JPH01280800A - Coding system - Google Patents

Abstract

PURPOSE:To perform encoding which is efficient in sound tone quality and compression by coding the maximum bit value by a completely reproducible coding method and vector-quantizing a difference value when coding a prescribed value of difference in sound signals divided into blocks and the maximum bit digit of the absolute value of the difference in each block. CONSTITUTION:Difference values of sound signals are taken at a difference calculating section 1 which inputs digital sound signals and output of the calculating section 1 is stored in a difference block buffer 2 by one block quantity. Simultaneously, the difference from the calculating section 1 is inputted to a scale value detecting section 3 where the maximum value of the absolute value of the difference and the maximum bit position is outputted as a scale value. The scale value is inputted to a vector-quantizing (VO) code book 4 and the output of the code book 4 is selected by using the scale value. Then a code book A is selected when the scale value is smaller than a prescribed value THA and another code book B is selected when the scale value is larger than the prescribed value THA but smaller than another prescribed value THB. When the scale value is larger than the prescribed value THB but smaller than a prescribed value THC, a code book C is selected.

Description

TECHNICAL FIELD The present invention relates to a coding system, and more particularly to an adaptive VQ coding system in digital voice processing such as voice response, voice dialing, etc.

BACKGROUND OF THE INVENTION Currently, various audio signal encoding (compression) methods are being researched that provide high quality audio at a relatively low bit rate for audio storage and audio signal transmission such as voice mail. For example, telephone quality audio (0,3-3,4KHz) is 32k
CCITT Recommendation A, D P C encoded in bps
M method, A P C-A B (Ad
Aptive Prediction Codingw
j,th Adaptive Bit A11ocat
ion) method, multipulse method, and LSP (Lj, ne Spectrum Pa
1r) method. However, although these methods have good audio quality, the encoding and decoding processes are complex and require a considerable amount of hardware.

On the other hand, one of the high-quality PCM audio transmission systems for broadcasting satellites is the quasi-instantaneous companding system. Although the encoding process is simple in this method, the compression is not sufficiently efficient. Therefore, one method to improve the compression ratio of quasi-instantaneous companding is
A combination of differential PCM method and quasi-instantaneous companding may be considered. Generally, if quasi-instantaneous companding is simply applied to the differential PCM method, missing bits during compression will cause transmission errors, and the errors will accumulate in the integrator on the receiving side, making reception impossible. Therefore, Takaran, Kuninaka, and others proposed to solve this problem with a method called "missing bit accumulation" (hereinafter referred to as DC-PCM method). , Theory of Faith, V
ol, J67-BNαlO) reported that it was possible to significantly reduce quantization noise.

Apart from the method of Takanara et al., the present applicant proposed a method that combines the differential PCM method and quasi-instantaneous companding. We investigated two types of audio ossification methods that can reproduce high-quality audio through processing (small heart volume).

The basic idea of the method is to correct the compressed difference data obtained during quasi-instantaneous companding within its quantization pitch. This is a method in which the compressed difference data is sequentially decoded and compared with the original signal to correct each sample point so that the difference data has fewer errors within the number of compressed bits. Furthermore, method (2) is a method in which the compressed difference data with the minimum error power from the original signal is selected by changing the scale value and changing the quantization step radius in method I. Here, we will explain the principles and simulation results of these two types of audio signal encoding methods. First, we will explain the outline of quasi-instantaneous companding and the problems when applying quasi-instantaneous companding to the differential PCM method. Then, quasi-instantaneous companding when compressing 12 hits of audio data to 3 bits involves dividing the PCM data into, for example, 8 samples per plotter, and finding the maximum value from this block, whether positive or negative. 3 bits are transmitted from the binary digit to the lower order (the most significant bit indicates the sign). The positions of the transmission bits within one flock consisting of eight samples are the same, and this position is transmitted as a scale value.

At the time of decoding, the position of the truncated bit is padded with an O, and the bits higher than the transmission hint are padded with code bins 1 to 1, and a pip I- is expanded. FIG. 3 shows the situation when this quasi-instantaneous companding is applied to the difference molecule 3 CM.

For the audio signal (solid line) shown in Figure 3(a), find the difference value between adjacent samples (thin lines (#1 to #8) in Figure 3(b')), and add it to the quasi-instantaneous When applying companding, the third
This becomes the thick line in figure (b). Here, the differential data (12 bits 1 to signal), which constitutes one block of 8 samples, is subjected to quasi-instantaneous companding and compressed to 3 points. At this time, each sample value becomes ① to ① by truncating the lower P1-component.

On the other hand, on the receiving side, the compressed difference data (
When the audio signal is decoded using
A negative DC shift occurs in the decoded signal, and the signal cannot be accurately reproduced.

For this reason, a method of accumulating missing bits (DC-PCM method) has already been proposed as one method for preventing negative DC shifts. This is achieved by replenishing the transmission bits with the missing bit components caused by quasi-instantaneous companding in the form of accumulation.
This method suppresses shifts.

By the way, in the DC-PCM method, although missing bits are compensated for by acculsion, the decoded waveform always follows the lower side of the original waveform, and the error within a block is at most 1 missing bit.
− is the maximum value.

The applicant basically applies quasi-instantaneous companding to the differential PCM method, and takes into account that the decoded waveform wraps around the original waveform, and even if the error is maximum, the transmission bit quantum We have proposed an audio signal encoding method that provides relatively high quality audio with a low bit rate of less than 1/2 of the encoding width and simple processing (minimum hardware). Here, this method will be referred to as the "differential companding PCM method using optimal pi1-" and its principle will be explained below.

Quasi-instantaneous companding has the following characteristics.

(1) The quantization step of intra-block transmission bits is determined by the scale value.

(2) The value of the transmission bit at each sample point takes the smaller quantized value.

Therefore, considering (1) and (2) above, we proposed the following correction method.

FIG. 4 shows the principle of method I. Sample point #1 with sample point #O as a reference. #2. 13 differences for #3 each
Consider the case where the data is formed from bits and compressed to 3 bits by quasi-instantaneous companding. The absolute value of the difference value in this case is sample point #
] becomes the maximum, and 3-bit compressed difference data is formed based on sample #1. The scale value is the digit position of the most binary digit of the pattern to the pin 1 of #]-. Each sample #1, 32. The value #3 is replaced with data that can be expressed with this quantization width. For example, the compressed difference data of sample #1- is a value P12 (=
(010)). By the way, among the data that can be expressed with this quantization width, a value P13 that is one value larger than PI3
The data corresponding to (=(01, l)) is closer to the actual pH value of sample #1. Therefore, if this PI3 is used as the compressed difference data of sample #1, the error in the audio signal when decoding can be reduced. The error in the decoded value at this time can be suppressed to at most 1/2 of the quantization width of this compressed difference data.

Similarly, sample #2. For #3, compressed difference data that is closest to its decoded value or the value of the signal before encoding (P21 for sample #2, P31 for sample #3) may be selected. In this case, for sample #2,
The decoded value based on I-'23, which is larger than P21, is closer to P21 than the decoded value based on P22, which is smaller than P21, so the difference between PI3, which is the decoded value of sample #1, and P23 is Set the difference (=(], ], O)) to the compressed difference data. P2 for this sample #3 as well.
The difference between 3 and P31 (=(001-)) is set as compressed difference data. In this way, it is possible to form compressed difference data that has better followability with respect to the original audio signal. As an operation for this purpose, an operation of adding and subtracting the LSB to the compressed difference data is repeatedly performed so that the error between the decoded value and the true value becomes small.

By the way, when the change in the difference data is large and the sample point cannot be expressed by the above-mentioned 3 bins 1~, the maximum value (011) that can be expressed by the 3 bins 1~ is sent as a substitute. Similarly, when this happens on the negative side, the maximum negative value (
100) to be substituted. Furthermore, the occurrence of a missing histo l- between blocks is dealt with by calculating the difference at the first sample point of the next block using the reproduced value at the last sample point of the previous block.

FIG. 5 is a diagram showing the configuration of method I, in which 11 is L
PF, 12 is A/D, 13 is maximum value limiting section, 14 is quasi-instantaneous companding section, 5 is optimization processing section, ]-6 is scale value setting section, 17 is register, ]8 is integrating section , 19 is a quasi-instantaneous expansion unit, 20 is a multiplexer, and first, the input signal is 8
The data is sampled at KHz and becomes 2-bit original data, and the difference value from the data one sample before is calculated.

For example, 8 samples constitute one block of difference values, and quasi-instantaneous companding is performed. Method I is applied to the compressed difference data obtained thereby, and correction is performed for each sample point.

The optimized difference bit routine of FIG. 5 performs this processing. During transmission, a scale value is added to the beginning of one block by a multiplexer, and the corrected compressed difference data is transmitted sequentially. At the time of decoding=8-, the compressed differential data is expanded and decoded based on this scale value.

Based on the method shown in Method I, we proposed a method in which scale values and differential data are selected so that the error power of the decoded waveform within each block is minimized. This method (■) first sets two scale values for one block in addition to the scale value determined by normal quasi-instantaneous companding, and then applies method I to each scale value. In this method, the difference data with the minimum error in the block is determined, and the difference data with the minimum error power with respect to the original audio signal within the block is selected from among these difference data groups.

Here, the three scale values are the scale value normally found by quasi-instantaneous companding, and the scale value -1 and scale value +1 based on this scale value. FIG. 6 shows how the decoded waveform follows when the scale value is changed.

FIG. 7 is a diagram showing the configuration of method H, in which 21 is the LP
F, 22 is A/D, 23 is maximum value limiter, 24 is register, 25 is scale setting unit, 26 is method ■ adaptive logic, 27 is method ■ adaptive logic, 28 is method I adaptive logic, 29 is comparison logic , 30 is a selector, 31 is an integrator, and 32 is a multiplexer.The difference from method I shown in FIG. The difference is that comparison logic for selecting items is added.

This method (2) requires more complicated processing than method I, but it is thought that the error power of the decoded signal will be smaller and the voice quality will be better.

Note that the error power (RMSk) calculated by the comparison logic
is the square root of the sum of squares of the difference between the original signal in the block and the value obtained by decoding the difference data (this is referred to as RMSk)
It is defined in That is, ``d aJ the original signal at each sample point in the block.

When the value decoded from the compressed difference data for each scale value is dasXJ, it is calculated as follows.

Here, k is a subscript for the three scale values, j is a sample point, and n is the number of samples within two blocks.

The differential compression PCM method using the above-mentioned optimal differential PCM blocks the digital audio and encodes the scale value and the difference, but according to the above technique, the compression rate cannot be lower than 2 bits/sample. Also, the sound quality deteriorates significantly at that time.

1-The present invention was made in view of the above-mentioned circumstances,
In particular, in an encoding method that blocks a digital audio signal and encodes the difference value and the digit of the maximum difference value (absolute value) within the block, the deterioration of sound quality is suppressed to a certain level t.
This was done with the aim of making it possible to achieve a low bit rate of 1- (1 bit/sample or less), thereby improving the compression ratio while maintaining good sound quality.

[Structure] In order to achieve the object described in item 1, the present invention provides a method for dividing a digital audio signal into blocks and encoding a predetermined number of bits of the difference and the maximum bit digit (scale value) of the absolute value of the difference within the block. The scale value is encoded using a fully recoverable encoding method, and the difference is encoded using vector quantization (VQ
). Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a block diagram for explaining one embodiment of the present invention. In the figure, 1 is a difference calculating section which calculates the difference between digital audio signals. Reference numeral 2 denotes a difference block buffer, which stores the output of the difference calculation section 1 for one block, and outputs a predetermined number of bits from the bit digit detected by the scale value detection section 3. 3 is a scale value detection unit that detects the maximum absolute value of the output of the difference calculation unit 1 from within the block, and outputs the bit position of the maximum digit. 4 is the VQ codebook,
A codebook is selected based on the output of the scale value detection section 3. 5 is a maximum similarity detection unit which inputs a predetermined number of bits lower than the bit digit detected by the scale value detection unit 3 from the difference block buffer 2, calculates the distance to each code in the codebook 4, and calculates the minimum distance. Prints the index of the given coat.

Reference numeral 6 denotes an output selection section, and when the output of the scale value detection section 3 is larger than a predetermined value, the output of the scale value detection section 3 and the difference block buffer 2 is used as a code; otherwise, it is similar to the scale value detection section 3. The output of the degree maximum detection unit 5 is assumed to be a code.

In FIG. 1, a digital audio signal is now input,
A difference calculation section calculates the difference.

The difference is stored in the difference block buffer for each block and simultaneously input to the scale value detection section, the maximum absolute value is detected, and the maximum bit position is output as the scale value. Depending on the scale value, the difference block buffer outputs a difference value (the lower part is rounded) obtained by cutting out a predetermined number of bits from the bit position of the scale value from the stored difference value. Also, the scale value selects the codebook to be used. In FIG. 1, there are three codebooks. Codebook A is used when the scale value is smaller than a predetermined value THΔ, and codebook A is used when the scale value is greater than or equal to THA.
If smaller than 8, codebook B, scale value TH
If the value is smaller than the predetermined value THc, codebook C is selected. The maximum similarity detection unit calculates the distance between the output of the difference block buffer and each coat of the code hook, sets the coat that provides the minimum distance as the maximum similarity, and outputs its index.

The output selection section outputs the output of the difference block buffer and the scale value when the scale value is greater than or equal to THc, and otherwise outputs the index and scale value of the maximum similarity detection section.

FIGS. 2A and 2B are diagrams for specifically explaining the operating principle of the present invention. In the figures, the number of blocks is 4 samples, and the number of difference value extraction bits is 4 bits (including the sign). Now, if the input string is 5.60, 78°70, the difference value is +55.
+18. -8 is calculated. The maximum bit position 5 of the absolute value is the scale value. If a codebook as shown in the figure is selected based on this scale value, the maximum similarity detection unit calculates the distance to the output of the difference output buffer. That is, in the maximum similarity detection unit 5, the distance from the code of index O = + Tatami 1 cost 2 - 0) 2 + (-1 - 〇)': = 3.74
Distance to the code of index 1 = i"4rσy+(-1-C-3))" = 2.83 Distance to the coat of index 2 = f■ ": C tuft ()" + (2-0) 2+ (-1-3)
2=5.39 is calculated, the distance or the code with the smallest index 1 is considered to have the maximum similarity, and the index code is output. Next, since the scale value is smaller than the output selection threshold, the output selection unit 6 outputs the index and the scale value.

As is clear from the explanation in Section 2, according to the present invention, the compression ratio can be improved, especially when the scale value is small, using a small coat book, and when the scale value is large. Using a large coatbook is efficient in terms of both sound quality and compression.

Also, since VQ is not used when there are rapid changes, there is little deterioration in sound quality. Furthermore, since the small codebook is a subset of the large codebook, it has the advantage of requiring only a large codebook in terms of memory.

[Brief explanation of the drawing]

FIG. 1 is a block diagram for explaining one embodiment of the present invention, FIG. 2 is a diagram for concretely explaining the operating principle of the present invention, and FIGS. 3 to 7 are diagrams showing the prior art. FIG. 1...Difference calculation unit, 2-...Difference block buffer, 3
・Scale value detection unit, 4...Book to VQ code 1,
5. Maximum similarity detection section, 6. Output selection section. 16-Figure 6 (a) Figure 6 (C)

Claims (1)

[Claims] 1. In a method in which a digital audio signal is divided into blocks and a predetermined number of bits of the difference and the maximum bit digit (scale value) of the absolute value of the difference within the block are encoded, the scale value is encoded using a completely recoverable encoding method. An encoding method characterized by vector quantization (VQ) for the difference.