JPS63179399A - Voice encoding system - 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 Technical Field The present invention relates to a multi-pulse driven audio encoding system, and more particularly to a method of searching for pulses with excellent encoded audio quality under a limited amount of calculation.

As one method for encoding audio at a bit rate of 8 to 16 kbps, it is possible to encode audio at a bit rate of 8 to 16 kbps.
There is a multi-pulse driven speech encoding method that drives a linear predictive filter with 30 pulses.

In this method, it is necessary to find pulses so that the error between the input signal and the composite signal is minimized, but as a method to simplify the calculation, the absolute value of the cross-correlation between the input signal and the impulse response of the linear prediction filter is maximized. For the second and subsequent pulses, a new cross-correlation is obtained by subtracting the impulse response from the position where the cross-correlation pulse was set (a method has been proposed).

The cross-correlation is based on the cross-correlation between the input signal and the impulse response.In order to remove the influence of the previous frame, the input of the synthesis filter is set to zero, the influence signal is found and subtracted from the input signal, and the noise is masked using auditory characteristics. For auditory weighting, it is common to obtain a signal passed through a filter and use the cross-correlation between this signal and the impulse response.

The pulse position is determined from the position of the sample where the absolute value of the cross-correlation is the maximum, but conventionally, in order to determine the position of the absolute MR, the magnitude of the absolute value was compared for each sample. For more information on this conventional method, see
Study of multi-valse-driven speech coding method J March 23, 1983, Proceedings of the R-Communication Society P, 115-P
, 122. Such conventional methods have the following drawbacks.

Here, if we consider that the original audio is sampled at 8 llz and encoded in a frame of 2 Q m5ec, the number of samples of the original audio in the frame will be 160 samples, and the cross-correlation will also be 160 samples. Furthermore, in pulse search, it is known that the encoded sound quality is better if the pulse search is performed not only within the encoded frame but also outside the frame, than when the pulse search is performed only within the frame. Therefore, the cross-correlation used for pulse search is about 200 samples.

Absolute Ia is determined by the conventional method of comparing the size of each sample.
When searching for the position of R size, 199 comparisons are required for the cross-correlation of 200 samples.

Also, this position search for the maximum absolute value is required once for each pulse, so 30
When calculating the pulse, the comparison process is 199x 30-
The processing time is 5,970 times/201 seconds, which is a drawback that is a considerable obstacle to processing in real time.

Purpose of the Invention The present invention has been made to eliminate the drawbacks of the above-mentioned conventional methods, and its purpose is to provide a speech encoding method that can significantly reduce the number of times of magnitude comparison processing in maximum value search. It's about doing.

Structure of the Invention According to the present invention, an input signal is linearly predictively analyzed, an impulse response of the linear prediction filter is determined, a cross-correlation between these input signals and the impulse response is determined, and the maximum absolute value of the cross-correlation is calculated. The first pulse with the magnitude of the cross-correlation is set at the position, and the new cross-correlation is obtained by subtracting the autocorrelation of the impulse response at the position where the cross-correlation pulse was set, normalized to the pulse magnitude. Similarly, from the cross-correlation, a predetermined number of pulses are determined within a predetermined range of the cross-correlation, and the coefficients of the linear prediction filter and the positions and magnitudes of the predetermined number of pulses are transmitted. A speech encoding method, in which cross-correlation is thinned out at intervals of two or more samples, the position of one or more samples with a large absolute value is determined in the thinned out signal, and the position of the sample with the maximum value is determined, A speech encoding method is obtained in which the position is set as the position where the absolute value of the cross-correlation is maximum.

1 ri 1 Hereinafter, the present invention will be explained in detail using the drawings.

FIG. 1 is a block diagram of a multi-pulse driven speech encoding method applied to an embodiment of the present invention. The input signal X is applied to a weighted G filter 1 to obtain an auditory weighted signal XW.

Further, the input signal X is subjected to upstream linear prediction analysis by the LPG analysis unit, and a linear prediction coefficient A is obtained. The impulse response R of the linear prediction filter having the linear prediction coefficient A as a coefficient is expressed as LP
The G parameter is determined by the collapsing/encoding unit 4. Then, the cross-correlation section 2 and the autocorrelation section 5 calculate the cross-correlation φx,r between the weighted signal XVI and the impulse response R, and the autocorrelation φ'r of the impulse response R, respectively. Correlation φxr and autocorrelation φr
The position and amplitude of the obtained excitation pulse are encoded and transmitted together with the linear prediction coefficient A in the encoding section 7.

Next, a pulse search method will be explained. FIG. 2 shows an example of the internal configuration of the pulse search section 6 shown in FIG. 1. Inside the pulse search section 6, there is a cross-correlation holding means 61, which has a length l.
The cross-correlations φxr(i) of x, i-1 to lx are maintained. In the maximum absolute value search unit 62, the length ix
A value φxr(is) at which the absolute value of the cross-correlation φxr(i) of i=1 to lx is maximum is determined, and a pulse of magnitude φxr(is) is generated at the position il.

From the cross-correlation φX', the autocorrelation φ' of length ±lr is set at the position il and the magnitude is determined by the position/amplitude correction unit 63, φX'(i
(2), and then the subtraction unit 64 performs subtraction as shown in the following equation.

φxr(im+i)=φXr(llll+1)−(φ
xr(is)/φrr(0))×φrr(i)
(i--Ir to Ir) This maximum value search and the process of subtracting autocorrelation from cross-correlation are performed once every time one pulse is obtained. Then, the result of subtracting the autocorrelation from the crosscorrelation is again set in the crosscorrelation holding means 61 and used together with the crosscorrelation in the next pulse search.

Maximum search unit 62 for the absolute value of cross-correlation in the above processing
The processing in is performed by the method shown in FIG. first,
A 1/N decimation section 621 decimates the cross-correlation φxr(i) to N samples, and a detection section 622 detects the maximum absolute value of the decimated signal. Next, centering on the position of the maximum value obtained, a total of 2
Maximum value detection is further performed in the partial detection unit 623 using (N-1)+1 samples.

That is, this method first performs coarse detection of the maximum value by thinning out by 1/N, and then performs more precise detection around the maximum value position obtained thereby.

Assuming that the maximum value is detected by comparing the size of samples,
The number of times of magnitude comparison processing is Ix/N-1 times in rough detection and 2x(
N-1) times, total lx /N+2 (N-1) -1
times.

Comparing this with the number of size comparisons of 1x-1 in the conventional method of comparing all samples in size without thinning out, we get (Ix /N+2 (N-1) -1)/(lx-1)
#1/N (lx>N), and when IX is sufficiently larger than N, it is approximately 1/N.

In this embodiment, only one maximum value is detected in the coarse detection of the maximum value of cross-correlation, but two maximum values are detected in order from the maximum value.
It is also possible to detect more than one and use it as a candidate for the maximum value position.

In this case, precise detection is performed on all candidates for the maximum value position determined by rough detection.

DETAILED DESCRIPTION OF THE INVENTION According to the present invention, in the maximum value search for cross-correlation, the number of cross-correlation samples is reduced from N by first thinning out to 1 N samples and performing a coarse search for the maximum value. When the value is sufficiently large, there is an effect that the number of times of size comparison processing in the maximum value search can be reduced to about 1/N of the conventional value.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of the present invention, FIG. 2 is a block diagram showing the configuration of the pulse search section of FIG. 1, and FIG.
FIG. 2 is a block diagram showing the configuration of a maximum absolute value search unit shown in the figure. Explanation of symbols of main parts

Claims (1)

[Claims] Linear prediction analysis is performed on the input signal, the impulse response of the linear prediction filter is determined, the cross-correlation between these input signals and the impulse response is determined, and the position of the maximum absolute value of the cross-correlation is set at the position of the magnitude of the cross-correlation. A new cross-correlation is obtained by subtracting the autocorrelation of the impulse response at the position where the cross-correlation pulse was set, normalized to the pulse size, from the cross-correlation. A speech encoding method that detects a predetermined number of pulses within a predetermined range of cross-correlation and transmits the coefficients of the linear prediction filter and the positions and magnitudes of the predetermined number of pulses. Thin out the correlation at intervals of two or more samples, find the position of one or more samples with a large absolute value in the thinned out signal, find the position of the sample with the maximum value, and set this position as the maximum absolute value of the cross-correlation. A speech encoding method characterized by a position of .