JP2709386B2 - Spectrogram normalization method - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for normalizing a spectrogram, and more particularly, to normalizing a spectrogram between different speakers by using vector quantization to recognize an unspecified speaker. Spectrogram Normalization Method Applicable to Speaker Adaptation and Property Conversion Techniques [Problems to be Solved by Conventional Techniques and Inventions] Automatic translation telephones use speech as input,
The voice is a voice of an unspecified speaker, and it is necessary to accurately recognize the voice of such an unspecified speaker. As one means for unspecified speaker recognition, there is a method of normalizing a spectrogram between different speakers, but the conventional means for normalizing a spectrogram between different speakers mainly includes normalization of a vowel section. And there were only methods such as deterministic changes in spectral frequency. Therefore, a method of normalizing a spectrogram between different speakers using vector quantization is considered. However, in the conventional vector quantization, the spectral distortion measure used for vector quantization has been improved in order to suppress the increase in the amount of calculation and the memory and to improve the recognition performance. Although a composite spectral distortion scale of a combination of various features has been used, in this method, various types of feature amounts are mixed in the spectral distortion scale, and the dependency between them is used as a constraint, thereby improving recognition performance. It was meaningful to map features to a good space. However, this method has the following problems. In order for the dependency between the features to be statistically valid in the codebook of vector quantization, an extremely large number of learning samples and an enormous amount of calculation time for this are required. In terms of the codebook size, the codebook size required for each feature is reduced by using the dependency between features as a constraint. However, the overall codebook size is still a product of the codebook size required for each feature, and becomes very large, requiring an enormous amount of memory. When a codebook for vector quantization is generated using the composite spectrum distortion measure, the reproducibility of the spectrum is reduced due to the correlation between various features. Therefore, a main object of the present invention is to express a spectrum as a finite vector for each individual using vector quantization, and then obtain a correspondence between vectors between different speakers, thereby obtaining a spectrogram between different speakers. It is an object of the present invention to provide a spectrogram normalization method that can be normalized. Means for Solving the Problems The present invention digitizes speech, extracts a spectrogram as a feature of the speech, and normalizes the extracted spectrogram between an unknown speaker and a reference speaker. In the normalization method, a codebook for each feature of speech is generated, the speech of the unknown speaker is vector-quantized with reference to the codebook, and each feature generated by vector quantization using dynamic programming. Create a histogram of the correspondence between the code sequence and the standard pattern sequence for learning of the reference speaker, and constrain the path of the correspondence using the sum of the inter-code distances of various features to the local distance of matching by dynamic programming. By learning the correspondence vector by rewriting the feature vector of the unknown speaker by linear combination of the feature vector of the reference speaker. This is to normalize the program. [Operation] The spectrogram normalization method according to the present invention digitizes speech, generates a codebook for each feature of speech, refers to the codebook, vector-quantizes the speech of an unknown speaker, and performs dynamic programming. A histogram of the correspondence between the code sequence for each feature generated by vector quantization and the standard pattern sequence for learning of the reference speaker is created using the Learning the correspondence by constraining the correspondence path using the sum of the inter-code distances, and rewriting the feature vector of the unknown speaker by linear combination of the feature vector of the reference speaker, the normalization of the spectrogram Is performed. Embodiments of the Invention Hereinafter, embodiments of the present invention will be described in more detail with reference to the drawings. FIG. 1 is a schematic block diagram of a speech recognition apparatus to which the present invention is applied. In FIG. 1, the speech recognition device comprises an amplifier 1, a low-pass filter 2, an A / D converter 3, and a processing device 4. The amplifier 1 amplifies an input audio signal, and the low-pass filter 2 removes aliasing noise from the amplified audio signal. The A / D converter 3 converts an audio signal into a 16-bit digital signal using a 12 kHz sampling signal. The processing device 5 includes a computer 5, a magnetic disk 6, terminals 7 and a printer 8. The computer 5 performs voice recognition based on the digital voice signal input from the A / D converter 3. FIG. 2 is a flowchart showing the entire flow from the input of speech to the output of a normalized spectrogram according to one embodiment of the present invention. Next, the operation of one embodiment of the present invention will be described with reference to FIGS. The input audio signal is amplified by the amplifier 1 and the aliasing noise is removed by the low-pass filter 2, and then, in a step (abbreviated as SP in the drawing) SP1 shown in FIG. 2, the A / D converter 3 is input. The audio signal is converted to a 16-bit digital signal. In step SP2, the computer 5 of the processing device 4 extracts features of the voice converted into the digital signal. This feature extraction is performed using a technique such as linear analysis (LPC analysis). In step SP3, it is determined whether or not a codebook is to be generated. If a codebook is to be generated, in step SP4, a codebook is generated based on the characteristics of the extracted voice. As the codebook generation, for example, the LBG algorithm is used, generated for each feature, and stored in the separate codebook of the magnetic disk 6 in step SP5. Note that LBG
For the algorithm, see Linde, Buzo, Gray; “An algo
rithm for Vector Quantization Design "IEEE COM
−28 (1980-01). When performing quantization, it is determined in step SP3 that a codebook is not generated, and the above-described step
In step SP6, the features of the voice determined in SP2 are
The separate vector quantization is performed with reference to the separate codebook. Then, in step SP7, it is determined whether or not it is learning of the transformation vector. If the transformation vector is learning, in step SP8, the code string for each feature generated by the separate vector quantization is trained by the standard speaker. Standard pattern series and D by Double Split
P (Dynamic Programming: Dynamic Programming) matching. This learning standard pattern sequence is registered in the magnetic disk 6 in advance in step SP9. Step SP10
In, a transformed vector is generated using the histogram of the correspondence between the vectors as a result of the DP matching. This conversion vector is registered on the magnetic disk 6 in step SP11. When it is determined in step SP7 that it is not learning of a transformation vector, that is, when it is determined that normalization is performed, a code sequence for each feature generated by separate vector quantization is determined in step SP11 in step SP12. The permutation is performed frame by frame using the already stored transform vector, and a normalized spectrogram is generated and output. FIG. 3 is a flowchart for explaining the operation of spectrogram normalization using vector quantization, FIG. 4 is a flowchart for explaining the operation of separate vector quantization, and FIG. FIG. 6 is a flowchart for explaining a vector learning algorithm, FIG. 6 is a spectrogram normalization algorithm, and FIG. 7 is a flowchart for explaining a matching method. Next, spectrogram normalization using vector quantization will be described with reference to FIG. The spectrogram normalization using the vector quantization according to the present invention is mainly composed of two functions. One is vector quantization in step SP23. This vector quantization is separate vector quantization in which vector quantization is performed separately for each type of feature. In step SP22, different codebooks are generated for each feature. The second is spectrum conversion (normalization) in step SP24. In step SP24, an unknown speaker utters a learning word to associate a vector. Here, the histogram of the association obtained for all the learning words is obtained, and the feature vector of the codebook of the unknown speaker is represented by a linear combination of the feature vector of the codebook of the standard speaker using the weight as a weight. Stored as a book in step SP25,
At the time of normalization, the input spectrum is replaced using a conversion codebook for each input, and the spectrum is normalized. Here, the separate vector quantization will be described in detail. According to the present invention, speech is divided into two types of characteristics, power and spectral information (autocorrelation coefficient, LPC cepstrum coefficient), and vector quantization is performed separately for each of them. However, since the power is a scalar, non-uniform scalar quantization is performed. In FIG. 4, step SP
In step 31, the speech signal converted into a 16-bit digital signal is subjected to LPC analysis by a 14th-order autocorrelation analysis, and the power and autocorrelation coefficient L
Extract PC cepstrum coefficients. In step SP32, it is determined whether or not power codebook generation is performed,
If a power codebook is to be generated, in step SP33, the power of the input voice is scalar-quantized. In scalar quantization, a non-uniform quantization technique is used to
Generate a power codebook in SP33, step SP
At 34, the generated code book book is stored on the magnetic disk 6. When the power codebook is not generated, that is, at the time of quantization, quantization is performed in step SP35 using the power codebook in step SP34, and a power-related code sequence is output. On the other hand, in step SP36, the LPC correlation coefficient and the LPC
If it is determined that the codebook generation is a cepstrum coefficient codebook, in step SP37, a codebook is generated based on the WLR scale by the LBG algorithm, and in step SP38, the generated codebook is stored on the magnetic disk 6. Is stored. Here, the WLR scale is a measure that emphasizes the characteristics of speech, and exhibits high performance in word speech recognition. “LPC spectrum matching scale with weighted peaks” by Sugiyama and Kano, IEICE Transactions (A) It is described in J64-A5 (198-05). Note that when the codebook generation of the LPC correlation coefficient and the LPC cepstrum coefficient is not performed, that is, at the time of quantization, for the autocorrelation coefficient and the LPC cepstrum coefficient of the input voice, using the spectrum codebook in step SP38,
In step SP39, vector quantization is performed, and a code sequence related to spectrum information is output. Here, the spectrum distortion measures used for codebook generation and quantization are as follows. d _power = P / P '+ P' / P-2 (1) d _spectrum = Σ (C (n) -C '(n)) (R (n) -R' (n)) ... (2) _Where d _power is the distortion measure of the power term, d _spectrum is the spectral distortion measure, R (n) is the autocorrelation coefficient of the nth order in the codebook, and R ′ (n) is the nth order of the input. C (n) is the nth-order LPC cepstrum coefficient of the codebook, C ′ (n) is the nth-order LPC cepstrum coefficient of the input, and P is the power of the codebook. , P 'is the input power. Next, referring to FIG. 5, step SP shown in FIG.
24, the normalization of the spectrum and the generation of the conversion codebook in step SP25 will be described in detail. First,
When generating the conversion codebook, the unknown words are uttered by the unknown speaker. In step SP41, the input speech is subjected to separate vector quantization using the codebook already stored in step SP42. Step SP
In step 43, the quantized code sequence is DP-matched by the Double Split method with the standard pattern for learning the same word of the standard speaker already stored in step SP44, and the unknown speaker and the same Find the correspondence between vectors using learning words. Then, correspondences are obtained for all the learning words, and stored in the form of a histogram.
In step SP45, using the obtained histogram,
The feature vector of the unknown speaker is represented by a weighted sum using the histogram of the association of the feature vector of the codebook of the standard speaker stored in step SP46 as a weight. This load sum can be expressed by the following equation. k: code number of the codebook of the standard speaker n: code number of the codebook of the unknown speaker a ′: conversion vector from unknown speaker to standard speaker b (k): feature vector of the codebook of standard speaker h _n (k): histogram of the code n of the standard speaker with respect to the code n of the unknown speaker obtained by the association by the DP matching Next, in step SP48, the codebook of the unknown speaker is Exchange, step SP43, SP4
5 and SP47 and SP48 are repeated. This repetition is repeated a fixed number of times or until the DP distances for all the learned words converge. When it is determined in step SP47 that the convergence has been achieved, a final conversion vector from the unknown speaker to the standard speaker is obtained. Next, spectrum normalization will be described with reference to FIG. In step SP51, the input speech of the unknown speaker is subjected to separate vector quantization using a codebook. Here, the codebook of the unknown speaker is stored in step SP52.
Are stored in advance. Then, the codebook of the unknown speaker is replaced in step SP53 by the conversion vector from the unknown speaker to the standard speaker obtained in step SP54, the spectrum is replaced in a frame-wise manner, and a normalized spectrogram is output. Next, with reference to FIG. 7, a matching operation for obtaining association will be described. Matching is Double Split
It is performed using the method. In step SP61, the code sequence generated by separately vector-quantizing the power and spectrum by separate vector quantization is matched with the standard pattern stored as the code sequence. In step SP62, a standard pattern of power and spectrum coded by separate vector quantization is stored in advance as the standard pattern. And step SP
In the matching in 61, the distance between the codes is determined in advance by creating a distance matrix in step SP63 and performing this appearance. In this way, the correspondence between the input voice and the vector of the standard pattern obtained by sequentially matching with the standard pattern is output to the histogram generation unit in step SP64. Then, using the histogram obtained by the histogram generation unit as a weight, the feature vector of the unknown speaker is represented by a linear combination of the feature vector of the standard speaker to obtain a conversion vector. Next, the matching method will be described in detail. In the conventional matching, both the input and the standard pattern are one feature sequence or code sequence. However, in the separate vector quantization, generally, the input and the standard pattern are configured by a plurality of code sequences. In the present invention, a description will be given of an example of a matching method of two sequences of a power code sequence and a spectrum code sequence. PW as a distance scale when considering both power and spectrum information
There is an LR scale. This is shown by the following equation. d _PWLR = Σ (C (n) −C ′ (n)) (R (n) −R ′ (n)) + a (P / P ′ + P ′ / P−2) (3) a = 0.01 Conventional In the matching of the code strings by the Double Split method, as described above, the distance between all the representative points is obtained in advance by utilizing the fact that all the spaces are vector-quantized and represented by a finite number of points. It is stored in a matrix. _{Therefore, d PWLR (i, j)} = DL (A (i), B (j)) DL (A (i), B (j)) = Σ (C K (n) -C L (n)) ( _{R K (n) -R L (} n)) + a · (P K / P L + P L / P K -2) a (i) is i-th frame of the code number B of the input speech (j) is the standard The code number DL (K, L) of the j-th frame of the pattern is obtained by expressing the distance between the codes K and L from the distance matrix K and L are the code numbers of A (j) and B (j) However, since separate vector quantization has two sequences, the distance is obtained as follows. d _{[p] [WLR]} (i, j) = DL _spect (A _spect (i), B _spect (j)) + a · DL _power (A _power (i), B _power (j)) where DL _{spect (A spect (i), B} spect (j)) = Σ (C K (n) -C L (n)) (R K (n) -R L (n)) DL power (A power (i), B _power (j)) = P _{K ′} / P _{L ′} + P _{L ′} / P _{K ′} −2 K, L is the code number of A _spect (i), B _spect (j), and K ′, L ′ is A _power (i) and B _power (j). This is obtained by separately coding the first and second terms of the PWLR scale, calculating the distance, and calculating the sum. Using this local distance measure, a distance is determined by DP matching. As described above, it is possible to achieve a very high-performance normalization method using vector quantization. [Effects of the Invention] As described above, according to the present invention, a speech is vector-quantized, a spectrogram is extracted, and a codebook of vector quantization is associated between different speakers. Since the spectrogram is normalized based on the
Fewer learning samples and less computation. It should be noted that the amount of calculation is increased by a certain amount because another vector quantization system is configured by separating, but the amount of calculation can be sufficiently reduced because the number of learning samples is small. In the separate vector quantization, the codebook size is the sum of the codebook sizes required for each feature, so that the entire codebook size can be drastically reduced. Moreover, since the dependence term of each feature is ignored, optimal quantization can be performed within the features of the codebook, and therefore, the spectrogram can be faithfully reproduced.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a schematic block diagram of a speech recognition apparatus to which an embodiment of the present invention is applied. FIG. 2 is a flowchart showing the flow of the entire process from input of voice to normalization. Third
The figure is a flowchart for explaining the operation of spectrogram normalization using vector quantization. FIG. 4 is a flowchart for explaining the operation of the separate vector quantization. FIG. 5 is a flowchart for explaining the algorithm of the transformation vector learning. FIG. 6 is a flowchart showing an algorithm for spectrogram normalization. FIG. 7 is a flowchart for explaining the matching operation. In the figure, 1 is an amplifier, 1 is a low-pass filter, 3 is
An A / D converter 4 is a processing device, and 5 is a computer.

   ────────────────────────────────────────────────── ─── Continuation of front page    (56) References JP-A-61-261799 (JP, A)                 JP-A-61-123889 (JP, A)                 Tokiko 56-51637 (JP, B2)                 International Publication 86/5618 (WO, A1)

Claims (1)

(57) [Claims] A spectrogram normalizing method for digitizing a voice, extracting a spectrogram as a feature of the voice, and normalizing the extracted spectrogram between an unknown speaker and a reference speaker; Is generated, and the speech of the unknown speaker is vector-quantized with reference to the codebook, and the code sequence for each feature generated by vector quantization using the dynamic programming method, the reference speaker, and the standard pattern sequence for learning are generated. A correspondence histogram is created by constraining the correspondence path by using the sum of the inter-code distances of various features to the local distance of the matching by the dynamic programming method, and learning of the correspondence is performed. Spectrogram normalization is now performed by rewriting the feature vector of an unknown speaker with a linear combination of the speaker's feature vectors. , Normalization method of spectrogram. 2. Separate vector quantization is performed using two types of power and autocorrelation coefficient as features of the speech, and a learning histogram is created by learning certain learning words, and a feature of each codebook of the unknown speaker is created. 2. The spectrogram normalization method according to claim 1, wherein the normalization is performed by replacing the vector with a feature vector of a reference speaker using a histogram as a weight by a linear combination.