JPH05313698A - System and device for multi-pulse voice encoding - Google Patents

Abstract

PURPOSE:To improve the precision of a pulse search by calculating the autocorrelation of impulse responses and making the pulse search by using them. CONSTITUTION:A voice signal X(n) which is applied to an input terminal 10 is supplied to an interpolator 1 and interpolated there. The interpolated voice signal XI(n) is supplied to an LPC analyzer 2, a pitch extracter 3, and a weighting filter 4. The weighting filter 4 finds a voice XIW(n) by auditory correction based upon the interpolated voice signal and a linear predictive filter coefficient and supplies it to a cross-correlation calculator 6. Then, the cross correlation RHX(n) and autocorrelation RHH(n) are supplied to a pulse searcher 8. The pulse searcher 8 searches for pulses by using the cross correlation RHX(n) and autocorrelation RHH(n). Information obtained by the pulse search is supplied to an encoder 9.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a multi-pulse speech coding system and its apparatus, and more particularly to a multi-pulse speech coding system and its apparatus for improving sound quality when pitch prediction is involved.

[0002]

2. Description of the Related Art In the conventional multi-pulse speech coding system with pitch prediction, since the pitch is calculated at the sampling interval of the input speech, the accuracy of the pitch is the sampling interval of the input speech.

[0003]

In this conventional multi-pulse speech coding system with pitch prediction, the accuracy of the pitch is the sampling interval of the input speech, but the pitch is not always accurately determined. Impulse response of synthesis filter with pitch prediction, cross-correlation of impulse response of synthesis filter with pitch prediction and speech signal, autocorrelation of impulse response of synthesis filter with pitch prediction, and pulse search accuracy are not sufficient, improving sound quality However, there is a drawback that there is a limit.

[0004]

According to the multi-pulse speech coding method of the first invention, a speech signal having an increased number of samples is created by interpolating an input speech, and pitch information is extracted from the speech signal. Calculating the impulse response of the synthesis filter with pitch prediction from the speech signal and the pitch information, and the cross-correlation of the impulse response of the synthesis filter with pitch prediction and the speech signal, and the autocorrelation of the impulse response of the synthesis filter with pitch prediction And the cross-correlation is used as an evaluation function, a first pulse is set at the maximum value of the values standing at the sample position of the input voice in the evaluation function, and the pulse amplitude is normalized from the evaluation function at the pulse position. Search for a predetermined number of pulses by sequentially repeating the generation of the converted autocorrelation function as a new evaluation function. And wherein the Rukoto.

A multi-pulse speech coding apparatus according to a second aspect of the invention is an interpolator for interpolating an input speech to generate a speech signal having an increased number of samples, and a pitch extractor for extracting pitch information from the speech signal. An impulse response calculator that calculates the impulse response of the synthesis filter with pitch prediction from the voice signal and the pitch information; and a cross-correlation calculator that calculates the cross-correlation of the impulse response of the synthesis filter with pitch prediction and the voice signal. And, an autocorrelation calculator that calculates the autocorrelation of the impulse response of the pitch-predictive synthesis filter, and the cross-correlation as an evaluation function, with the maximum value of the values standing at the sample position of the input voice in the evaluation function. A first pulse is set, and a function obtained by subtracting the autocorrelation normalized by the pulse amplitude from the evaluation function at the position of the pulse is given. And a pulse searcher for searching for a predetermined number of pulses by repeating Do evaluation function that sequential to have.

[0006]

The present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a multi-pulse encoding device showing an embodiment of the present invention, FIG. 2 is a diagram showing an example of signals of respective parts in FIG. 1, and FIG. 3 is a flowchart showing a flow of a pulse search operation in FIG. is there.

In the multi-pulse speech coding apparatus of this embodiment, as shown in FIG. 1, the speech signal X (n) (n = 0, 1, ..., N-1) applied to the input terminal 10 is interpolated. 1 and is interpolated here. The interpolated voice signal X _I (n) (n = 0, 1, ..., 2N−1) is supplied to the LPC analyzer 2, the pitch extractor 3, and the weighting filter 4. In the LPC analyzer 2, the interpolated audio signal X _I (n)
Is subjected to LPC analysis, the coefficient α of the linear prediction filter is obtained, and the coefficient α is supplied to the weighting filter 4, the weighting impulse response calculator 5, and the encoder 9. The pitch extractor 3 extracts pitch information from the interpolated voice signal and supplies it to the weighted impulse response calculator 5 and the encoder 9. In the weighting filter 4, the audio signal X _IW (n) (n = 0, 1, ..., 2N−1) that is aurally corrected from the interpolated audio signal XI (n) and the coefficient α of the linear prediction filter is output. It is obtained and supplied to the cross-correlation calculator 6. In weighting the impulse response calculator 5, the weighted impulse _response of the pitch prediction synthesis filter obtained by the auditory correction of the coefficients of the linear prediction filter α and pitch information _{H IW (n) (n =} 0,
1, ..., M-1) are obtained and supplied to the cross-correlation calculator 6 and the auto-correlation calculator 7. In the cross-correlation calculator 6, the voice signal X _IW (n) and the weighted impulse response H _IW (n)
Cross-correlation R _HX (n) (n = 0, 1, ..., 2N−1) is calculated, and the auto-correlation calculator 7 calculates the auto-correlation R _HH (n) (n) of the weighted impulse response H _IW (n). = 0, 1,
, M-1) is required. And the cross-correlation R _HX (n)
And the autocorrelation R _HH (n) is supplied to the pulse searcher 8. The pulse searcher 8 performs a pulse search using the cross-correlation R _HX (n) and the auto-correlation R _HH (n). The pulse information obtained by the pulse search is supplied to the encoder 9. The encoder 9 encodes the linear prediction coefficient α, pitch information, and pulse information and supplies the encoded information to the output terminal 11.

Next, the processing of each part in this embodiment will be described with reference to FIG. 2 and FIG.

(1) Pitch extraction: Pitch extraction by the pitch extractor 3 is performed by using the input speech X (n) (n shown in FIG. 2 (a).
= 0,1, ..., N-1) is performed on the interpolated audio signal X _I (n) (n = 0, 1, ..., 2N-1) shown in FIG. Finer, more accurate pitch is required.

(2) Weighted impulse response of synthesis filter with pitch prediction: Since the pitch is more accurately obtained, the weighted impulse response H _IW (n) (n = 0) of the synthesis filter with pitch prediction shown in FIG. , 1, ..., M
-1) is also obtained more accurately than in the case where the input voice X (n) is not interpolated.

(3) Cross-correlation: Weighted impulse response H _IW (n) (n =
, M-1) and the weighted speech signal X _IW (n) (n = 0, 1, ..., 2N-1) shown in FIG. 2C, R _HX (n) (n) = 0, 1, ..., 2N-1) (FIG. 2
(Shown in (e)) is also more accurately calculated as compared with the case where the input voice X (n) is not interpolated.

(4) Autocorrelation: Weighted impulse response H _IW (n) (n =
0, 1, ..., M−1) autocorrelation R _HH shown in FIG.
(N) (n = 0, 1, ..., 2N-1) is also the input signal X
It can be calculated more accurately than when (n) is not interpolated.

(5) Pulse search: First, the number of pulses 1 is cleared (step 21 in FIG. 3). Next, the cross correlation (evaluation function) R _HX (n) (n = 0, 1, ..., 2) shown in FIG.
The maximum value R _HX (m) is searched from among the values standing at the sample position of the input voice X (n) before interpolation in N-1) (step 22 in FIG. 3). And the pulse amplitude P (m) = R
Calculate _HH (m) / _RHH (0) (step 2 in FIG. 3)
3). Where R _HH (0) is the zeroth order autocorrelation of the weighted impulse response H _IW (n) of the pitched synthesis filter. Next, the evaluation function R _HX (n) (n =
0, 1, ..., 2N-1) to weighted impulse response H
_IW (n) autocorrelation R _HH (n) (n = 0, 1, ..., M−
R obtained by subtracting 1) normalized by the pulse amplitude P (m)
_{HX (n) = R HX (} n) -P (m) · R HH (│n-m│)
(Here, n = m−M, m−M + 1, ..., M + M−1, m
+ M) as a new evaluation function (step 2 in FIG. 3)
4). Then, the pulse number 1 is counted (step 25 in FIG. 3), and the pulse search is repeated until the pulse number 1 reaches a predetermined number L (step 26 in FIG. 3).

When the pulse search is performed in this manner, the cross-correlation R _HX (n) and the auto-correlation R _HH (n) are accurately obtained as compared with the case where the pulse search is performed without interpolating the input speech X (n). Therefore, the pulse search can be performed more accurately, and the synthesized speech can be brought closer to the original sound. Further, since the position where the pulse is set is only the sample position of the input voice X (n) before the interpolation, the amount of pulse information to be transmitted does not change. As the transmission information, only the pitch information amount increases, but the pitch information is much smaller than the pulse information and the transmission information amount is almost unchanged.

[0015]

As described above, according to the present invention, the pitch information, the impulse response of the synthesis filter with pitch prediction, and the mutual response of the impulse response and the speech signal are calculated from the speech signal in which the number of samples is increased by interpolating the input speech. correlation,
Since the autocorrelation of the impulse response is calculated and the pulse search is performed using these, the pitch information, the impulse response of the synthesis filter with the pitch prediction, and the impulse response are compared to the case where the pulse search is performed without interpolating the input speech. It is possible to improve the cross-correlation between the sound signal and the voice signal, the autocorrelation of the impulse response, and the accuracy of the pulse search, and to improve the sound quality.

[Brief description of drawings]

FIG. 1 is a block diagram of a multi-pulse encoding device showing an embodiment of the present invention.

FIG. 2 is a diagram showing an example of signals of respective units in FIG.

FIG. 3 is a flowchart showing a flow of a pulse search operation in FIG.

[Explanation of symbols]

 1 Interpolator 2 LPC Analyzer 3 Pitch Extractor 4 Weighting Filter 5 Weighting Impulse Response Calculator 6 Cross Correlation Calculator 7 Autocorrelation Calculator 8 Pulse Searcher 9 Encoder 10 Input Terminal 11 Output Terminal

Claims (2)

[Claims] 1. In a multi-pulse speech coding method with pitch prediction, an input speech is interpolated to create a speech signal with an increased number of samples, pitch information is extracted from the speech signal, and the speech signal Calculate the impulse response of the pitch prediction with synthesis filter from the pitch information, calculate the cross-correlation of the impulse response of the pitch prediction with synthesis filter and the voice signal and the autocorrelation of the impulse response of the pitch prediction with synthesis filter,
The cross-correlation is used as an evaluation function, the first pulse is set to the maximum value of the values standing at the sample position of the input voice in the evaluation function, and the pulse amplitude is normalized from the evaluation function at the pulse position. A multi-pulse speech coding method characterized in that a predetermined number of pulses are searched by sequentially repeating a function obtained by subtracting autocorrelation as a new evaluation function. 2. A multi-pulse speech coding apparatus with pitch prediction, comprising an interpolator for producing a speech signal with an increased number of samples by interpolating an input speech, and extracting pitch information from the speech signal. A pitch extractor, an impulse response calculator that calculates an impulse response of a synthesis filter with pitch prediction from the speech signal and the pitch information, and a mutual calculation that calculates a cross-correlation between the impulse response of the synthesis filter with pitch prediction and the speech signal. Correlation calculator, an autocorrelation calculator that calculates the autocorrelation of the impulse response of the synthesis filter with pitch prediction, and the cross-correlation as an evaluation function, of the value standing at the sample position of the input speech in the evaluation function A first pulse is set at the maximum value, and the autocorrelation normalized by the pulse amplitude is subtracted from the evaluation function at the pulse position. And a pulse searcher for searching a predetermined number of pulses by sequentially repeating the existing function as a new evaluation function.