JPH0761044B2 - Speech coding method - Google Patents

Abstract

PURPOSE:To minimize a waveform distortion under a constant total information quantity by following a sound signal for an analyzing section to divide the sound signal for the number of plural samples and distributing adaptably the information quantity of a spectrum enveloping information code and the information quantity of a waveform information code. CONSTITUTION:Encoders 31, 32 and 33 composed of an estimation coefficient extracting part 12, an estimation residual extracting part 13, quantizing parts 14 and 17 and a reverse quantizing part 16 are provided, and in respective encoders, an input sound signal is quantized, for example, by three types of the information distributing method set beforehand as shown in the figure. Here, a total information quantity T per one analyzing section is fixed. Respective encoders are encoded respectively by the first, second, third information distributing methods in the figure. Respective encoding outputs are respectively encoded by a local decoding deciding part 34, the most desirable distributing method is obtained and a spectrum enveloping information code 35, a waveform information code 36 and a distributing information code 37 are transmitted to a decode side. The most desirable method means that a local decode signal and a local signal are compared, the quantizing distortion is obtained and this comes to be minimum.

Description

Description: TECHNICAL FIELD The present invention relates to a method of separating a speech signal into spectral envelope information indicating a shape of a spectral envelope and a waveform information code indicating a waveform and encoding the speech signal.

“Conventional Technology” Conventionally, in order to efficiently encode a speech waveform, it has been considered to use a statistical correlation between adjacent sample values of a speech signal, and a linear prediction method is particularly effective. . This method basically divides a voice signal into sections for each number of samples, obtains a prediction coefficient that reflects the shape of the spectrum envelope of the voice signal for each section, encodes the coefficient as a spectrum envelope information code, and The prediction residual of the voice signal is encoded as a waveform information code indicating the waveform of the voice signal in the section, and these two codes are combined by a decoder.

The block diagram is shown in FIG. Digital audio signals indicating sample values of the audio signal at fixed time intervals are input from the input terminal 11 to the prediction coefficient extraction unit 12 and the prediction residual extraction unit 13, respectively, and are divided into analysis intervals of a fixed number of samples, respectively. The prediction coefficient indicating the shape of the spectrum envelope of the audio signal is extracted by the prediction coefficient extraction unit 12 for each analysis section, the extracted prediction coefficient is quantized by the quantization unit 14, and the spectrum envelope information code 15 is output. . The spectrum envelope information code is decoded by the dequantization unit 16, the prediction residual extraction unit 13 is controlled by the decoded prediction coefficient, and the prediction residual in the corresponding analysis section is extracted. The prediction residual extraction unit 13 is a so-called inverse filter having an inverse characteristic of the voice synthesis filter. This prediction residual signal is quantized by the quantizer 17 to obtain the waveform information code 18.

The spectrum envelope information code 15 and the waveform information code 18 are respectively decoded by the decoding units 21 and 22 on the decoder side, and the prediction residual signal decoded by the decoding unit 22 is supplied to the synthesis unit 23 of the synthesis filter as a driving excitation signal. The filter characteristic of the synthesis unit 23 is controlled by the prediction coefficient decoded by the decoding unit 21, and the sound signal is synthesized and output from the synthesis unit 23.

Various methods have heretofore been made for the expression method of the prediction coefficient, that is, the spectrum envelope information, the quantization method, the expression method of the prediction residual signal, that is, the waveform information, and the quantization method, and various coding methods have been proposed. For example, adaptive predictive coding (edited by The Institute of Electronics and Communication Engineers, "Application of Digital Signal Processing", Corona Publishing Company, 1981.
Annual 183p-), multi-pulse coding (US patent 4472832)
"Digital Speech Coder" (September 1984), adaptive transform coding (Japanese Patent No. 1258025 "adaptive transform coding system of speech"), etc. are all based on the configuration shown in FIG.

In all of these conventional methods, the spectrum envelope information code 15
The amount of information distributed between the waveform information code 18 and the waveform information code 18 is fixed to be most desirable on average.

However, in real speech, the optimal information distribution changes for each analysis interval due to the fact that the statistical properties of the waveform change from moment to moment and the quantization distortion changes. That is, with respect to a certain two analysis sections A and B, the spectrum envelope information code, the waveform information code, and the total amount of information are made constant, and the analysis sections A and B are The SNRs of the coded speech are the curves 25 and 26 in FIG. 7, respectively. (256 samples / analysis section, total
256 bits / analysis interval). The curve 25 shows that the SNR tends to increase as the information amount of the spectrum envelope code increases, but the information amount of the spectrum envelope code increases from 16 bits to 2 bits at 19 bits.
It is lower than B. Curve 26 is relatively flat,
Even a difference of 1 bit has a fluctuation of about 1 dB, and if the number of bits is relatively different, it is about 2 dB. Since the SNR is originally about 10 to 20 dB, the difference between 1 dB and 2 dB is quite large.

However, the conventional encoding method does not consider the fluctuation of the SNR due to the distribution of the information amount for both codes, which means that the distortion due to the encoding can be further reduced.

It is an object of the present invention to provide a speech coding method that minimizes waveform distortion or perceptually weighted waveform distortion under a constant total information amount.

"Means for Solving the Problem" The present invention adaptively distributes the information amount of the spectrum envelope information code and the information amount of the waveform information code in accordance with the voice signal which is changing every analysis section. .

That is, in the present invention, both the quantizer that encodes the spectrum envelope information and the quantizer that encodes the waveform information can be quantized with a plurality of types of information amounts, and the information amount and the waveform of the code of the spectrum envelope information A plurality of types of distribution with respect to the information amount of the information code are set in advance, and 1 is selected from the set plurality of distributions so that the quantization distortion in the analysis section becomes small.
One adaptively.

"Embodiment" FIG. 1 shows an embodiment of the present invention. In this example, the prediction coefficient extraction unit 12, the prediction residual extraction unit 13, and the quantization unit shown in FIG.
14, 17 and three encoders 31, 32, 33 including the inverse quantizer 16 are provided. In these encoders 31, 32, 33, for example, the second encoder
The audio signal from the input terminal 11 is quantized by three preset information distribution methods as shown in the figure. Here, the total amount of information T per analysis section is fixed and is 256 bits in this example, that is, T (total amount of information) = E (spectral envelope code information amount) + W
(Waveform code information amount) + D (distributed information amount). The encoders 31, 32, 33 are respectively the first, second, and third in FIG.
It is encoded by the third information distribution method.

Each of the encoded outputs of these encoders 31, 32, 33 is decoded by the local decoding determination unit 34 to obtain the most desirable distribution method, and its spectrum envelope information code 35, waveform information code 36, and distribution information code 37 are obtained. Transmit to the decoding side. The most desirable here means that the locally decoded signal and the input signal are compared to obtain the quantization distortion, and this is the minimum. In that case, quantization distortion weighted perceptually may be used depending on the application. As shown in FIG. 3A, the distribution information code 37 is first transmitted, and subsequently, the spectrum envelope information code 35 and the waveform information code 36 are sequentially transmitted. It is more effective to use a uniquely decodable variable length code as shown in the example of FIG. 2 as the distribution information code according to its frequency. That is, since there are three types of distribution methods in FIG. 2, two bits are required to display each distribution method. Of these three types, distribution information indicating the third type of distribution method with the highest frequency of occurrence. 1 bit of code "1", and 2 for the other two types of distribution information codes
Use bits.

On the other hand, on the decoding side, the distribution information code 37 is first decoded by the decoding unit 41. In accordance with the decoded distribution information, the spectrum envelope information code 35 and the waveform information code 36 are decoded by the decoding units 42 and 43, respectively.
The decoded output is supplied to the synthesizing section 44 and synthesized to obtain an output voice.

In order to determine the model of information distribution, the number of bits of the spectrum envelope information code with which the distortion becomes smaller on average is selected using the voice signal sample. At this time, the trade-offs caused by the following two points may be taken into consideration.

If the allocation information code is m bits, 2 ^m kinds of allocation templates can be set.

If the distribution information code is m bits, the total information amount T is constant, so that, for example, the waveform information code decreases by m bits. In this case, SNR is ₁₀ log _{10 on} average in the case of N samples in one analysis section.
(22m ^{/ N} ) [dB] = 6.02m / N [dB] decreases.

In FIG. 1, as the quantizers for quantizing the spectrum envelope information and the waveform information respectively with a plurality of kinds of information amounts, the quantizers for the respective information amounts are respectively provided.
One of each of the quantizers may be provided and the amount of information may be changed according to the information distribution. For this purpose, a quantizer capable of supporting an arbitrary number of bits is required. For this purpose, for example, vector / scalar quantization combining vector quantization and scalar quantization (Japanese Patent Application No. 57-204849 "Vector Quantization" The quantization method “) may be applied to the spectrum envelope information quantization and the waveform information quantization.

As shown in the above-mentioned patent application “Vector Quantization Method”, when a two-stage quantizer is used such that vector quantization is performed first and then scalar quantization is performed, for example, a portion corresponding to FIG. First, the quantization unit 14 vector-quantizes the extracted spectrum envelope information as shown with the same reference numeral,
From the provisional spectral envelope characteristic obtained from the quantized value,
The determination unit 51 determines the information amount of the spectrum envelope information code. For example, if the provisional spectrum envelope characteristic by vector quantization is flat, the information amount for the spectrum envelope is set to a small predetermined value, and the corresponding amount of information is used for the second-stage quantization for the spectrum envelope information. Performs scalar quantization. Further, the determination unit 51 supplies the remaining information amount of the predetermined information amount T to the quantizer 17, and quantizes the waveform information into the information amount. Therefore, in this case, Fig. 3B,
As shown in C, a fixed number of bits from the beginning in the spectrum envelope information code 35 is the first-stage quantization (vector quantization) code, and when the spectrum envelope characteristic is flat, FIG. As shown in B, the amount of quantization information in the second stage is set to a small predetermined value, and when the spectral envelope characteristic has a large change, as shown in FIG. Is a large predetermined value. In this way, since the amount of information of the spectrum envelope information code is determined by the first-stage quantization, it is not necessary to send the distribution information.

The setting of the information amount of the second-stage quantization may be performed by collecting learning samples having the same provisional spectral characteristics and performing statistical processing.

On the decoding side, a code having a constant length from the beginning in the received code string, that is, the information amount (code length) of the second-stage quantized code in the determination unit 52 from the first-stage quantized code of the spectrum envelope information. And the decoding unit 42 performs the second-stage decoding in accordance with this. Of course, before that, the first-stage decoding is performed.
At the same time, the determination unit 52 determines the information amount of the waveform information code 36 by the decoding unit.
It gives it to 43 and decodes it.

In this way, when the spectrum envelope information is quantized in two stages, the amount of the spectrum envelope information code may be controlled from the quantized code of the one stage as follows. That is, an information distribution table according to the spectrum envelope characteristic is prepared in advance as shown in, for example, FIGS.
Is for a flat spectral characteristic and B is for a varying spectral characteristic. The code 35 ₁ obtained by the first-stage quantization (vector quantization) for the spectrum envelope information is used to determine which of the information distribution tables in FIGS. 5A and 5B is to be used, and the determined information distribution table is used. In this example, the three types of information distribution are performed, the second-stage quantization (scalar quantization) for the spectral envelope information and the waveform information are performed, and the three types of quantization codes in the quantization code are used. The one with the smallest distortion is determined and its code is shown in Fig. 3D, E.
The first-stage quantized code of the spectral envelope information as shown in
35 ₁ , distribution information code 37, second-stage quantization code 35 _{2 of} spectrum envelope information, and waveform information code 36 are transmitted in this order. On the decoding side, the information allocation table to be used is determined from the code 35 _1, and the information allocation in the information allocation table is known from the allocation information code 37,
The codes 35 ₂ and 36 are taken out and decoded respectively.

When the spectrum envelope information is quantized in multiple stages, the multistage quantization may be performed only by vector quantization or scalar quantization, not limited to vector quantization and scalar quantization.

[Advantages of the Invention] As described above, according to the method of the present invention, encoding with a smaller final distortion is realized under a constant amount of information.

For a voice signal of 8 kHz sample and 256 samples of 1 analysis section, 4.8 to 9.6 kilobits / second encoding with 3 bits of allocation information code is 0.5 to 1.0 dBSNR from fixed allocation encoding.
Was confirmed to improve. SNR was originally 10-20d
Since it is about B, this improvement is quite good. In the fluctuation of SNR as shown in FIG. 7, allocation information code 3
By using bits, it is possible to use the optimum one among the eight types, and by reducing the waveform information code by 3 bits, the final SNR is about ₁₀ log ₁₀ (2 ^{2 ×} 3/256) = 0.07 [on average]. dB)
The improvement in the SNR can be expected from the fact that it is only decreased.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of an encoding method of the present invention, FIG. 2 is a diagram showing an example of information distribution, and FIG. 3 is a diagram showing various examples of encoded output by the encoding method of the present invention. FIG. 4 is a block diagram showing another embodiment of the encoding method of the present invention, FIG. 5 is a diagram showing an example of information distribution, and FIG. 6 is a block diagram showing a conventional encoding method using linear prediction. FIG. 7 is a diagram showing an example of changes in SNR according to the amount of spectrum envelope information code while keeping the total amount of information constant.

Claims (1)

[Claims] 1. A speech signal is divided into an analysis section for each of a plurality of samples, a spectrum envelope information code indicating a shape of a spectrum envelope of the speech signal in the analysis section, and a waveform information code indicating a waveform of the speech signal in the analysis section. In the method of separating into and encoding, a quantizer for the spectrum envelope information code that can be quantized with a plurality of preset information amounts, and a quantizer with a plurality of preset information amounts A quantizer for the above waveform information code is provided, and a plurality of types of distribution of the information amount for both quantizers are set in advance, and the information amount is set so that the quantization distortion becomes small for each analysis section. A speech coding method characterized by adaptively deciding the distribution of the values from the plurality of values set above.