JP2899024B2 - Vector quantization method - Google Patents

Description

The present invention is used in a speech analysis / synthesis system and speech waveform coding, and has a small number of LPC spectrum parameters representing a spectrum envelope characteristic of a time-series signal such as a speech signal. The present invention relates to a vector quantization method for encoding efficiently with an information amount.

[Background Art] Encoding a speech spectral parameter with a minimum amount of information is a basic technique in a speech analysis / synthesis system and speech waveform encoding. A vector quantization method has been known as an effective method for low-bit coding of speech spectral parameters. In the vector quantization method, a plurality of parameters are grouped into one and regarded as a vector to determine the distance from the standard pattern vector, and a standard pattern having the smallest distance is selected from among the plurality of standard patterns to perform quantization. Output. In this case, a basic measure of vector quantization is a distance scale between the input vector and the standard pattern vector and a search method for finding a standard pattern that minimizes the distance. For the distance measure and search method used in the conventional vector quantization, see RMGray's “Ve
ctor Quatization ", IEEE ASSP Magazine Vol.1No.2 198
4) or in the literature (Furui, "Digital Speech Processing", Tokai University Press, 1985).

As a distance measure used in vector quantization of spectral parameters, a mean square error (Euclidean distance) between LPC parameter vectors, a weighted mean square error, a log likelihood ratio, and the like have been used. The quality of these distance measures is usually measured by the spectral distortion of the coded speech and ultimately determined by the subjective quality of the coded speech. The spectrum distortion is calculated as an average distortion obtained by calculating a square error between a logarithmic spectrum of an input speech and a coded speech for each frequency component, and adding this over the entire band. On the other hand, the spectral distortion for each frequency component has a different contribution to the audibility. There is a mel-cepstral scale as a spectrum distortion scale utilizing such auditory characteristics. The mel-cepstral scale is defined as the spectral distortion obtained by calculating the spectral distortion at each frequency at non-uniform frequency intervals and adding this over the entire band. Specifically, the mel-cepstral coefficients are obtained by Fourier series expansion of the log-amplitude spectrum with respect to the logarithmic frequency axis, and the mel-cepstral coefficients are minimized so that the mean square error between the mel-cepstral coefficients of the input speech and the coded speech is minimized. Is quantized. A specific example of applying the mel-cepstral scale to a speech analysis / synthesis system is described in the literature, Kitamura and Imai, "Speech analysis-synthesis system using mel-cepstrum and quality of synthesized speech", IEICE, J68-A, 9,198.
It is described in 5. In this document, a method of scalar quantizing individual mel-cepstral coefficients is used. However, it can be easily inferred that the method is extended to a method of vector quantizing these coefficients collectively. In this method, the logarithmic scale of the frequency axis is used to accurately reproduce the amplitude spectrum components in the lower frequency range than the high frequency range, and the coding is performed at the same bit rate as compared with the method using a uniform frequency scale. The subjective quality of the reproduced speech can be improved. However, in the method using the mel-cepstrum, the weighting of the spectrum distortion along the frequency axis is fixed independently of the spectrum characteristics of the input voice, so that the signal such as a voice signal whose spectral characteristics change with time is fixed. On the other hand, the audible nature of the spectral distortion is not sufficiently reflected. As a result, there is a problem that the bit rate required for coding the spectrum parameter cannot be sufficiently reduced. Further, when vector quantizing the mel-cepstral coefficient, it has been difficult to create a standard pattern of the mel-cepstral coefficient that guarantees the stability of the synthesis filter.

An object of the present invention is to encode a spectral envelope parameter of a time-series signal such as a voice or an acoustic signal, particularly an LPC parameter at a low bit rate, and use an LPC parameter using a spectral distance measure that utilizes an audible property regarding spectral distortion. Is to provide a method of vector quantizing.

According to the present invention, a spectrum distance weighted according to the intensity of the spectrum envelope of the input signal as a spectrum distance between the LCP parameter of the input signal and the standard pattern,
A standard pattern with such a minimum spectral distance is selected, and the standard pattern is used as a quantized output.

The spectral distance weighted according to the intensity of the spectral envelope of the input signal is obtained as a weighted cepstrum distance as follows. Assuming that the power spectra of the two all-pole types are f (λ) and (λ), the weighted log spectrum distortion is defined by the following equation.

The weighting characteristic W (λ) is determined from the prediction coefficient α _K of the input speech such that the weight characteristic W (λ) takes a large value for a frequency component having a strong amplitude intensity of the spectrum f (λ). The calculation method will be described in an embodiment. The power spectrum f, is subjected to Fourier series expansion as follows.

Here, c _k and _k are cepstrum coefficients, and c ₀ =
₀ = 0, _ck = c- _k , _k = _-k . The cepstrum coefficient for the all-pole spectrum is obtained from the prediction coefficient using the following recurrence formula.

However, in the above equation, a _k = 0 (k> p). _k
Is similarly obtained from _k using the above equation. At this time, the weighted logarithmic spectral distortion is rewritten as follows using the cepstrum coefficient. Hereinafter, this distortion measure is referred to as a weighted cepstrum distance.

Where γ is the cepstrum truncation order, c _k
^w is a weighted cepstrum coefficient, which is given by the following equation by a convolution operation of the cepstrum coefficient and the weighting coefficient.

A vector having cepstrum coefficients as elements is represented by c = (c
_−γ ,…, c ₀ ,…, c _γ ) ^t , = (− _γ ,…, ₀ ,…,
_{Assuming that γ} ) ^t (t represents an inverted value), the weighted cepstrum distance is approximately expressed by the following equation.

d ² = (c−) ^t V (c−) (2) where V is a 2γ + 1 × 2γ + 1 weight matrix,
The matrix element v _ij is given by the following equation as a correlation function of the weight coefficient w _k .

Here, w _k = 0 (k <0, k> q).

"Example" Next, an example of the present invention will be described. FIG. 1 shows a configuration example of vector quantization of LPC parameters according to the present invention. A sample sequence of a sampled audio signal is input from an input terminal 1. The LPC analysis unit 2 performs a linear prediction analysis on the sample value of the audio signal included in the analysis window of 20 to 30 ms for each frame period of 10 to 20 ms, and obtains p prediction coefficients α _i (i = 1, 2 ,…, P). The cepstrum coefficient calculator 3 calculates γ cepstrum coefficients c _k (k = 1,
2, ..., γ). The weight coefficient calculation unit 4 calculates a weight coefficient w _k at the weighted cepstrum distance in the following procedure. The following three methods are conceivable as calculation methods of the weight coefficient.

-Method 1 Linear prediction filter A (γ
The inverse characteristic of z ^-1 ) is approximated by a low-order all-pole filter, and the prediction coefficient at that time is used as a weight coefficient. In particular, And the prediction coefficients are α ′ _k = α _k γ ^k (k = 1, 2,..., P),
α ′ ₀ = 1. Next, the autocorrelation function of the prediction coefficient α ′ _k is And solve the following simultaneous equations to calculate the weighting coefficients.

However, w ₀ = 1.

Method 2 Linear prediction filter A (γ
The weighting factor is obtained by the following equation as an impulse response to z ^-1 ).

Here, w _k = 0 (k <0), δ _k is a Kronecker delta function, and δ ₀ = 1 and δ _k = 0 (k ≠ 0).

Method 3 A weighting factor is obtained by the following equation as an impulse response of the filter A (z ^-1 ) / A (γz ^-1 ).

In each of the methods 1 and 2, the spectral distortion at each frequency is weighted according to the contour of the LPC spectrum of the input voice. The weight constant γ is a constant that changes the bandwidth of the LPC spectrum, and takes a value in the range of 0γ.
When γ = 1, the degree of weighting is the largest, and as it approaches 0, the degree of weighting decreases. When γ = 0, the weighting factor is w ₀
= 1, w _k = 0 (k> 0), and the weighting of the spectrum distortion at each frequency becomes uniform. Usually, a value of 0.4-0.6 is used. On the other hand, in method 3, weighting is applied to the strain distortion near the peak of the LPC spectrum. When γ = 0, the degree of weighting is the largest, and as it approaches 1, the degree of weighting decreases. When γ = 1, the weighting factors are w ₀ = 1 and w _k =
0 (k> 0), and the weighting of the spectrum distortion at each frequency becomes uniform. Usually, a value of 0.4-0.8 is used.

Vector quantization is performed in multiple stages using a plurality of codebooks. In the embodiment of FIG. 1, a three-stage configuration is shown. The codebook 6 includes M ₁ code vectors (standard patterns) θ having LSP parameters as elements.
_i ⁽¹⁾ (i = 1, 2,..., M ₁ ) is stored. In addition, the codebook 9, 12, M _n-number of code vectors ^{_{θ i (n) (i =}} 1,2, ..., M n) to the normal random entries are stored.

In the multi-stage vector quantization procedure, first, each LSP parameter included in the codebook 6 is converted into a cepstrum coefficient _i ⁽¹⁾ by the conversion unit 7. The minimum distance search unit 8 calculates the weighted cepstrum distance between the input vector c having the cepstrum coefficient as an element and the code vector _i ^{(1) according} to the equation (2), and calculates the minimum distance among all the code vectors. Is determined. The optimal code vector number at that time is set to ₁ and the first stage LS
P parameter quantization output Output as In the second stage, the optimal LS in the first stage
After converting the P parameter and each code vector included in the codebook 9, the conversion section 10 converts the cepstrum coefficient _i
⁽²⁾ , and the minimum distance search unit 11 determines the code vector with the minimum weighted cepstrum distance by the same procedure as in the first stage. Optimal code number at this time
₂ and the quantized output of the LSP parameter Output as The quantization procedure at each stage is represented by the following equation.

By performing N-stage multi-stage vector quantization, the prediction coefficient of the input speech is represented by the optimal set of code vector numbers { ₁ , ₂ ,..., _N } in each stage. The quantization output of the LPC parameter is converted to the optimum LSP parameter Is obtained by converting The procedure for obtaining LPC parameters such as prediction coefficients and Percoll coefficients from LSP parameters is described in the above-mentioned document by Furui. Decoding is performed by using a number _n (n = 1, 2,...,
The code vectors are read according to N), and they are added to obtain the LSP parameter. If another LPC parameter such as a Percoll coefficient or a prediction coefficient is to be obtained, it is obtained by converting the LSP parameter.

The code vector of each stage is created by the following procedure. A plurality of LSP parameter samples obtained from the learning speech are clustered by the LBG algorithm to create a first-stage code vector. The details of the algorithm are described in the aforementioned RMGray document. Each sample of the LSP parameter is vector-quantized using this code vector, and a quantization error vector at that time is obtained.
A plurality of random number vectors each having a normal random number having a variance equal to the variance of each element of the error vector are generated, and these vectors are used as a second-stage code vector. next,
This error vector is vector-quantized using the second-stage code vector, and a third-stage code vector is created from the quantized error vector at that time in the same procedure as the second-stage code vector.
By repeating this procedure, code vectors up to the N-th stage are created.

[Effects of the Invention] The vector quantization method of LPC parameters according to the present invention was applied to a speech analysis / synthesis system, and the quantization efficiency of the method of the present invention was examined by evaluating the quality of synthesized speech. Analysis is at sampling frequency 8kHz, analysis window length (Humming window) 30ms, frame period
The spectrum was smoothed using a lag window (cutoff frequency 100 Hz) with 10 ms and analysis order 10. The voice samples used were short sentence voices uttered by 21 male and female speakers, 19 of which were used for standard pattern learning (6600 frames), and the voices of one other male and female were used for evaluation. The weight coefficient w _k is calculated by the method 1, and the time constant γ is set to 0.4. The order of the cepstrum is γ
= 30. After the LPC parameters extracted from the input speech were quantized by the following five methods, a synthesized speech was created using an LSP synthesizer.

・ Method 1 Max-Lloy using Euclidean distance of LSP parameter
d Scalar quantization (SQ) ・ Method 2 Multistage vector quantization (VQ-LSP) ・ Method 3 Multistage vector quantization using cepstrum distance (VQ-LSP)
CD) ・ Method 4 Multi-stage vector quantization using weighted cepstrum distance (VQ-WCD) ・ Method 5 Difference multi-stage vector quantization using weighted cepstrum distance (DVQ-WCD) ・ When LPC parameters are not quantized ( ORG) Among the above-mentioned quantization methods, method 1 is a scalar quantization method that has been often used conventionally. In the method 1, the LSP parameters are quantized with three kinds of quantization bit numbers of 2, 3, 4 bits per parameter (20, 30, 40 bits per frame). Methods 2, 3, and 4 all use multi-stage vector quantization and differ in the spectral distance measure. Also,
Method 5 is a method in which the vector quantization method according to the present invention is applied to a difference vector between a predicted value linearly predicted using a parameter two frames before the LSP parameter and a cepstrum coefficient of the frame. Method 2-
In 5, the number of code vectors is 8 bits per stage (256)
Then, three types of quantization bit numbers of 16 (two stages), 24 (three stages), and 32 bits (four stages) per frame were compared. The evaluation of speech quality was performed using a paired quality evaluation method in which two synthesized speech samples were presented to the subject, and which speech quality was subjectively better determined.

FIG. 2 shows, by preference values, the subjective quality of the analyzed and synthesized sound when the LPC parameters are quantized using each of the quantization methods described above. The numbers in the figure indicate the number of quantization bits per frame. Comparing the number of quantization bits of each quantization method when the voice quality becomes equal, the vector quantization method using the weighted cepstrum distance according to the present invention (method 4) determines the conventional cepstrum distance without weighting. The number of quantization bits per frame can be reduced by about 8 bits compared to the method 3 used and by about 16 bits compared to the vector quantization method using the Euclidean distance of the LSP. Also, in the quantization method (method 5) combined with the difference vector quantization, the quality of the synthesized speech when quantized at 24 bits per frame is equivalent to that when no quantization is performed, and the spectral distortion due to the quantization is reduced. Can be less than the detection limit.

As described above, the vector quantization method of LPC parameters based on the weighted cepstrum distance according to the present invention can obtain much higher voice quality than the conventional method at the same number of quantization bits. Conversely, there is an effect that the number of quantization bits required to obtain the same sound quality as the conventional method can be reliably reduced. Further, the LPC parameter vector quantization method according to the present invention can be applied as a vector quantization method in speech recognition using vector quantization, in addition to speech coding.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an example of a vector quantization method according to the present invention, and FIG. 2 is a diagram comparing the subjective qualities of synthesized speech for various quantization methods.

────────────────────────────────────────────────── ─── Continued on the front page (58) Fields surveyed (Int. Cl. ⁶ , DB name) G10L 3/00-9/18 H03M 7/30 H04B 14/04 JOIS file (JICST)

Claims (3)

(57) [Claims] 1. A process for calculating an LPC parameter representing a spectrum envelope characteristic by performing linear predictive analysis on a time-series input signal such as a voice or an audio signal, and an LPC parameter of the calculated input signal.
For the parameter and its standard pattern, a process of calculating a spectral distance weighted according to the intensity of the spectral envelope of the input signal, and selecting a standard pattern that minimizes the spectral distance from a plurality of standard patterns,
In the LPC parameter vector quantization method having the process of using the standard pattern as the LPC parameter quantization output, the prediction coefficient α ′ _K of the all-pole filter A (γZ ⁻¹ ) is calculated as α ′ _K = α _K γ ^K Α _K is the LPC parameter of the input signal, (K = 1, 2,
··· P) The autocorrelation coefficient φ _K of 0 <γ <1 α ′ _K is obtained from φ _K = Σα ′ _i α ′ _{i + K} , and the weighting coefficient W _K is calculated as の φ _Ki W _K = −φ _K A vector quantization method for LPC parameters, wherein the vector is obtained by solving an equation, and the weighting coefficient W _K is used as a weight when calculating the spectrum distance. 2. A process for calculating an LPC parameter representing a spectrum envelope characteristic by performing linear predictive analysis on a time-series input signal such as a voice or an acoustic signal, and calculating an LPC parameter of the calculated input signal.
For the parameter and its standard pattern, a process of calculating a spectral distance weighted according to the intensity of the spectral envelope of the input signal, and selecting a standard pattern that minimizes the spectral distance from a plurality of standard patterns,
In the vector quantization method of the LPC parameters with a step of the reference pattern and the quantization output of the LPC parameters, the weighting factor W _K as an impulse response for linear prediction filter _{^{A (γZ -1), W K}} = -Σα i γ ^K W _Ki + δ _K α _i is the LPC parameter of the input signal, δ ₀ = 1, δ _K
= 0 (K is other than 0), W _K = 0 (K <0), and the weighting factor W _K is used as a weight when calculating the spectrum distance. Vector quantization of LPC parameters Method. 3. A process of calculating an LPC parameter representing a spectrum envelope characteristic by performing linear predictive analysis on a time-series input signal such as a voice or an acoustic signal, and calculating an LPC parameter of the calculated input signal.
For the parameter and its standard pattern, a process of calculating a spectral distance weighted according to the intensity of the spectral envelope of the input signal, and selecting a standard pattern that minimizes the spectral distance from a plurality of standard patterns,
In a vector quantization method of LPC parameters having a process of using this standard pattern as a quantization output of LPC parameters, a weight coefficient W _K is used as an impulse response of a filter A (Z ⁻¹ ) / A (γZ ⁻¹ ), and W _K = −Σα _i γ ^K W _Ki + Σα _i δ _Ki α _i is the LPC parameter of the input signal, 0 <γ <1, δ ₀ = 1, δ _Ki = 0 (K−i is other than 0), W _K = 0 (K
<0), wherein the weighting factor W _K is used as a weight for calculating the spectrum distance.