JPH01997A - Speech recognition method using vector quantization - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition method using vector quantization, and particularly to a speech recognition method using template matching using vector quantization.

[Prior Art and Problems to be Solved by the Invention] Automatic translation telephones use speech as input, but it is necessary to recognize the input speech. As voice recognition,
Conventionally, methods using vector quantization have been considered. In conventional speech recognition using vector quantization, improvements have been made to the spectral distortion measure used in vector quantization in order to improve recognition performance by suppressing increases in computational complexity and memory. A spectral distortion measure has been proposed. The significance of this method is to mix various types of features in the spectral distortion scale, use the dependencies between them as constraints, and map the features to a space with better recognition performance.

However, this method has two major problems as described below.

■ In order for the dependencies between each feature to have statistical validity in the food book of vector quantization, a large number of learning samples and an enormous amount of calculation time are required.

■ In terms of codebook size, the codebook size required for each feature can be reduced by using the dependencies between features as a constraint. However, the overall codebook size is still the product of the codebook sizes required for each feature, resulting in a very large size and requiring a huge amount of memory.

Therefore, the main purpose of the present invention is to solve the conventional problems by using separate vector quantization, which generates a codebook separately for each feature and performs separate vector quantization. The object of the present invention is to provide a speech recognition method using vector quantization.

[Means for Solving the Problems] The present invention is a speech recognition method using vector quantization that performs recognition by comparing input speech with a standard pattern stored as a vector quantized code string. , has a plurality of codebooks depending on the type of characteristics of input speech, performs vector quantization on each codebook, and performs recognition using the plurality of code strings obtained.

[Operation] In the speech recognition method using vector quantization according to the present invention, by performing vector quantization on a codebook according to the type of feature of input speech, the codebook size can be changed to the codebook size required for each feature. becomes the sum of
The overall codebook size can be reduced.

[Embodiments of the Invention] Hereinafter, embodiments of the invention will be described in more detail with reference to the drawings.

FIG. 1 is a schematic block diagram of a speech recognition device according to the present invention.

In FIG. 1, the speech recognition device includes an amplifier 1, a low-pass filter 2, an A/D converter 3, and a processing device 4. The amplifier 1 is for amplifying an input audio signal, and the low-pass filter 2 is for removing aliasing noise from the amplified audio signal. The A/D converter 3 converts the audio signal into a 16-bit digital signal using a 12 kHz sampling signal. The processing device 4 includes a computer 5, a magnetic disk 6, a terminal 7, and a printer 8. The computer 5 performs voice recognition based on the voice digital signal input from the A/D converter 3.

FIG. 2 is a flowchart showing the overall flow from inputting an audio signal to outputting a recognition result in an embodiment of the present invention, and FIG. 3 is a flowchart for explaining the operation of separate vector quantization. FIG. 4 is a flowchart for explaining the matching operation.

Next, the operation of one embodiment of the present invention will be described with reference to FIGS. 1 to 4. The input audio signal is amplified by the amplifier 1, and after aliasing noise is removed by the low-pass filter 2, the input audio signal is input to the A/D converter 3 in step SPI (abbreviated as SP in the diagram) shown in FIG. Converts the audio signal into a 16-bit digital signal.

In step SP2, the computer 5 of the processing device 4 extracts features of the audio converted into a digital signal. In this feature extraction, for example, linear child n1 analysis (LPG
This is done using methods such as analysis).

Speech recognition is performed in step SP3.

That is, the feature string for each feature generated in step SP2 is matched with the standard pattern already stored in the matching section using the 5-plit method, and the matching distance is sent to the result determining section 5 in step SP5. Note that in the case of specific speaker recognition, the standard pattern in step SP4 is created by the user uttering the recognized word in advance, and in the case of non-specific speaker recognition, it is created by analyzing a database of voices uttered by a large number of speakers. The representative pattern is created as a multi-template and stored on the magnetic disk 6. The result determination unit in step SP5 determines whether the result is appropriate for the recognition candidate and outputs the recognition result.

Next, with reference to FIG. 3, the operations of feature extraction and separate vector quantization shown in FIG. 2 will be described in more detail. In feature extraction, in step 5PII, the audio signal converted to a 16-bit digital signal is subjected to LPG analysis using 14th order autocorrelation analysis, and the power and autocorrelation coefficient, which are the characteristics of the input audio, are extracted.
Extract LPC cepstral coefficients. Step 5P12
In step 5, it is determined whether the power codebook generation is to be performed, and if it is the power codebook generation, step 5 is performed.
In P13, the power of the human voice is scalar quantized. In the scalar quantization, a power codebook is generated in step 5P13 using an inconsequential-m childization method, and the generated power codebook is stored on the magnetic disk 6 in step 5P14.

When not generating a power codebook, that is, at the time of sonization, quantization is performed at step 5P15 using the power codebook at step 5P14,
Outputs a code string related to power.

On the other hand, if it is determined in step 5P16 that the codebook generation is for LPC correlation coefficients and LPC cepstral coefficients, then in step 5P17, LBG
The algorithm generates a codebook based on the WLR measure, and the generated codebook is stored on the magnetic disk 6 in step 5P18.

Here, regarding the LBG algorithm, Linde,
Buzo, Gray; “An alg.

rithm for Vector Quant
ization Design-IEEE C0M
-28 (1980-01).

In addition, the WLR scale is a scale that emphasizes the characteristics of speech and shows high performance in word speech recognition.
Sugiyama and Kano, “LPG spectral matching measure with peak weight”, Institute of Electronics and Communication Engineers paper (A) J64-
A5 (1981-05).

Note that when the codebook of LPG correlation coefficients and LPC cepstrum coefficients is not generated, that is, at the time of quantization, the spectral codebook in step 5P18 is used for the autocorrelation coefficients and LPC cepstrum coefficients of human speech, and step 5P19 is performed. Vector quantization is performed at , and a code string related to spectral information is output.

Here, the spectral distortion measure used for codebook generation and quantization is as follows.

d -P/P'10P' /P-2...(1)o
wer d5. . o, rUll-'i (C(n)-C' (n
)) (R(n)-R' (n))...(2) d: Distortion measure of power term lower dspectrum 'Spectral distortion measure 1? (n)
: n-th autocorrelation coefficient R'(n
): Human-powered nth-order autocorrelation coefficient C(n): Codebook n-th-order LPC cepstrum coefficient C'(n): Human-powered nth-order LPC cepstrum coefficient As mentioned above, when creating a standard pattern, this The code string is stored as a standard pattern, and during recognition, ma-soching with the code string of the standard pattern is performed.

Next, the ma-notching method will be explained with reference to FIG. In step 5P21, matching is performed by the 5prit method based on the code string generated by vector quantizing the power and spectrum separately and the standard pattern stored as the code string. The standard pattern in step 5P22 stores a standard pattern of power and spectrum coded by separate vector quantization. Then, in the matching in step 5P21, Dr (dynamic
For local distances in matching, a time-distance matrix of the feature vector series and food book vectors is created in advance in step 5P24, and matching is performed by performing this table search. In this way, the distance between the input voice and the standard pattern obtained by sequential matching with the standard pattern is output to the result determination section in step 5P25.

Here, the matching method will be explained in more detail.

In conventional matching, both the input and the standard pattern are one feature string or code string, but in separate vector quantization, they are generally composed of a plurality of code strings. This embodiment also uses a two-series matching method of a power code string and a spectral code string, as in the previous embodiment, and an example thereof will be described below.

There is a PWLR measure as a distance measure when both power and spectrum information are considered, but this measure is based on the following (3)
It is shown by the formula.

'PWLI? -Σ (C(n)-C' (n))(
R(n)-1? ' (n)), +a・(P/P'
+P'/P-2) -(3)a-0
,01 In matching code strings using the conventional 5plit method,
Only the standard pattern side is represented by finite points by vector quantization. For the feature vector series of input speech, the distances to all codebooks are calculated in advance and stored in a time-distance matrix. Therefore, dpylR(1゛J) one Σ(CI(n)−CK(j)(n))(J(n)
-RK(j)<n))+1°(PI/PK(j)+PK
(j)/Pi-2) Ri (n), CI (n),
Pi: n-th self-ill relation coefficient, LPC cepstral coefficient, power of i-frame of input audio, 1? K(j)(n"CK(j)(n)PK(j)'
These are the nth-order autocorrelation coefficient, LPC cepstral coefficient, and power of the first code in the standard power code string.

However, since separate vector quantization has two sequences, the distance is determined as follows.

d[Pl[Wl,R](+,,+)'-Σ(C
, (n)-CK(j)(g)) R,(n)-RK(j
, (n)) + 3° (PI/PI, (j) + PL
(j)/Pi-2)Pl, (j): Power of the j-th chord in the standard pattern string.

This is obtained by separately encoding the first and second terms of the PWLR scale, calculating the distance, and finding the sum. Using this local distance measure, the distance is determined by DP matching.

[Effects of the Invention] As described above, according to the present invention, a plurality of codebooks are provided according to the types of characteristics of input speech, vector quantization is performed for each codebook, and a plurality of obtained code strings are Since recognition is performed using , the dependence term of each feature can be ignored, the number of learning samples can be reduced, and the amount of calculation can be reduced. However, by separating, a separate vector quantization system is constructed, so the amount of calculation increases somewhat, but since the number of learning samples is small, the amount of calculation can be sufficiently reduced. Furthermore, in separate vector quantization, the codebook size is the sum of the codebook sizes required for each feature, so the overall codebook size can be drastically reduced.

[Brief explanation of the drawing]

FIG. 1 is a schematic block diagram of a speech recognition device according to the present invention. FIG. 2 is a flow diagram showing the overall flow of speech recognition using the 5-plit method. FIG. 3 is a flow diagram for explaining the operation of separate vector quantization. FIG. 4 is a flow diagram for explaining the matching operation. In the figure, 1 is an amplifier, 2 is a low-pass filter, 3 is an A/D converter, 4 is a processing device, 5 is a computer, 6 is a magnetic disk, 7 is a terminal, and 8 is a printer. Patent applicant A.T.R. Automatic Translation Telephone Research Institute

Claims (1)

[Claims] In a speech recognition method that performs recognition by comparing input speech with a standard pattern stored as a vector quantized code string, a plurality of codes are recognized according to the types of characteristics of the input speech. A speech recognition method using vector quantization, characterized in that the method has a book, performs vector quantization on each codebook, and performs recognition using a plurality of obtained code strings.