and a codebook including a plurality of representative vectors is generated and stored. obtaining a weighted quantization based on a second weight between a corresponding one of the plurality of representative vectors stored in the codebook and each speech feature vector of a second sequence of speech feature vectors generated from the speaker's speech to be recognized distance. Each speech feature vector in the second sequence of speech feature vectors is quantized based on the obtained weighted quantization distance. and obtaining a weighted quantization based on a third weight different from the first weight and the second weight between a corresponding one of the plurality of representative vectors stored in the codebook and each speech feature vector of the second speech feature vector sequence distortion. Speaker identification is performed based on the quantization distortion.

According to the present invention, highly accurate speaker recognition can be realized.

Description of drawings

1 is a block diagram showing the structure of a speaker recognition device of the present invention;

Fig. 2 schematically shows a pattern diagram of clustering in order to obtain representative vectors from the speech feature vector sequence;

FIG. 3 is a block diagram showing the structure of a speaker recognition section provided in the speaker recognition device;

FIG. 4 is a schematic diagram showing the structure of a feature vector; and

FIG. 5 is a block diagram showing an example structure of the speaker recognition device of the present invention realized by software.

Detailed ways

A first embodiment of the present invention will be described with reference to FIGS. 1 to 4 . FIG. 1 is a block diagram showing the structure of a speaker recognition device 100 of the first embodiment.

As shown in FIG. 1, the speaker recognition device 100 includes a microphone 1, a low-pass filter 2, an A/D converter 3, a feature vector generation part 4, a speaker recognition part 5, a speaker model generation part 6, and a memory Part 7. With these parts and elements, various means (or steps) can be performed.

The microphone 1 converts the input speech into an electrical analog signal. The low-pass filter 2 removes frequencies higher than a predetermined frequency from the input analog signal. The A/D converter 3 converts an input analog signal into a digital signal at a predetermined sampling frequency and quantization bit number. Voice input section 8 includes microphone 1 , low-pass filter 2 , and A/D converter 3 .

The feature vector generation section 4 performs discrete analysis of the input digital signal and generates and outputs an M-order speech feature vector sequence (feature parameter time series). Furthermore, the feature vector generation section 4 includes a switch (not shown) for selecting a registration mode and an identification mode. According to the selected mode, in the registration mode, the feature vector generation part 4 is electrically connected to the speaker model generation part 6, and outputs the M-order speech feature vector sequence to the speaker generation part 6, while in the identification mode, the feature vector generation The section 4 is electrically connected to the speaker identification section 5, and outputs the M-order speech feature vector sequence to the speaker identification section 5. In this embodiment of the present invention, the M-order speech feature vector sequence is a 16-order speech feature vector sequence (M=16), and the feature vector includes 1 to 16-order LPC cepstral coefficients, but is not limited to this example.

The speaker model generating section 6 generates a codebook as a speaker model from the speech feature vector sequence generated at the feature vector generating section 4 in the registration mode. The storage section 7 is a dictionary that stores (registers) the codebook generated at the speaker model generation section 6 .

The speaker identification section 5 calculates the distance between the codebook pre-stored in the storage section 7 and the speech feature vector sequence generated at the feature vector generation section 4, then identifies the speaker based on the distance, and outputs the result as the speaker recognition result.

Next, the speaker model generating section 6 will be described with reference to FIG. 2 , which is a pattern diagram schematically showing clustering performed to obtain representative vectors (centroids) from a sequence of speech feature vectors.

As shown in FIG. 2 , the speaker model generation section 6 clusters the M-order speech feature vector sequence generated from the speaker's voice to be registered at the feature vector generation section 4 in the registration mode into a plurality of corresponding predetermined codebooks clusters of size. The speaker model generating section 6 then obtains, for each cluster, the centroid as the weighting center of the cluster as a representative vector of the cluster, and registers a plurality of representative vectors (centroids of each cluster) into the storage as codebook elements. Part 7 (Dictionary). A codebook is generated for each enrolled speaker.

Here, clustering is performed by using an N-order (N<M) speech feature vector sequence (shaded area in FIG. 2 ) out of an M-order speech feature vector sequence, and an M-order representative vector is obtained. The sequence of N-order speech feature vectors is a group of low-order speech feature vectors.

The vector distance D1 between vectors used in clustering can be obtained from the following formula (1). In this embodiment of the invention, N=8, M=16, and the codebook size is 32.

$\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>\underset{\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ]</span>}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

where D1: vector distance

X _K , Y _K : M order eigenvectors

N<M

That is, the speaker model generating section 6 uses the N-order speech feature vector sequence among the M-order speech feature vector sequences generated at the feature vector generating section 4 in the registration mode, obtains the vector distance D1 according to formula (1), and then The M-order speech feature vector sequence is clustered based on the obtained vector distance D1, and a codebook composed of a plurality of M-order representative vectors is generated.

Next, the speaker recognition section 5 will be explained with reference to FIG. 3, which is a block diagram showing the structure of the speaker recognition section 5. FIG.

As shown in FIG. 3 , the speaker recognition section 5 includes a speaker distance calculation section 11 and a recognition section 12 .

The speaker distance calculation section 11 calculates the distance between the plurality of representative vectors stored in the codebook and the M-order speech feature vector sequence generated at the feature vector generation section 4 from the speaker's voice to be recognized (in the codebook The distance between this and the sequence of feature vectors). That is, the speaker distance calculating section 11 calculates the feature vector from the feature vector producing section 4 and the code in the storing section 7 for each feature vector in the sequence of M-order speech feature vectors produced at the feature vector producing section 4. The distance between the representative vectors of this (the distance between the representative vector and the feature vector).

Here, the distance between the codebook and the feature vector sequence can be obtained as follows: (a) based on the quantization distance D2 between the representative vector and the feature vector calculated by using the N-order elements, quantize the speech feature vector sequence Each of the M-order speech feature vectors; and (b) obtaining the distortion distance D3 (quantization distortion) between the representative vector and the feature vector by using the M-order speech feature vector. Therefore, the distance between the codebook and the sequence of feature vectors is calculated as the average value of the obtained quantization distortions. Here, the sequence of N-order speech feature vectors is a group of low-order speech feature vectors, and the sequence of M-order speech feature vectors is a group of high-order speech feature vectors.

The quantization distance D2 between the representative vector and the feature vector used in the quantization process can be obtained from the following formula (2), and the distortion distance D3 can be obtained from the following formula (3).

$\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>\underset{\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ]</span>}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Among them, D2: represents the vector-feature vector distance (quantization distance)

C _K : represents the vector

X _K : M order eigenvector

$\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>\underset{\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ]</span>}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

Among them, D3: represents the vector-eigenvector distance (distortion distance)

C _K : represents the vector

X _K : M order eigenvector

Speaker distance calculation section 11 obtains quantized distance D2 according to formula (2). D2 is the quantization distance between representative vectors and feature vectors, that is, each speech feature vector in the sequence of M-order speech feature vectors generated at feature vector generating section 4 and stored in storage section 7 in the registration mode Multiple representations in the codebook represent quantized distances between vectors. Then, based on the obtained quantization distance D2, quantization of the M-order speech feature vector sequence is performed by using the N-order speech feature vector sequence. That is, each speech feature vector in the sequence of M-order speech feature vectors is quantized. Then, the distortion distance D3 between the representative vector and the feature vector is calculated according to formula (3). D3 is the distortion distance between the plurality of representative vectors stored in the codebook of the storage section 7 and the sequence of M-order speech feature vectors generated at the feature vector generation section 4 .

In this embodiment, quantization distortion is obtained by using an M-order speech feature vector sequence, but it is not limited to this example. For example, quantization distortion can be obtained by using a speech feature vector sequence including a (m to M) order (N<m<M) speech feature vector sequence (higher-order speech feature vector sequence). A speech feature vector sequence including (m to M) order (N<m<M) speech feature vector sequences should be a high-order speech feature vector group. If the set of high-order speech feature vectors comprises a sequence of high-order speech feature vectors, it should be high enough. (m to M) order (N＜m＜M) speech feature vector sequence can be following any one: only comprise the speech feature vector of (m to M) order cepstral coefficient shown in Fig. 4 (b) shaded area Sequence; Include (m to M) order cepstral coefficient shown in Fig. 4 (c) shaded area and (1 to N) the voice feature vector sequence of a part in order N) order cepstral coefficient; Or comprise Fig. 4 (d) shadow The speech feature vector sequence (M-order speech feature vector sequence) of (1 to M) order cepstral coefficients shown in the region. Here, the (1 to N) order cepstral coefficients (shaded area in Fig. 4(a)) are low-order cepstral coefficients, and the (m to M) order cepstral coefficients are high-order cepstral coefficients. High-order cepstral coefficients contain more personal information than low-order cepstral coefficients, and low-order cepstral coefficients contain more phonological information than high-order cepstral coefficients. In this embodiment, N=8 and M=16, but not limited to these values.

The identification section 12 identifies the speaker based on the average value of the quantization distortion obtained at the speaker distance calculation section 11, and outputs the identification result as a speaker identification result. When the speaker recognition device 100 is used as a speaker verification device, the speaker distance calculation section 11 calculates a sequence of speech feature vectors generated from the speech of the speaker to be recognized and stored in the codebook of the speaker to be recognized The distance (average value of quantization distortion) between multiple representative vectors of . The recognition section 12 recognizes the speaker by comparing the distance with a threshold. In addition, when the speaker identification device 100 is used as a speaker identification device, the speaker distance calculation section 11 calculates the sequence of speech feature vectors generated from the speech of the speaker to be recognized and stored in the codebooks of all registered speakers. The multiples in represent the distances between vectors, and then the speaker is identified by selecting the shortest distance from the multiple distances.

According to the first embodiment of the present invention, in the registration mode, it is possible to obtain each speech feature Vector-to-vector distance of vector D1. The sequence of M-order speech feature vectors is clustered based on the vector distance D1, and a codebook composed of multiple M-order centroids is generated. Furthermore, in the identification mode, quantization is performed based on the quantization distance D2 between each M-order speech feature vector generated from the speaker's voice to be recognized and the N-order vector element in each representative vector of the codebook. For each speech feature vector in the M-order speech feature vector sequence, the distortion distance D3 using the M-order vector elements is obtained, and speaker recognition is performed based on the average value of the quantization distortion. With the above structure, the trade-off relationship between the phonetic discrimination performance and the speaker discrimination performance can be eliminated, and their good balance can be ensured. Therefore, highly accurate speaker recognition can be realized.

In this embodiment of the present invention, the first sequence of speech feature vectors generated from the speaker's speech to be registered and the second sequence of speech feature vectors generated from the speaker's speech to be recognized are M-order speech features The vector sequence, the low-order speech feature vector group is an N-order (N<M) speech feature vector sequence, the codebook includes M-order representative vectors, and the high-order speech feature vector group is an M-order speech feature vector sequence. Therefore, stable recognition performance can be assuredly ensured.

Alternatively, according to this embodiment of the present invention, both the first speech feature vector sequence and the second speech feature vector sequence are M-order speech feature vector sequences, and the low-order speech feature vector group is N-order (N<M) speech features The vector sequence, the codebook includes M-order representative vectors, and the high-order speech feature vector group is a speech feature vector sequence including (m to M) order (N<m<M) speech feature vector sequences. Therefore, stable recognition performance can be assuredly ensured.

A second embodiment of the present invention will now be described. The second embodiment is a modification of the speaker recognition section 5 and speaker model generation section 6 of the first embodiment of the present invention. Therefore, parts of the same structure appearing in the second embodiment will be denoted by the same reference numerals as those in the first embodiment, and their explanations will be omitted except for the speaker recognition section 5 and the speaker model generation section 6 .

The speaker model generating section 6 according to the second embodiment will be explained with reference to FIG. 2 . The speaker model generation section 6 clusters the M-order speech feature vector sequence generated from the speaker's voice to be registered at the feature vector generation section 4 in the registration mode into a plurality of clusters corresponding to a predetermined codebook size, obtaining as The centroid of the center of each cluster is weighted so as to make the centroid a representative vector of the cluster, and a plurality of representative vectors are registered to the storage section (dictionary) 7 as a codebook. A codebook is generated for each enrolled speaker.

Here, clustering is performed by using an M-order speech feature vector sequence, and an M-order representative vector is obtained. The weighted vector distance D1 between vectors used in clustering can be obtained from the following formula (4). In this embodiment, N=8, M=16, and the codebook size is 32.

$\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>\underset{\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ]</span>}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

where D1: vector distance

U _K : Weight $u_K=\left\{\begin{array}{cc}1& K\&le;N\\ 0& K>N\end{array}\right.$

X _K , Y _K : M order eigenvectors

The speaker model generation section 6 obtains each weighted vector distance D1 according to formula (4) using the M-order speech feature vector sequence generated at the feature vector generation section 4, and clusters the M-order speech based on the obtained weighted vector distance D1 feature vector sequence, and generate a codebook composed of multiple M-order representative vectors.

Next, the speaker recognition section 5 (see FIG. 3) according to the second embodiment will be explained. The speaker recognition section 5 basically has a structure similar to that of the first embodiment, and includes a speaker distance calculation section 11 and a recognition section 12 .

The speaker distance calculation section 11 calculates the distance between the plurality of representative vectors stored in the codebook of the storage section 7 and the M-order speech feature vector sequence generated at the feature vector generation section 4 from the speaker's voice to be recognized (distance between codebook and sequence of feature vectors). That is, the speaker distance calculation section 11 calculates, for each feature vector in the sequence of M-order speech feature vectors generated at the feature vector generation section 4, the distance between the feature vector and the representative vector of the codebook (in the representative vector vector and the distance between the eigenvectors).

Here, by quantizing each M-order speech feature vector in the speech feature vector sequence based on the weighted quantization distance D2, and then obtaining a weighted distortion distance D3 representing the distance between the vector and the feature vector by using the M-order speech feature vector ( quantization distortion), and obtain the distance between the codebook and the feature vector as the average value of the quantization distortion.

According to the second embodiment, the weighted quantization distance D2 of the representative vector and the eigenvector used in quantization can be obtained from the following formula (5), and the weighted distortion distance for obtaining the quantization distortion can be obtained from the following formula (6) D3.

$\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>\underset{\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ]</span>}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

Among them, D2: represents the vector-feature vector distance (quantization distance)

U _K : Weight $u_K=\left\{\begin{array}{cc}1& K\&le;N\\ 0& K>N\end{array}\right.$

U _K : represents the vector

X _K : M order eigenvector

$\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>\underset{\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ]</span>}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

Among them, D3: represents the vector-eigenvector distance (distortion distance)

V _K : Weight (VK=1)

C _K : represents the vector

X _K : M order eigenvector

Therefore, the speaker distance calculation section 11 obtains the weighted quantization distance D2 between the representative vector and the feature vector according to the formula (5), which is a plurality of representative vectors stored in the codebook of the storage section 7 in the recognition mode and in The distance between each speech feature vector in the sequence of M-order speech feature vectors generated at the feature vector generation section 4 . Then, quantization of the sequence of M-order speech feature vectors is performed based on the obtained weighted quantization distance D2. That is to say, quantize each speech feature vector in the sequence of M-order speech feature vectors, obtain a plurality of representative vectors stored in the codebook of the storage part 7 and the M generated at the feature vector generation part 4 according to formula (6). The weighted distortion distance D3 between each speech feature vector in the first-order speech feature vector sequence, and the average value of the obtained weighted distortion distances D3 (average value of quantization distortion) is obtained.

The identification section 12 identifies the speaker based on the average value of the quantization distortion obtained at the speaker distance calculation section 11, and outputs the identification result as a speaker identification result. When the speaker recognition device 100 is used as a speaker verification device, the speaker distance calculation section 11 calculates a sequence of speech feature vectors generated from the speech of the speaker to be recognized and stored in the codebook of the speaker to be recognized A plurality of representative vectors of , and the recognition section 12 verifies the speaker by comparing the distance with a threshold. In addition, when the speaker recognition device 100 is used as a speaker identification device, the speaker distance calculation section 11 calculates a speech feature vector generated from a speaker's speech to be recognized and stored in codebooks of all registered speakers. The distance between multiple representative vectors of , (the average value of the quantization distortion), and the speaker is identified by selecting the shortest distance from the obtained distances.

According to the second embodiment of the present invention described above, in the registration mode, the weighted vector-to-vector distance D1 of each vector of the M-order speech feature vector sequence generated from the speaker's speech to be registered is obtained, based on the obtained The weighted vector distance D1 clusters the M-order speech feature vector sequence, and generates a codebook including a plurality of M-order representative vectors. In the identification mode, each is quantized based on the weighted quantization distance D2 between each speech feature vector in the sequence of M-order speech feature vectors generated from the speaker's speech to be recognized and each representative vector of the codebook A speech feature vector, quantization distortion is obtained by using an M-order speech feature vector sequence based on the distortion distance D3, and then speaker recognition is performed based on an average value of the quantization distortion. With the above structure, the trade-off relationship between the phonetic discrimination performance and the speaker discrimination performance can be eliminated, and their good balance can be ensured. Therefore, highly accurate speaker recognition can be achieved.

In the second embodiment of the present invention, formulas (4), (5) and (6) are used, but not limited to these formulas. For example, formula (4) can be replaced by the following formula (7) (weight: U _K =1), formula (5) can be replaced by the following formula (8) (weight: U _K =1), and formula (6 ) can be replaced by the following formula (9) (weight: V _K =1/S _K ). Here, the standard deviation S _K which is the dispersion value of each speech feature vector is obtained statistically in advance.

$\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>\underset{\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ]</span>}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$

where D1: vector distance

U _K : weight (U _K =1)

X _K , Y _K : M order eigenvectors

$\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>\underset{\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ]</span>}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$

Among them, D2: represents the vector-feature vector distance (quantization distance)

U _K : weight (U _K =1)

C _K : represents the vector

X _K : M order eigenvector

$\mathrm{<span\; class="notranslate">\; D.</span>}\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>\underset{\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; K</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ]</span>}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 9</span>}<span\; class="notranslate">\; )</span>$

Among them, D3: represents the vector-eigenvector distance (distortion distance)

V _K : Weight (V _K =1/S _K )

C _K : represents the vector

X _K : M order eigenvector

S _K : K order standard deviation

In the second embodiment of the present invention, the first sequence of speech feature vectors generated from the speech of the speaker to be registered and the second sequence of speech feature vectors generated from the speech of the speaker to be recognized are both of order M A sequence of speech feature vectors. The weighted vector distance and the weighted quantized distance are respectively obtained by using weights satisfying the following relation,

U _K =1(k≦N), 0(k>N), where N<M.

Here U _K is the first weight and the second weight. The weighted distortion distance can be obtained by using weights satisfying the following relation,

V _K =1(k≤M), wherein the third weight is V _K .

Therefore, highly accurate recognition performance can be achieved.

Alternatively, in the second embodiment of the present invention, both the first speech feature vector sequence and the second speech feature vector sequence are M-order speech feature vector sequences. The weighted vector distance and the weighted quantized distance are respectively obtained by using weights satisfying the following relation,

U _K ＝1(k≤M)

Here U _K is the first weight and the second weight. The weighted distortion distance can be obtained by using weights satisfying the following relation,

V _K =1/S _K (k≤M), wherein the third weight is V _K .

Therefore, highly accurate recognition performance can be realized.

The hardware structure is not limited to the above-mentioned specific structure, and it can be realized by software. The speaker recognition section 5 or the speaker model generation section 6 can be realized by software. FIG. 5 is a block diagram showing the speaker recognition device 100 implemented by software.

As shown in FIG. 5, the speaker recognition device 100 includes: a CPU 101 connected via a bus to a ROM storing a BIOS and the like; and a memory 102 including a ROM and a RAM to constitute a microcomputer. The CPU 101 is connected via I/O (not shown) to an HDD 103, a CD-ROM drive 105 that reads a computer-readable CD-ROM 104, a communication device 106 that communicates with the Internet, etc., a keyboard 107, such as a CRT or LCD, via a bus like display 108, and microphone 1.

The CD-ROM 104 as a computer-readable storage medium stores a program for realizing the speaker recognition function of the present invention, and the CPU 101 can realize the speaker recognition function of the present invention by installing the program. Also, the voice input through the microphone 1 is stored in the HDD 103 or the like. Then, when the program is running, voice data stored in the HDD 103 or the like is read to perform speaker recognition processing. The speaker recognition processing realizes a function similar to that of each of the feature vector generation section 4, speaker recognition section 5, speaker generation section 6, etc., and thus effects similar to those described above can be obtained.

As the storage medium, various optical disks such as DVD, various optical magnetic disks, various magnetic disks such as floppy disks, semiconductor memories, and the like can be used. Furthermore, the present invention can be realized by downloading a program from a network such as the Internet and installing the program into the HDD 103 as a storage section. In this case, the storage device storing the program at the server on the sending side becomes the storage medium of the present invention. The program may run on a given OS (Operating System), and in that case, the program may allow the OS to perform some part of the above-mentioned processing, and the program may include prescribed applications such as word processor software A part of a program file group including software or an OS.

Claims (8)

1, a kind of speaker identification equipment (100) is characterized in that comprising:

Be used for obtaining based on the low order speech feature vector group first mentioned speech feature vector sequence that produces from speaker's voice that will be registered between the speech feature vector of first mentioned speech feature vector sequence distance, first mentioned speech feature vector sequence of trooping, and produce the device of the code book that comprises a plurality of representation vectors based on the distance that is obtained;

Be used to store the device of the code book that is produced;

Be used for obtaining each speech feature vector in (a) second mentioned speech feature vector sequence and (b) be stored in quantized distance between of correspondence in a plurality of representation vectors in the code book based on the low order speech feature vector group second mentioned speech feature vector sequence that produces from speaker's voice that will be identified vector, quantize each the described speech feature vector in second mentioned speech feature vector sequence based on the quantized distance that is obtained, and obtain each the described speech feature vector in second mentioned speech feature vector sequence based on the high-order speech feature vector group in second characteristic vector sequence and be stored in the device of the quantizing distortion between one corresponding in a plurality of representation vectors in the code book; And

Be used for carrying out the device of speaker identification based on the quantizing distortion that is obtained.

2, speaker identification equipment as claimed in claim 1, wherein, in first mentioned speech feature vector sequence and second mentioned speech feature vector sequence each all is M rank mentioned speech feature vector sequence, low order speech feature vector group is the N rank (mentioned speech feature vector sequence of N＜M), corresponding vector is M rank representation vectors, and high-order speech feature vector group is M rank mentioned speech feature vector sequence.

3, speaker identification equipment as claimed in claim 1, wherein, in first mentioned speech feature vector sequence and second mentioned speech feature vector sequence each all is M rank mentioned speech feature vector sequence, the low order mentioned speech feature vector sequence is the N rank (mentioned speech feature vector sequence of N＜M), representation vector is M rank representation vectors, and high-order speech feature vector group is to comprise that m is to the M rank mentioned speech feature vector sequence (mentioned speech feature vector sequence of N＜m＜M).

4, a kind of speaker identification equipment (100) is characterized in that comprising:

Be used for obtaining between the speech feature vector of first mentioned speech feature vector sequence that produces from speaker's voice that will be registered the weighing vector distance based on first weight, first mentioned speech feature vector sequence of trooping, and produce the device of the code book that comprises a plurality of representation vectors based on the weighing vector distance that is obtained;

Be used to store the device of the code book that is produced;

Weight quantization distance between each speech feature vector in second mentioned speech feature vector sequence that is used for obtaining corresponding in a plurality of representation vectors of code book storage one and from speaker's voice that will be identified, produces based on second weight, quantize each the described speech feature vector in second mentioned speech feature vector sequence based on the weight quantization that obtained distance, and in a plurality of representation vectors that obtain in code book, to store corresponding one with second mentioned speech feature vector sequence in each described speech feature vector between the device based on the weight quantization distortion of three weight different with second weight with first weight; And

Be used for carrying out the device of speaker identification based on this quantizing distortion.

5, speaker identification equipment as claimed in claim 4, wherein, each in first mentioned speech feature vector sequence and second mentioned speech feature vector sequence all is M rank mentioned speech feature vector sequence;

Wherein second weight of first weight of weighing vector distance and weight quantization distance all is U _K, wherein:

U _K＝1(k≤N)，0(k＞N)，

And N＜M; And

Wherein the 3rd weight of weighted distortion distance is V _K, wherein:

V _K＝1(k≤M)。

6, speaker identification equipment as claimed in claim 4, wherein, each in first mentioned speech feature vector sequence and second mentioned speech feature vector sequence all is M rank mentioned speech feature vector sequence;

Wherein second weight of first weight of weighing vector distance and weight quantization distance all is U _K, wherein:

U _K=1 (k≤M); And

Wherein the 3rd weight of weighted distortion distance is V _K, wherein:

V _K＝1/S _K(k≤M)，

And the deviation value on each M rank is S _K

7, a kind of speaker identification method is characterized in that comprising:

Be used for obtaining the vector distance between the speech feature vector in first mentioned speech feature vector sequence, first mentioned speech feature vector sequence of trooping based on the vector distance that is obtained and produce the step of the code book that comprises a plurality of representation vectors based on the low order speech feature vector group first mentioned speech feature vector sequence that produces from speaker's voice that will be registered;

Be used to store the step of the code book that is produced;

Be used for obtaining each speech feature vector in (a) second mentioned speech feature vector sequence and (b) be stored in quantized distance between one corresponding in a plurality of representation vectors in the code book based on the low order speech feature vector group second mentioned speech feature vector sequence that produces from speaker's voice that will be identified, quantize each the described speech feature vector in second mentioned speech feature vector sequence based on the quantized distance that is obtained, and obtain each the described speech feature vector in second mentioned speech feature vector sequence based on the high-order speech feature vector group in second mentioned speech feature vector sequence and be stored in the step of the quantizing distortion between one corresponding in a plurality of representation vectors in the code book; And

Be used for carrying out the step of speaker identification based on the quantizing distortion that is obtained.

8, a kind of speaker identification method is characterized in that comprising:

Be used for obtaining the weighing vector distance between the speech feature vector of first mentioned speech feature vector sequence that produces from speaker's voice that will be registered, first mentioned speech feature vector sequence of trooping based on the weighing vector distance that is obtained and produce the step of the code book that comprises a plurality of representation vectors based on first weight;

Be used to store the step of the code book that is produced;

Weight quantization distance between each speech feature vector in second mentioned speech feature vector sequence that is used for obtaining corresponding in a plurality of representation vectors that code book is stored one and from speaker's voice that will be identified, produces based on second weight, quantize each the described speech feature vector in second mentioned speech feature vector sequence based on the weight quantization that obtained distance, and in a plurality of representation vectors that obtain in code book, to store corresponding one with second mentioned speech feature vector sequence in each described speech feature vector between the step based on the weight quantization distortion of three weight different with second weight with first weight; And

Be used for carrying out the step of speaker identification based on the quantizing distortion that is obtained.

Images