JPH01243099A - System and device for speech encoding - Google Patents

Abstract

PURPOSE:To improve a signal/noise and to obtain a reproduced voice of high quality by feeding the quantization error of a coefficient which is quantized in the past back to a coefficient to be quantized when quantizing coefficients successively, and improving the relation of quantization error-to-signal electric power relation. CONSTITUTION:A voice signal x(n) is inputted from an input terminal 10 and supplied to a coefficient extraction part 31. A coefficient corrector 60 corrects the amplitude of a coefficient which is not quantized yet according to the mutual correlation function phi(i, j) between the quantization error supplied from a quantization error calculator 70 and a base signal train supplied from a coefficient extractor 31 and a coefficient (gi). A quantizer 41 quantizes and supplies coefficients supplied from the coefficient corrector 60 to the quantization error calculator 70 through an output terminal 51, and the quantization error calculator 70 calculates the quantization error of the coefficients and supplies it to the coefficient corrector 60. Consequently, a reproduction signal error is reduced greatly and the reproduction signal of high quality is obtained.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a speech encoding method and an apparatus thereof, and relates to a speech encoding system that expresses a special trap speech Igo code by a linear combination of a plurality of base signal sequences, such as multipulses. Concerning the conversion method and its equipment.

[Prior art]

The audio signal x (n) in five sections is expressed as a base signal sequence (h 1 (n
) A nine-voice encoding method that approximates the linear sum of ) is widely available, and is known to be particularly effective at low bit rates.

Now, if the error signal is expressed as e (n), then in this encoding, x
(n) can be expressed as in the following equation (1).

0≦n≦N-1 (1) where gl (+=L..., K) is a coefficient, N is the number of samples in the short interval, and K is the number of signal sequences. Still (g
l) is determined as minimizing the mean square error of e (n) shown by the following equation (2).

The thread count (g+) thus obtained is quantized for encoding. FIG. 2 shows the principle of the conventional example.

The input terminal 10 takes in the audio signal x (n) and supplies it to the coefficient extractor 30 . The coefficient extractor 3o inputs (h, (n)) from the base signal string memory 20 and determines the coefficient (g+) so as to minimize J in equation (2).

The obtained (g,) is quantized by the successor generator 40 and then outputted from the output terminal 50 to the outside.

[Problem to be solved by the invention]

In the above-mentioned conventional audio encoding method, the error Q between the encoded audio signal and the reproduced signal is expressed as: (Ql) is the quantized coefficient of each coefficient, and (e,) is the quantized coefficient of each coefficient. It can be expressed as the following equation (3).

Therefore, in the configuration in which the coefficients are determined and then quantized as in the conventional method, the coefficients determined by optimizing the equation (2) are the optimal ones that minimize the equation (2) even after quantization. There is no holding pressure that it becomes a coefficient, and the n-th equation (3) is expressed in a quadratic form formed by the cross-correlation function between the basis functions that are selected to be mutually independent and correlated as described above. Since it is always positive, the quantization error of the coefficients definitely increases the reproduction error, and there is a drawback that it may significantly increase the error depending on the size K of the cross-correlation function between the base 1z numbers. .

An object of the present invention is to reduce reproduced signal errors by correcting the coefficients by an amount determined by the quantization error of the coefficients and the cross-correlation function between the base signals when quantizing the coefficients of base signals that are correlated with each other. It is an object of the present invention to provide a speech encoding method and an apparatus for the same.

[Failure to solve the problem]

The audio encoding method of the present invention is an audio encoding method in which an audio signal divided into a series of time intervals is expressed as a composite signal consisting of a linear sum of multiple base signal sequences. When performing 1ρ-order Amakoization, among the coefficients that have not yet been quantized, the quantization error of the previously quantized coefficients and the base signal sequence having the previously quantized coefficients and the quantized coefficients are added. In audio encoding 41C, which encodes a mutually divided audio signal with a base signal sequence having a coefficient that is not equal to a base signal sequence with a composite signal consisting of a linear sum of a plurality of base signal sequences, means for determining a cross-correlation function between the base signal sequences. a means for quantizing one coefficient of the base signal sequence; a means for determining a quantization error caused by the quantization; and means for correcting the values of coefficients that have not yet been quantized.

[Effect]

The relationship between (g,) that minimizes J in equation (2) is expressed as (h
, (n)). By γ differentiating both sides of equation (2) with respect to gl and setting it to zero, the following equation (4) is obtained.

This shows that the set (g,) that satisfies the relationship (4) f: , makes J in equation (2) the minimum i. Substituting this relationship into equation (2), the reproduced signal error Q is expressed by the following equation (5).

(-) Consider a configuration in which f, (gl) is quantized in order from i = 1. Now, if coefficients up to the kth are quantized, g+=gI+el ・・
・Since it can be written as (6), equation (5) becomes as follows.

Q=σ2-(g1e)'F(g0e)-
(7) Here, g and e are vectors, and F is a matrix.

g=[g11g2. ...1gK] e ”” (e 1 * 82 + ” ” * e
y r Or ” ” ' + O]F, j=φ(i,
D Also, consider the following subvectors and submatrices for each vector and matrix.

g=(ga, gb) e=[ealeb] ga ” G IQ2 +”’+Gk ]+ gb==
CQk+ r, ..., bell]F1-φ(1+J)
'+J"1+..., kj- F2-ψ(1+J) l"L..., kj- +1 to j=. ..., K (F3 = F2 t) F4ij=φ(i+j) +1 to i, j=. ...
, then equation (7) can be written as the following equation (8).

(8) If both sides of the above equation (8) are differentiated with respect to gb and set to zero, a relationship such as the following equation (9) is obtained.

(ga10a)tF2+F4gb=0
...(9) This formula is [g i ] r ””1
+..., and provides guidelines for correcting the remaining coefficients from +1 to K for i=. That is, in order to reduce the reproduction error even after quantization, the remaining coefficients are calculated using the previously quantized coefficients, the quantization error, and the cross-correlation function of each base signal, gb = F, 4-' ( g, 10e,)tF2
・(It is necessary to correct it with 1.

The principle of the invention is shown in FIG. The fF signal X(n) is input from the input terminal 10 and supplied to the coefficient extractor 31. The coefficient extractor 31 extracts the basis function sequence [
hl), calculate its cross-correlation function φ(+, j), find the coefficient (gl) so as to minimize J in equation (2), and send it to the coefficient corrector 60. ,j)
and outputs (gl). The coefficient corrector 60 calculates a cross-correlation function φ(i
, j) and the coefficient (gl), the amplitude of the coefficient that has not yet been quantized is corrected based on equation (10) of iIJ. The quantizer 41 quantizes the coefficients sequentially supplied from the coefficient corrector 60 and outputs them from the 8-input terminal 51 and also supplies them to the quantization error calculator 70 . Quantization error calculator 7
0 is supplied with the quantized coefficients and the output of the coefficient corrector 60, calculates the quantization error of the coefficients, and supplies it to the coefficient corrector 60.

〔Example〕

The invention will now be described in detail with reference to Figure f.

FIG. 1 is a block diagram of one embodiment of the present invention.

Embodiment f11 of FIG.
Mr. B, S., Atal): This is an example of a case where the proposed f2-4 ratio method is applicable.

An explanation of the audio encoding method proposed by Mr. B. Nis Attar can be found in Proceedings of the 1982 IJ3ν C.N.N.S.B. (ICASSP) 61.
4-617M, F7-New Model Op El Kenhi N-Excitety 73 N Fo Ichiban Production Swing Natural Sara/Ding Speech At Low Bit Lane (A new model
of LPCexcitation forprod
ucir+g natural sounding
5peech atlow bit rates)J
(Reference 1).

This encoding method converts audio signals into multiple aa (7
) This method is approximated by a set of impulse responses, and this is equivalent to the method of expressing the driving sound source signal sequence of the audio signal with a plurality of pulses.

Partly, the audio signal X(n) is converted into an impulse response sequence (b(n
-6, )], and the error J between the two is calculated as shown in the following equation (11).

Here, tllgl represents the digit 14 and amplitude of the sixth sound source pulse. In order to obtain a sound source pulse train that reduces equation (11), a method for finding K one by one from i = 1 for sound source pulses will be explained. Now, find the sound source pulses up to -1 at i== (end), i = k@th sound source pulse g
Suppose we want to find k and tk.

The evaluation function for determining gk and tk can be obtained from equation (11) as the following equation (12).

Differentiating equation (12) by lA with respect to gk and setting zero gives the following (1
3) Equation is obtained.

Here, . Equation (14) is the autocorrelation of the impulse response series,
Equation (15) is the cross-correlation between the original speech and the innoculus response sequence. By substituting into equation (12), the following equation (16) is obtained.

Jk=J, -8-gk2...(1
6) From this, tk that minimizes the error Jk is determined as the one that maximizes the magnitude of gk in equation (13). Therefore, the calculation procedure for obtaining the sound source pulse train is
The calculation of the right side of Equation 19 and the detection of its maximum value are repeated from i = 1 to K.

Returning again to FIG. 1, the description of the embodiment will be continued.

The configuration of the embodiment shown in FIG. 1 includes a linear prediction analysis unit 110 that analyzes and extracts linear prediction coefficients, a child izer 120 that quantizes the extracted linear prediction coefficients, and a quantized linear prediction coefficient. An impulse response calculation unit 130 that calculates the impulse response of the synthesis filter to be formed, a cross-correlation function calculation unit 160 that calculates the cross-correlation (-number) between the input voice and the impulse response, a pulse search unit 160 that searches for the sound source pulse train, and a search A pulse amplitude example positive section 170 that corrects the influence of quantization error of the sound source pulse, a pulse quantizer 180 that quantizes the sound source pulse, and a quantization error calculation section 190
In addition to the above, it also includes a multiplexer 200.

Next, the operation of this embodiment will be explained.

In FIG. 1, an input terminal 100 takes in a discretized audio signal x (n) and supplies it to a linear prediction analysis section 110 and a cross-correlation function calculation section 150.

The linear prediction analysis unit 110 uses P known as a linear prediction coefficient.
ARCOR (partial autocorrelation) coefficients are determined, and the determined PARCOR coefficients are output to the prediction coefficient quantizer 120.

The prediction coefficient quantizer 120 quantizes PARCOR coefficients, and there is a well-known da method.
``Optimal Code Configuration in Shape Speech Analysis and Synthesis'', Transactions of the Institute of Electronics and Communication Engineers J61-A, Nw2゜pp, 119-126 (
Nonlinear Amakoization as described in 1982-2) is adopted.

The quantized PARCOR coefficients are output to an impulse calculation section 130 and a multiplexer 140.

Impulse response calculation unit 130! i, quantized PAR
The impulse response h of the all-pole filter formed by the COR coefficients
Calculate (n). This impulse response h(n) is supplied to the cross-correlation function calculation section 150 and the autocorrelation Kan-Etsu Sanzan section 140.

The autocorrelation function calculation unit 140 calculates the autocorrelation function ψ(m,n) given by the above equation (14) so that O<m (N −1
, 0(n(N-1), and the calculated autocorrelation function ψ(m, n) is output to the pulse search section 160 and the pulse amplitude correction section 170.

The cross-correlation function calculation unit 150 calculates the autocorrelation function ψ(n) in the above equation (15) over 0 < n < N −1. The calculated φ(n) is the pulse search unit 1
160.

The pulse search unit 160 uses ψ(m, n) supplied from the autocorrelation function calculation unit 140 and the cross-correlation function calculation unit 15
The sound source pulse trains (gt) and (tl) are calculated by the sequential processing described above using φ(n) supplied from 0.

As mentioned above, the calculation procedure is to calculate (gel and (tt)) by repeatedly calculating the equation (13) and detecting its maximum value from i = 1 to K. The obtained sound source pulse train has a pulse amplitude. The signal is supplied to the correction section 170.

The pulse amplitude correction section 170 performs pulse amplitude correction so as to reduce the influence of the error due to the quantization of the pulse amplitude described in the above (Operation) section.

The correction of the pulse amplitude is determined by ψ (m, n) supplied from the autocorrelation function calculation section 140 and (gl), (ti) supplied from the pulse search section 160.
(tshitj) r' + J"1 +・-・, Nito,
The quantization error (ei) of the pulse amplitude supplied from the quantization error calculation unit 190 is used to correct the pulse amplitude as described in equation (10) above. That is, the elements of the autocorrelation function represented by F and F4 are FH
j=ψ(LH*13) i=1,...2
+1° to kj=,..., K F41j=ψ(Ll, ZJ) i, J=k
+ 1-..., is. In addition to this, the pulse amplitude (gl
) and the quantization error (e, ) into the above equation (10), the pulse amplitude (gl) is corrected by convergence correction.
+1 to i=..., bury K in words.

Of the amplitude-corrected (gl), the +1st glc+1 that has not yet been quantized is output to the pulse quantizer 180 and the quadrupled error calculation ratio section 190.

Further, the pulse position yttk+ corresponding to gk+1 is output to the pulse quantizer 180.

The pulse quantizer 180 quantizes gk+1 and t input from the pulse amplitude correction 170, and the quantized signal gk+4 quantization error calculation unit 19
0 and output to multiplexer 200, dtt, +1
The bt of gk+s is also output to the multiplexer 200.
There is a method of quantization, for example, normalizing by the maximum value of (gk) and then quantizing. Regarding the quantization of tk+I,
(t
By taking a configuration in which tk is supplied in ascending order of
By quantizing the difference between +I and 1, efficient quantization becomes possible.

1 Attraction - difference calculation section 190, pulse table width sleeve normal section 170
Calculate the difference eh++ between gk+1 that can be supplied from the pulse quantizer 180 and g,+, supplied from the pulse quantizer 180, and calculate the pulse'
r5' Output to the W4 correction section 170.

The multi-6 lexer 200 uses the PARCOR coefficients supplied from the prediction coefficient generator 120 and the sound source pulse train (gt) (2) supplied from the pulse Δ encoder 180.
, ) are multiplexed and output from the output terminal 21 to the outside.

〔Effect of the invention〕

As explained above, in an audio encoding method in which an audio signal is expressed as a linear sum of mutually independent and correlated base signal sequences, when coefficients are sequentially quantized, the quantization of previously quantized coefficients is The relationship between quantization error and signal power is improved by feeding back the quantization error to future quantization expansion, which significantly improves the signal-to-noise ratio compared to the conventional configuration that does not feed back the sweetening error. The effect is that high-quality playback audio can be obtained.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an implementation example of the present invention, and FIG. 2 is a block diagram of an example of the implementation of the present invention. , Table 10... is a linear combination of base terms of numbers.
...Input terminal, 20...Base signal sequence, 21
......Base signal sequence, 3o...Coefficient extractor, 31...Coefficient extractor, 40...Quantizer, 41... Quantizer, 50...output terminal, 51...output terminal, 60...-
Coefficient corrector, 70...Quantization error calculator, 10
0...Input terminal, 110...Linear prediction analysis unit, 120...Prediction coefficient quantizer, 130
... Impulse response calculation section, 140 ...
- Autocorrelation function calculation section, 150...- Cross correlation function calculation section, 160... Pulse search section, 170.-
. . . Pulse amplitude correction section, 180 . . . Pulse quantizer, 190 . . . Quantization error calculation section, 20
0...Multiplexer, 210...Output terminal. Agent Patent Attorney Susumu Uchihara 2nd and 3rd v

Claims (2)

[Claims] (1) In an audio encoding method in which an audio signal divided into a series of time intervals is expressed as a composite signal consisting of a linear sum of multiple base signal sequences, when the multiple coefficients of the base signal sequence are sequentially quantized, the A quantization error of a previously quantized coefficient among the coefficients that has not yet been quantized, a base signal sequence having the previously quantized coefficient, and a base signal sequence having the unquantized coefficient. A speech encoding method characterized by applying correction determined by the cross-correlation of (2) In an audio encoding device that encodes an audio signal divided into a series of time intervals with a composite signal consisting of a linear sum of a plurality of base signal sequences, means for determining a cross-correlation function between the base signal sequences; means for quantizing one coefficient of the base signal sequence; means for determining the quantization error; and a coefficient that has not yet been quantized in the base signal sequence from the calculated cross-correlation function and the quantization error. 1. A speech encoding device comprising: means for correcting a value of .