JPH0981196A - Vector quantizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the distortion of synthesized voice in voice coding. SOLUTION: On the decoding side of the vector quantizing device, a distribution coding book 109 which stores the distribution of the vectors clustered to the reproduced vector is installed in addition to a coding book 108 which stores reproduced vectors. From the correlation between this distribution and the decoded vectors determined by a maximum likelihood detector 106 in the previous frame, the likelihood as the decoded vector at each point of the vector space having the vector reproduced by a decoder 104 as the center, is calculated by a likelihood calculator 105, and the point where the likelihood is judged to be maximum by a maximum likelihood detector 106 is regarded as the decoded vector.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vector quantizer used in voice information compression required in digital telephones and digital recorders.

[0002]

2. Description of the Related Art In recent years, since the medium for transmitting or accumulating data is limited, it is desired to realize a voice information compression technique which enables a high compression rate and a high voice quality. Especially,
Vector quantization, which quantizes a vector consisting of a plurality of samples as a unit, is used as an element technology of high-efficiency compression, and realization of a vector quantizer with less distortion is desired.

A conventional vector quantizer will be described below with reference to the block diagram shown in FIG. In FIG. 6, reference numeral 401 is a voice analyzer that analyzes voice and extracts analysis parameters. An encoder 402 selects a vector closest to the extracted parameter group (hereinafter, referred to as a vector) from the codebook 403, and the code number thereof is set to the transmission line 4
07. A codebook 403 stores vectors obtained by analyzing learning sample voices in advance. The above is the configuration on the encoding side. Next, the configuration on the decoding side is shown. A decoder 404 obtains a reproduction vector from the code transmitted from the transmission line 407 using a codebook 406. A voice synthesizer 405 performs voice synthesis using the decoded reproduction vector. 406 is a codebook on the decoding side,
A reproduction vector corresponding to the codebook on the encoding side is stored. Reference numeral 407 is a transmission path connecting the device on the encoding side and the device on the decoding side.

The operation of the vector quantizer having the above configuration will be described below. First,
The voice analyzer 401 analyzes the voice and obtains a vector according to the analysis model. This vector is the encoder 402
Is vector quantized by. In vector quantization,
In advance, using a codebook 403 composed of a code vector obtained by statistically processing the learning vector obtained by analyzing the learning sample speech, the vector analyzed at the time of encoding is quantized to the closest code vector, The code number is transmitted to the transmission line 407. The above is the operation on the encoding side.
On the decoding side, the decoder 404 selects the code vector of the code number obtained via the transmission path 407 from the code book 406 as the reproduction vector. Like the codebook 403, the codebook 406 is composed of reproduction vectors obtained in advance using learning sample voices. The speech synthesizer 405 synthesizes speech using the reproduction vector thus obtained.

[0005]

However, since the conventional vector quantizer as described above does not consider the error between the vector obtained by the analysis and the reproduction vector, quantization distortion occurs. It had the problem of being lost.

The present invention solves the above-mentioned conventional problem. As a codebook created using a learning vector obtained by analyzing a learning sample speech, not only the reproduction vector but also the reproduction vector is clustered. An object of the present invention is to provide a vector quantum device capable of reducing vector quantization distortion by having a distribution of learning vectors.

[0007]

In order to achieve the above object, the first structure of the present invention is, in a quantizing device for speech coding, analyzes speech on the coding side to extract parameters. Equipped with a speech analyzer, an encoder that vector-quantizes the analyzed parameters, and a codebook that stores the code vector for vector quantization.The decoding side corresponds to the coding-side codebook. A codebook that stores the reproduced vector, a decoder that obtains the reproduced vector from the code obtained from the transmission line using the codebook, and a distributed codebook that stores the distribution of the vectors clustered in the reproduced vector. , A likelihood calculator that calculates a likelihood function in the parameter space from the correlation between the reproduction vector and the occurrence probability distribution of the vector obtained using the distribution codebook and the decoded vector obtained in the previous frame It is obtained by a speech synthesizer for synthesizing speech by using a maximum likelihood detector for determining a vector that maximizes the calculated likelihood as a decoded vector, the resulting decoded vector.

A second structure of the present invention is, in addition to the above-mentioned first structure, an error detector for detecting a code error of a transmission line on the decoding side, and a likelihood calculator according to the detected error. And a weight switching device for weighting the calculated likelihood.

A third configuration of the present invention is, in addition to the above first configuration, obtained on the decoding side from an error detector for detecting a code error in a transmission line and a distributed codebook according to the detected error. And a distribution adjuster that adjusts the occurrence probability distribution of the generated vector.

[0010]

According to the first aspect of the present invention, when the decoded vector is obtained, the occurrence probability distribution of the decoded vector with respect to the reproduction vector and the continuity from the previous frame are taken into consideration, thereby reducing the quantization distortion in the auditory sense. can do.
Further, according to the second configuration, when the likelihood is calculated according to the transmission path error, the weighting for the likelihood of the correlation with the reproduction vector or the previous frame is switched, so that the synthesized speech at the time of the code error on the transmission path is obtained. It is possible to reduce deterioration of quality. Further, according to the third configuration of the present invention, by changing the distribution of the reproduction vector according to the transmission path error, it is possible to avoid synthesizing the voice with an incorrect reproduction vector, and it is possible to avoid the code error in the transmission path. It is possible to reduce the deterioration of the synthesized voice quality.

[0011]

【Example】

(Embodiment 1) Hereinafter, an embodiment of the vector quantization apparatus of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram of a vector quantizer according to the first embodiment of the present invention. In FIG. 1, reference numeral 101 is a voice analyzer that analyzes voice and extracts parameters. 10
Reference numeral 2 is an encoder which vector-quantizes the analyzed parameters. A codebook 103 stores a code vector for vector quantization. The above is the configuration on the encoding side. Next, the configuration on the decoding side will be described. A decoder 104 obtains a reproduction vector from the code obtained from the transmission line 110. A likelihood calculator 105 calculates likelihoods for all points on the vector space from the occurrence probability distribution of the obtained reproduction vector and decoded vector and the decoded vector obtained in the previous frame. Reference numeral 106 denotes a maximum likelihood detector, which obtains a vector that maximizes the calculated likelihood as a decoded vector. Reference numeral 107 denotes a voice synthesizer, which performs voice synthesis using the decoded vector thus obtained. A codebook 108 stores reproduction vectors corresponding to the codebook 103 on the encoding side. 109 is a distributed codebook, which is the codebook 10.
The distribution of learning vectors clustered into eight reproduction vectors is stored. Reference numeral 110 is a transmission line connecting the device on the encoding side and the device on the decoding side.

FIG. 2 shows an example of these codebooks in a two-dimensional vector space. L _i (i = 1, ..., 4)
Assuming that the vectors are normally distributed independently in each axis when the four learning vectors of are clustered in the same cluster, the variance is 1/2 {1 ² + (-
1) ² } = 1, the variance with respect to the y-axis is 1/2 {2 ² + (-
2) ² } = 4. Therefore, the average (0,0) is stored in the codebook 108 as the reproduction vector, and (1,4) is stored in the distribution codebook 109 as the corresponding distribution. By using such distribution codebook 109, since x and y are independent, the occurrence probability distribution p (x, y) of the decoding vector with respect to the reproduction vector (0, 0) is obtained by the following equation.

[Equation 1] Instead of such modeling, the distribution of learning vectors may be accumulated as a table at the time of creating the codebook and used as the occurrence probability distribution.

The operation of the vector quantizer having the above configuration will be described below. First, the vector analyzed by the speech analyzer 101 is the codebook 103
Is encoded by the encoder 102. The decoder 104 which has obtained this code through the transmission line 110 generates a reproduction vector using the codebook 108. The likelihood calculator 105 acquires the occurrence probability distribution corresponding to the reproduction vector using the distribution codebook 109, and from this and the decoded vector of the previous frame, using the following equation, for the point v on the vector space,
The likelihood l (v) is calculated by the following formula.

[Equation 2] From the likelihood calculated in this way, the maximum likelihood detector 1
The vector having the maximum likelihood at 06 is decoded as a decoding vector. The speech synthesizer 107 synthesizes speech using the decoding vector obtained by the above decoding process.

As described above, according to this embodiment, it is possible to reduce the audible quantization distortion in consideration of the connection between frames, and it is possible to realize efficient vector quantization.

(Embodiment 2) FIG. 3 is a block diagram of a vector quantizer according to a second embodiment of the present invention. In FIG. 3, 201 is a voice analyzer for analyzing voice and extracting parameters. Reference numeral 202 denotes an encoder that vector-quantizes the analyzed parameters. A codebook 203 stores a code vector for vector quantization. The above is the configuration on the encoding side. Next, the configuration on the decoding side will be described. A decoder 204 obtains a reproduction vector from the code obtained from the transmission line 212. 205
Is a likelihood calculator, and the obtained reproduction vector and distribution codebook 2
09, the likelihood is calculated for the point on the vector space from the occurrence probability distribution obtained from No. 09 and the decoded vector obtained in the previous frame, considering the likelihood weight by the weight switching unit 211. A maximum likelihood detector 206 obtains a vector that maximizes the calculated likelihood as a decoded vector. 207
Is a speech synthesizer, which synthesizes the decoded speech by using the decoded vector thus obtained. Reference numeral 208 denotes a codebook, which stores reproduction vectors corresponding to the codebook 203 on the encoding side. A distribution codebook 209 stores the distribution of learning vectors to be clustered with respect to the reproduction vector of the codebook 208. Reference numeral 210 denotes an error detector, which is a transmission line 21.
The code error at 2 is detected. A weight switching unit 211 weights the likelihood calculated by the likelihood calculator 205 according to the detected error. Reference numeral 212 is a transmission line connecting the device on the encoding side and the device on the decoding side.

The operation of the vector quantizer having the above configuration will be described below. First, Example 1
Similarly to, the parameters obtained by the voice analyzer 201 are encoded by the encoder 202 using the codebook 203. The decoder 204 receiving this code through the transmission path 212 obtains a reproduction vector using the codebook 208. The likelihood calculator 205 calculates the likelihood using the equation (2) from the occurrence probability distribution of the reproduction vector and the decoded vector and the decoded vector decoded in the previous frame. At this time, the parameter α of the equation (2) is determined by the weight switching unit 211. The maximum likelihood detector 206 obtains a vector having the maximum calculated likelihood as a decoded vector,
The speech synthesizer 207 synthesizes speech using this decoded vector. On the other hand, when the error detector 210 detects a code error in the transmission line 212, the weight switching unit 211 brings α close to 1, weights the correlation with the decoded vector of the previous frame in likelihood calculation, The influence of the reproduced reproduction vector on the likelihood is reduced.

As described above, according to this embodiment, when a code error occurs in the transmission path, it is possible to perform maximum likelihood decoding in consideration of the error, and it is possible to reduce the deterioration of the synthesized voice quality due to the error. .

(Embodiment 3) FIG. 4 is a block diagram of a vector quantizer according to a third embodiment of the present invention. In FIG. 4, reference numeral 301 is a voice analyzer that analyzes voice and extracts parameters. An encoder 302 vector-quantizes the analyzed parameters. A codebook 303 stores code vectors for vector quantization. The above is the configuration on the encoding side. Next, the configuration on the decoding side will be described. A decoder 304 obtains a reproduction vector from the code obtained from the transmission line 312. 305
Is a likelihood calculator, and the obtained reproduction vector and distribution adjuster 3
The likelihood is calculated for a point on the vector space from the occurrence probability distribution obtained in 11 and the decoded vector obtained in the previous frame. A maximum likelihood detector 306 obtains a vector that maximizes the calculated likelihood as a decoded vector.
A speech synthesizer 307 synthesizes decoded speech using the decoded vector thus obtained. A codebook 308 stores a reproduction vector corresponding to the codebook 303 on the coding side. A distribution codebook 309 stores the distribution of learning vectors clustered with the reproduction vector of the codebook 308. Reference numeral 310 denotes an error detector, which is a transmission line 3
A code error at 12 is detected. 311 is a distribution adjuster,
The occurrence probability distribution obtained from the distribution codebook 309 when an error is detected is changed. Reference numeral 312 is a transmission line connecting the device on the encoding side and the device on the decoding side.

The operation of the vector quantizer having the above configuration will be described below. First, Example 1
Similarly to, the parameters obtained by the speech analyzer 301 are encoded by the encoder 302 using the codebook 303. The decoder 304 that has received this code through the transmission line 312 obtains a reproduction vector using the codebook 308. The likelihood calculator 305 calculates the likelihood from the reproduction vector, the occurrence probability distribution obtained from the distribution adjuster 311, and the decoded vector decoded in the previous frame using the equation (2). The maximum likelihood detector 306 obtains a vector having the maximum calculated likelihood as a decoded vector, and the speech synthesizer 307 synthesizes speech using this decoded vector.
On the other hand, when the error detector 310 detects a code error in the transmission line 312, the distribution adjuster 311 causes the distribution codebook 30
Adjust the distribution obtained in 9.

Here, an example of adjusting the occurrence probability distribution in the distribution adjuster 311 will be described with reference to FIG. However, FIG. 5 shows an example on a one-dimensional vector space. FIG.
In, the occurrence probability when there is no transmission path error is 0 at the bisector of the distance from the adjacent vector in the codebook.
This indicates that the vector is quantized at a distance closer to the bisector, so that the decoded vector exists at a point closer to the reproduction vector than the bisector. On the other hand, when an error is detected, there is a possibility that a decoding vector exists regardless of the bisector of the adjacent vector. give.

As described above, according to this embodiment, the second embodiment
Similarly, when a code error occurs in the transmission path, it is possible to perform maximum likelihood decoding in consideration of the error, and it is possible to reduce deterioration of synthesized speech quality due to the error.

[0022]

As described above, according to the present invention, at the time of decoding a vector-quantized code, the probability that the decoded vector is distributed to each code vector, and the previous frame The distortion can be reduced by considering the correlation from Also, by adjusting the likelihood calculated when a code error occurs in the transmission path, it is possible to reduce the deterioration of the synthesized voice quality.

[Brief description of drawings]

FIG. 1 is a block diagram of a vector quantization device according to a first embodiment.

FIG. 2 is a schematic diagram of a decoding-side codebook configuration method in each embodiment.

FIG. 3 is a block diagram of a vector quantization device according to a second embodiment.

FIG. 4 is a block diagram of a vector quantization device according to a third embodiment.

FIG. 5 is a schematic diagram of a probability distribution adjustment method according to the third embodiment.

FIG. 6 is a block diagram of a vector quantization device in a conventional example.

[Explanation of symbols]

 101, 201, 301 Speech analyzer 102, 202, 302 Encoder 103, 203, 303 Codebook 104, 204, 304 Decoder 105, 205, 305 Likelihood calculator 106, 206, 306 Maximum likelihood detector 107, 207, 307 Speech synthesizer 108, 208, 308 Codebook 109, 209, 309 Distribution codebook 110, 212, 312 Transmission line 210, 310 Error detector 211 Weight switcher 311 Distribution adjuster

─────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] October 16, 1995

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] Fig. 6

[Correction method] Change

[Correction contents]

FIG. 6

Claims (3)

[Claims] 1. A vector quantizer in speech coding, wherein a speech analyzer for analyzing speech to extract parameters, a coder for vector quantizing the analyzed parameters, and a vector quantizer on a coding side. A codebook that stores a code vector for performing the decoding is provided, and on the decoding side, the codebook that stores the reproduction vector corresponding to the coding side codebook and the codebook obtained from the code obtained from the transmission path. , A distribution codebook that stores the distribution of the vectors clustered into the reproduction vector, the occurrence probability distribution of the vector obtained using the reproduction vector and the distribution codebook, and the previous frame From the correlation with the obtained decoding vector, the likelihood calculator that calculates the likelihood function in the parameter space and the vector that maximizes the calculated likelihood as the decoding vector. A vector quantizer comprising a maximum likelihood detector to be obtained and a speech synthesizer for synthesizing speech using the obtained decoded vector. 2. The decoding side is provided with an error detector that detects a code error in a transmission path, and a weight switcher that weights the likelihood calculated by the likelihood calculator according to the detected error. The vector quantizing device according to claim 1. 3. The decoding side is provided with an error detector that detects a code error in a transmission line, and a distribution adjuster that adjusts a probability distribution of occurrence of a vector obtained from a distributed codebook according to the detected error. The vector quantizing device according to claim 1.