JPH0286231A - Voice prediction coder - Google Patents

Abstract

PURPOSE:To ensure the articulation of talking even if a transmission error of 10<-2>-10<-1> takes place without increasing the bit rate by adding an error check code to energy information most important in side information sets and primary and secondary linear prediction coefficients. CONSTITUTION:The device is provided with a down-sampler 35 sampling down a quantized residual signal, an interpolation device 55 interpolating a missing sample by the down-sampler 35 by an autocorrelation coefficient of the prediction residual signal before down-sampling and an error correction code adder 54 adding an error correction code to energy information most important in side information sets and primary and secondary linear prediction coefficients from the quantity of information reduced by the down-sampler 35. Thus, the error correction information is added without increasing the total transmission information quantity while minimizing the deterioration in the reproduced voice and the articulation is ensured even to a transmission error of 10<-2>-10<-1>.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention: - A voice prediction code that can be used even in a bad environment with a transmission error rate of 10=~ICr'' and a non-t<kζ, particularly in the field of land mobile radio communications. Converting device:・It is related to two things.

BACKGROUND ART FIG. 2 is a schematic block diagram showing an example of the configuration of a conventional speech predictive coding device. In FIG. 3 is a reproduced signal output terminal that outputs a reproduced signal obtained by reproducing the input signal; 10 is a difference device that subtracts a predicted value of an adder 22, which will be described later, from the input signal; 11 is a A coefficient unit 12 C multiplies the signal from the difference unit 10 by a coefficient G. A quantizer 13 quantizes the signal from the coefficient unit 11. This is an encoder that outputs.

14 is a coefficient unit that multiplies the signal from the quantizer 12 by a coefficient G; 15 is an adder that adds the signal from the coefficient unit 14 and the signal from the adder 22, and outputs the result to terminal 3; 16.17
. . . 18 are delay memories for delaying the signal from the adder 15 every cycle, 19. 20 . . . 21
are coefficient units that multiply the signals from the delay memories 16, 17...18 by coefficients α1, α2...αP, and the adder 22 multiplies the signals from the coefficient units 19, 20...21. The added predicted value is output to the difference unit 10 and the adder 15.

Next, the operation of the conventional speech predictive coding apparatus described above will be explained.

Now, the discrete audio signal sequence is Xl (+=1.2...)
Let x be the forward predicted value of x, then the differentiator 1
0 takes the difference between Xl and X and outputs the prediction error E1. The coefficient unit 11 is for normalizing the level of the prediction error E1, and the coefficient G is, for example, G=2° (n=0.1.
2.3.4...).

The quantizer 12 converts the normalized signal G-Ei into stages 2 to 6.
It quantizes in 4 steps, and the quantization step size is a level distribution characteristic? It is determined so that the quantization distortion is minimized.

The encoder 13 is used to encode the signal quantized in the quantization step 12 into quantized values c-E'lt of 1 bit to 6 bits. The coefficient unit 14 multiplies the quantized value GE'1 by G and converts it to the original level El1. Adder 15 adds predicted value X to signal E'i to generate an approximate value x'1 of signal input x, and sends it to delay memory 16. The approximate value is the predicted value X at the next sampling period as Xl, -1
The calculation of l (C is used and transferred to the delay memory 17. Similarly, after the P period, the approximate value XI, -, -1
is transferred to delay memory 18.

The predicted value x, is α1 and α for the contents of each delay memory, respectively.
2,... It is obtained by adding all the products multiplied by the coefficient of α2. The coefficients α1, α2, ..., α are called linear prediction coefficients, and ignoring quantization errors, xl=x
, the sum of squares of prediction errors pE+ is determined to be the minimum.

Further, the coefficients α·, α2, and °°° α2 are not fixed values, but are updated every 1Qms to 25m5.

Problems to be Solved by the Invention However, when decoding a signal encoded by a conventional speech predictive encoding device, in addition to the prediction residual signal, side information such as linear prediction coefficients, amplitude information, energy information, etc. Requires. For this reason, for example, the sampling period is 5.4 K.
When the residual signal is encoded with 1 bit per sample at Hz, Pinot Tray 1 does not achieve 6.4 Kbps, but together with side information, it becomes about 8 Kbps. Therefore, when configuring a codec for 8 KbpS, there is a problem that error correction information necessary to ensure intelligibility cannot be added in areas where there are many errors with an error rate of about 10.

In addition, the linear prediction coefficient is set to 4 to reduce the bits of side information.
I kept it to about the same level. For this reason, in the conventional speech predictive coding apparatus, when a transmission error occurs, there is a problem in that the phonological properties are significantly deteriorated.

The present invention solves these conventional problems, and
Excellent speech prediction that can add error correction information without increasing the total amount of transmitted information while minimizing deterioration of reproduced speech, ensuring intelligibility even with transmission errors of about 10 to 10 The object of the present invention is to provide an encoding device.

Means for Solving the Problems In order to achieve the above-mentioned problems, the present invention provides a downsampler that downsamples a quantized residual signal, and a sample that is omitted by the downsampler and a sample of the predicted residual signal before downsampling. An interpolator performs interpolation using a correlation coefficient or an experimentally determined interpolation coefficient, and the amount of information reduced by the down sampler is error-corrected into energy information of side information and first- and second-order linear prediction coefficients. In addition, in order to improve intelligibility, voiced/unvoiced judgment is determined experimentally by selectively using the input speech energy and the value of the first-order short-term linear prediction coefficient. The circuit includes a delay memory that performs interpolation using the calculated interpolation coefficients, a coefficient multiplier, and an adder.

Effects Since the present invention is configured as described above, it has the following effects. In other words, the speech predictive coding device according to the present invention can increase the linear prediction order contributing to intelligibility to 8 or more, and can encode the most important energy information among the side information and the first-order and second-order linear predictions. It is now possible to add error correction codes to coefficients. As a result, it is possible to ensure the intelligibility of a call even in a state where about 10 to 10 transmission errors occur without increasing the bit rate.

Embodiment FIG. 1 is a block diagram showing an embodiment of a speech predictive coding apparatus according to the present invention.

In the figure, 30 is a signal input terminal into which audio, control signals, etc. are input, 31 is a linear prediction analyzer that analyzes and branches side information and linear prediction information, and 32 is a predicted value of an adder 52, which will be described later, from the input signal. The subtraction device 33 is the subtraction device 32
34 is a quantizer that quantizes the signal from the coefficient multiplier 33. 35 is a quantizer that multiplies the signal from the quantizer 34 by a coefficient G. 35 is a quantizer that multiplies the signal from the quantizer 34 by a coefficient G. Down sampler 36 samples by percentage, down sampler 35
This is an encoder that encodes a signal from the terminal 57 and outputs a quantized residual signal to a terminal 57.

37, 38...39 are down samplers 35, respectively.
a delay memory for delaying the signal from 4 for each period;
0.41...42 are respectively delay memories 37.38
Coefficients β1, β2, etc. are added to the signal delayed by 39.
・Coefficient unit 43 for multiplying βP is coefficient unit 40.41
The adder 44 which adds the signals from 42 is an interpolation switch which works in conjunction with the down sampler 35 and selectively outputs the signal from the down sampler 35 and the signal from the adder 43. 37, 38, . . . , 39, coefficient units 40, 41, .

53 is a coefficient multiplier that multiplies the signal from the interpolation switch 44 by a coefficient G; 45 is a signal from the coefficient multiplier 53 and an adder 52;
an adder 46 that adds the signals from and outputs it to a terminal 58;
．． 47...48 are delay memories for delaying the signal from the adder 45 every cycle, 49, 50...
- 51 are coefficient multipliers that multiply the signals from the delay memories 46, 47...48 by coefficients α1, α2...αP, and the adder 52 multiplies the signals from the coefficient multipliers 49, 50...51. A predicted value obtained by adding the signals is output to the difference unit 32 and the adder 45.

54 is an error correction code addition circuit that adds an error correction code to the energy information of the side information and the first-order and second-order linear prediction coefficients in place of the information reduced by the down sampler 35, and outputs the error correction code to the terminal 56. .

Next, the operation of one embodiment of the present invention will be explained. The prediction error signal Ei is obtained by the following equation.

However, Xi is an input signal and X'+ -H is an interpolation signal.

The coefficient unit 33 multiplies this prediction error signal Ei by a coefficient G,
After that, the quantizer 34 quantizes this G-Ei and G-E'
Let it be i. The down sampler 35 generates a smoothed prediction error signal G.
- Downsample E'i to 1/Q.

Therefore, the encoder 36 encodes the quantized prediction residual signal GE'i that has been downsampled to 1/Q once every Q times. Therefore, the quantized predictive residual encoded signal obtained by encoding the quantized predictive residual signal G·E'1 output to the quantized residual code output terminal 57 is different from the conventional speech predictive encoded signal. , the signal is compressed to 1/Q.

On the other hand, the downpla 35 sends the quantized predicted residual signal G-E''1 to the encoder 36, and at the same time sends the quantized predicted residual signal G-E''1 to the delay memory 37.
The quantized predicted residual signal GE''1 compressed to /Q is sent out. If 1/Q is now 1/4, the delay memory 37 will lose information on 3 out of 4 samples.

Here, at the sample point where the down sampler 35 does not send the quantized predicted residual signal G-E'i to the delay memory 37 (when the down sampler 35 and the interpolation switch 44 are connected downward), the delay memory From 37 onwards, the memory contents of delay memory 39 are shifted sequentially. At this time, "0" is sequentially stored from the delay memory 37 to the delay memory 39.
' will be stored.

The interpolator 55 interpolates the signal missing by the down sampler 35 as shown in equation (2) below to obtain a quantized predicted residual signal GE.
Output the approximate value G−E”i of 'i.

G-E"1 = ぢβ-E'+-K...
-(2) K=Q However, βo=1 Here, first, there is a method of using the autocorrelation coefficient of the residual signal before downsampling as the coefficient β for interpolation. That is, the autocorrelation coefficient of the residual signal before quantization is calculated for each processing frame, and the interpolation coefficient β1 is a first-order correlation coefficient, β2 is a second-order correlation coefficient, and βK is a second-order correlation coefficient. This method uses correlation coefficients. This method requires interpolation coefficients to be transmitted for each processing frame, leading to an increase in side information. Therefore, when the purpose is low bit transmission, an interpolation coefficient βK common to all frames, which will be described below, is used. On the other hand, what about the interpolation coefficient β that is common to all frames? The method of determining the value using auditory evaluation or physical evaluation such as SNRseg is insufficient, and in this case, the best evaluation value obtained experimentally is used.

Furthermore, in the present invention, based on the fact that the prediction error signal is different between voiced parts and unvoiced parts, voiced/unvoiced judgment is performed using the input audio energy and the value of the first-order short-term linear prediction coefficient, and the voiced parts and unvoiced parts are Different interpolation coefficients are selected in the interval.

As explained above, an extra amount of information can be secured by downsampling the quantized predicted residual signal. For example 5
．． If 4 KHz sampling is downsampled by 1/3, the amount of information will be reduced by 4.2 KHz.

Therefore, the error correction code adder 54 uses this reduced amount of information? It can be used as an error correction code for side information.

Here, side information includes elements such as energy encoded values and linear prediction coefficients, but as a result of evaluating which elements among these are likely to be affected by errors, we found that the energy information and
It has been found that when the next and second-order linear prediction coefficients are erroneously transmitted, colored noise is generated and intelligibility is significantly degraded. On the other hand, even if third-order and subsequent coefficients of the prediction coefficients were erroneously transmitted, there was no significant deterioration.

From the above results, does the present invention efficiently correct errors? In order to
Add an error correction code to the next linear prediction coefficient.

On the other hand, the linear prediction coefficients α1, α2, .

However, if an error occurs during transmission at an order of around 4th order, the phonology will deteriorate significantly, for example, (mj) will sound strange to [n'']. Essentially, in order to correctly reproduce phonology, it is necessary to correctly transmit the speech spectrum.

For this purpose, the prediction coefficient should be of order 8 or higher.

In the present invention, the linear prediction coefficient is made sufficiently high, and even if a transmission error occurs, the linear prediction coefficient is transmitted as accurately as possible in order to improve the reproducibility of phonology.

As described above, in the above embodiment, the quantized predicted residual signal is downsampled, and the missing samples are interpolated using the autocorrelation coefficient of the predicted residual signal before downsampling or the experimentally determined interpolation coefficient. The amount of information reduced by downsampling is added to the energy information of the side information and to the first- and second-order linear prediction coefficients as an error correction code. Furthermore, in order to improve the reproducibility of phonemes, the linear prediction coefficient is set to 8th order or higher.

Therefore, in the present invention, by adding an error correction code without increasing the total transmission bit rate, it is possible to obtain reciprocated speech with good intelligibility even under very poor conditions with an error rate of about 10 to 10. can.

Effects of the Invention As is clear from the above embodiments, the present invention downsamples the quantized residual signal, interpolates missing samples using the autocorrelation coefficient of the predicted residual signal before downsampling, and Voiced/unvoiced judgment is performed using the first-order short-term linear prediction coefficient value, and interpolation is performed using an experimentally determined interpolation coefficient. Error correction codes can be added without increasing the total transmission pit rate, and linear prediction can be performed. Since the coefficients can be of 8th order or higher, the error rate is 10~
It has the effect of being able to obtain playback audio with good intelligibility even in poor conditions with a score of about 10.

[Brief explanation of drawings]

FIG. 1 is a schematic block diagram showing an actual example of a speech predictive coding device according to the present invention, and FIG. 2 is a schematic block diagram showing a conventional speech predictive coding device. FIG. 30...Signal input terminal, 31...Linear prediction analyzer, 3
2... Differential unit, 33... Coefficient/number unit, 34... Quantizer,
35... Down sampler, 36... Encoder, 37.
38.39...delay memory, 40, 41.4.2...
Interpolation coefficient unit, 43...Adder, 44...Interpolation switch, 45...Adder, 46.47.48...Delay memory, 49, 50.51...Linear prediction coefficient unit, 52...・
- Adder, 53... Coefficient unit, 54... Error correction code adder, 55... Interpolator, 56... Error correction code additional information output terminal, 57... Quantization residual code output terminal, 58
...Reproduction signal output terminal.

Claims (2)

[Claims] (1) A linear prediction analyzer that analyzes and branches side information and linear prediction information, a difference machine that obtains a prediction error signal, a quantizer that quantizes the prediction error signal, and a quantizer that quantizes the prediction error signal. a downsampler that downsamples the quantized residual signal; and an interpolator that interpolates the samples missed by the downsampler using an autocorrelation coefficient or an experimentally determined interpolation coefficient of the predicted residual signal before downsampling; It is characterized by being equipped with an error correction code adder that adds an error correction code to the energy information of the side information and the first-order and second-order linear prediction coefficients in place of the amount of information that can be reduced by the downsampler. A speech predictive coding device. (2) A circuit consisting of a delay memory, a coefficient unit, and an adder that performs voiced/unvoiced judgment using the input audio energy and the value of the first-order short-term linear prediction coefficient, and performs interpolation using different interpolation coefficients obtained experimentally. The speech predictive coding device according to claim 1, further comprising: a speech predictive coding device;