JPS5942320B2 - Audio processing method - Google Patents

Description

[Detailed description of the invention] The present invention relates to a method for predictive encoding and demodulation of audio signals.

The predictive encoding method is characterized by being able to encode data into a smaller number of bits than the pulse code modulation method (hereinafter referred to as PCM).

The present invention further enhances this effect.

This will be explained below with reference to the drawings.

Figure 1 shows the differential pulse code modulation method (Differential pulse code modulation method).
alPulseCodeModulation, DPC
FIG. 2 is a configuration diagram of a conventional audio processing method using M). In Fig. 1, Si is a discrete audio input signal, Si is a predicted signal, Θ is a subtracter, el is a difference signal, QUAN is a quantizer, ei゜ is a quantized difference signal, 4 is an adder, Si゜is the addition signal, DPRED is the predictor, COD is the encoder, and Di is the encoded output color (where i=1, 2, 3, . . .
・・）.

Si is usually a past value, Si-0, Si-, , Si-O,
......Predicted from Si-n. If the prediction coefficients are a0, a2, a3, ... an, then △
Si■a4Si-1+ a2si-2+a3si-3+
・・・・・・+anSi-n・・・・・・・・・
・・・・・・・・・・・・・・・・・・・・・・・・
It is expressed as (1).

Also, △ ei.

Si-Si3゜゜゜゜゜゜'゜'゜゜゜'゜'゜゜゜゜゜゜゜゜゜゜゜゜010゜゜゜゜゜゜゜゜ (2), and e
The prediction coefficients al, a2,
a3, ... an is determined. For example, the long-term autocorrelation coefficient of an audio signal is R for a delay that is an integral multiple of the sampling period. , R0, R2, R3,...
- When Rn, a0, a2, a3...an are obtained as solutions of the following simultaneous linear equations. Using the coefficients obtained from equation (3), RI is obtained as shown in (1), and Ei is therefore sequentially obtained from equation (2). The amount of e obtained by quantizing Ei is slightly larger7. '． :＾悴=;34;
".,.=-,:?weight=艷1tQ;←delay element (composed of a register etc.) equal to the pulling period (consisting of a register, etc.), Al,
a2, . . . An are coefficient multipliers.

Thus, by encoding Ei* with COD,
Signal sequence {Di} with reduced redundancy than the sequence of {Si}
is obtained. FIG. 2 shows a demodulation method for restoring the original audio signal sequence from the encoded signal sequence {Di}.

In FIG. 2, DEC is a decoder, and the rest is exactly the same as in FIG. Here, Di is E by DEC.
It is inversely converted to i8 and Si8 is derived from Ei8, but
Although this S value includes some errors due to quantization, it is approximately S
The one equal to i is reproduced. The above is the conventional predictive coding and its demodulation method, but because the predictive coefficients are fixed, no significant effect can be expected. for example,
In order to obtain the same variety of audio as 8-bit PCM, 6-bit predictive coding was necessary, and the information compression rate was limited to 3/4. The present invention eliminates the above-mentioned drawbacks and provides a predictive encoding method that can achieve an information compression rate of about 10.

Another problem with conventional predictive coding methods is that the prediction coefficients are determined from the long-term autocorrelation coefficients of speech, and the coefficients are fixed.

The information compression rate effect of the predictive coding method depends entirely on the statistical properties of the input signal sequence, and it is difficult for input signals such as audio signals that cannot be considered to be a stationary stochastic process over a long period of time. , it is not a good idea to keep the prediction coefficient constant. This is because, while the spectral distribution of the audio signal changes, if the prediction coefficient is constant, the reproducibility of the vector becomes worse. However, the change in the spectral distribution of speech is based on the movement of the articulator tube, and is relatively gradual, and can be regarded as an almost stationary signal over a short period of several or several + Ms (for example, 20 to 30 ms).

The present invention focuses on this point, determines a prediction coefficient from short-time autocorrelation coefficients every 20 to 30 ms, and dynamically controls the predictor to improve the reproducibility of the speech spectrum. In this way, the amount of information can be compressed to approximately 1/2 with the audio quality comparable to that of the 8-bit PCM system, resulting in highly efficient transmission and highly efficient recording. That is,
The number of multiplex lines is doubled in the same transmission system, and the same memory can accommodate twice as many words in a recording/editing speech synthesizer, making it possible to realize an extremely economical system. . Next, embodiments according to the present invention will be described with reference to the drawings. FIG. 4 is an explanatory diagram of the speech encoding method according to the present invention. In FIG. 4, Akn (k=1,
2, 3,...: n=1, 2, 3,...
) is a prediction coefficient in the k-th interval, MIX is a mixer, DPRE is a predictor shown in FIG. 6, and the others are exactly the same as in FIG. 1. Unlike the conventional encoding method, Akn is determined for each short period of 20 to 30 ms so that the root mean square of the difference signal Ei in that period is minimized. At this time, for example, Ak,,Ak2, in the kth section,
Ak3,...Akn are obtained as solutions of the following one-dimensional simultaneous equations. Just RkO, Rkl, Rk2,...
. . . Rkn is an autocorrelation coefficient in the k-th interval (delay time is an integral multiple of the sampling period).

Yotori: Kannoko? Using the predicted coefficients determined by i, the following is derived, and Di is derived in the same manner as in the conventional method.

Here, Di and Akn are mixed by MIX to obtain a mixed signal sequence {Mi}. Figure 6 shows the configuration of a predictor that performs the calculation of equation 5, where 2 is a multiplier and the others are 3rd
It is exactly the same as the one shown in the figure. In the multiplier, the past value Si
8l-n (n=1, 2,...) and prediction coefficient Ak
Derive the product of n (n=1, 2,...). At this time, Akn is updated every time the interval changes and is kept at a constant value within that interval. The sequence {Mi} is the encoded audio signal according to the present invention.

A method for demodulating this and reproducing the original audio signal is shown in FIG. In FIG. 5, SEP is a separator, DEC is a decoder, and the other components are exactly the same as those in FIG. SEP is {Mi
} is separated into a sequence of {Di} and a prediction coefficient Akn. The DEC transforms the code sequence {Di} into the original quantized difference signal sequence {
Ei◆} is inversely transformed into 2′I,′. ′) static, 1゛=-heavy τ:
:↑;! The sum of I, Si♦, is the reproduced audio signal, which includes some errors, but is approximately equal to Si. Next, an example will be described. Bandwidth now 4.81CIIz
Sampling the audio signal at 10KHz (8-bit PC
Let the discrete signal sequence obtained by M) be {Si}, and 25
Divide into intervals of 6 samples (25.6 ms length). Assuming that the prediction is made from the past 4 samples,
There are four prediction coefficients for each interval (Ak, ,Ak2,Ak,
, Ak4; k=1, 2, 3, . . . ). This prediction coefficient is the short-time autocorrelation coefficient (RkO, Rk,, Rk
2, Rk3, Rk4) using equation (4). The predicted signal sequence {Si} is derived from DPRE while updating the above prediction coefficient for each interval, and this {S
i} and the difference signal sequence {Ei} derived from the input signal sequence {Si} is sequentially converted into a nonlinear element in 15 steps using QUAN,
A quantized signal sequence {Ei◆} is obtained. Next, in COD {Ei
◆} is sequentially converted into a 4-bit code sequence {Di}, and MI
At X, it is mixed with Akn for each section and outputs an encoded audio signal sequence {Mi}. At this time, if one coefficient is made up of 12 bits, the total amount of information will be 4×12 bits. In MIX, this coefficient information is inserted between intervals. When reproducing the original voice from {Mi}, first extract four pieces of 12-bit coefficient information using SEP and separate it into {Di} and Akn. {Di} is sequential DE
At C, the 15-level quantized signal sequence {Ei◆} is inversely modified S::=heavy: 2/2? --6 is added and {Si◆} is output sequentially.

Although {Si◆} includes some errors due to quantization, it is a signal roughly equal to {Si}. If you do this, (
Information compression of 4 x 256 + 4 x 12) / (8 x 256) + 1/2 can be achieved, and the audio quality is almost the same as 8-bit PCM. can bring about effects.

Does Figure 7 show the relationship between prediction order and S/N improvement? ．． ′：
=,1J゜nie→:Lily:j21"゛Glaze~'2τ:=
=;゛:=This means that it can be encoded with the amount of information.

On the other hand, if the amount of information is kept constant, the effect can be evaluated in terms of S/N improvement. This S/N improvement degree (S/N)
IMP is based on the principle of DPCM and is expressed by the following equation.
--]~V凰' Figure 7 shows the overall logarithm of (S/N) IMP calculated every 30 ms from audio signals of approximately 9 seconds each generated by two men and two women. This is an S/N improvement degree curve obtained by averaging over the period.

As is clear from Fig. 7, the S/N improvement effect is large up to the fourth predicted order on the horizontal axis, but beyond the fourth, the S/N
The degree of N improvement slows down. If the prediction order is doubled to 8th, the number of bits allocated to the prediction coefficient will be doubled (from 12 bits x 4 to 12 bits x 8), but the S/N improvement will increase slightly by 1.2 dB. Only. Increasing the number of bits allocated to prediction coefficients increases the amount of information, reduces compression efficiency, and increases processing time. The present invention has the above-mentioned configuration, and by setting the prediction order to 4, it can be used with 8-bit PCM without increasing the amount of information, without decreasing compression efficiency, and without increasing processing time. It is possible to obtain the same level of voice variety.

[Brief explanation of drawings]

Fig. 1 is a block diagram of a conventional predictive coding system, Fig. 2 is a block diagram of a conventional predictive coding signal demodulation system, Fig. 3 is a block diagram of a conventional predictor, and Fig. 4 is a block diagram of a conventional predictive coding signal demodulation system. 5 is a block diagram of the encoded speech demodulation method according to the present invention, FIG. 6 is a block diagram of the predictor in the speech encoding method according to the present invention, and FIG. 7 is a block diagram of the speech processing according to the present invention. It is a figure which shows the relationship between the prediction order and S/N improvement degree in a method. QUAN...Quantizer, COD...Encoder, DPRE...Predictor, MIX...
・Mixer, SEP... Separator, DEC...
...Decoder.

Claims (1)

[Claims] 1 A predictor that derives a predicted value from past sample values, a subtracter that derives a difference signal between the sample value and the predicted value at a certain time, a quantizer that nonlinearly quantizes the difference signal, and a quantizer output It has an adder that approximately derives a past sample value (predictor input signal) from the sum of a signal and a predicted signal (predicted value), and an encoder that converts a quantizer output signal into a coded signal. In an audio processing method that performs predictive coding, a discrete input audio signal is divided into short intervals of several or tens of milliseconds, and for each interval, the sample value at a certain time within that interval and the four samples before that are divided. A predetermined prediction coefficient is set so that the root mean square value (average value over each interval) of the difference with the predicted value is the minimum, and each time the divided interval moves, it is calculated in that interval. While updating the corresponding prediction coefficient and processing the input signal within that interval,
The present invention is characterized in that it uses a predictor that maintains a constant constant value, and further includes a mixer that inserts the prediction coefficient between intervals of the encoded output signal corresponding to the divided intervals of the input signal. Audio processing method.