JPH0828668B2 - Audio signal encoding method - Google Patents

Description

TECHNICAL FIELD The present invention relates to a method for encoding a voice signal.

(B) Conventional technology ADPCM, vector quantization, etc. are known as audio signal encoding methods, but these methods perform approximation and truncation, so the original signal waveform is reproduced without distortion at all. You cannot do it.

Huffman coding is a high-performance coding method that can reproduce the original signal waveform. In this encoding, codewords having different code lengths are assigned according to the probability of occurrence of the codeword, and when the codeword is created, it is necessary to calculate the probability of occurrence for each frame. Delay) is required. Further, if the statistical properties of the signal when the code word is created and the statistical properties of the signal when actually encoded are different, the compression rate becomes worse. Therefore, for a signal whose statistical properties change with time, such as a voice signal, it is necessary to transmit to the decoding side the code word created at each statistical property change together with the coded information at that time. , The compression ratio becomes worse as a whole.

Adaptive arithmetic coding solves these problems. In this method, the probability of signal occurrence is calculated using the values encoded up to the present time, and the coding is performed while modifying the configuration of the encoder according to the probability of occurrence. Statistical properties change over time. It is possible to show high compression efficiency even for a thing.

In addition, since the encoder is constructed from the values encoded up to the present time, a similar encoding table can be constructed in the decoder without transmitting the encoding information, so that the compression rate does not decrease. It has the characteristics of

However, this encoding method has the drawback that it takes a lot of judgment as much as the number of code types to perform decoding, and it takes much time for decoding when the number of code types becomes large.

Further, the number of coding tables increases as the number of types of codes increases, and there is a drawback that a large memory capacity is required.

(C) Problems to be Solved by the Invention Therefore, the present invention eliminates the time loss of calculating the probability for each frame and sends the coding information every time the statistical property of the signal waveform changes. It is intended to prevent a decrease in compression rate and to encode a voice signal at high speed and with high efficiency.

(D) Means for Solving the Problems In the present invention, arithmetic coding is used. In this case, if the interval is divided into multiple values, the calculation becomes complicated. Therefore, multivalues are represented by m-ary numbers to form a sequence of m-values. That is, when the step of sampling an audio signal and the range of this sampled value are m ⁿ , this is the n-th order m
After conversion into a base number, the steps of arranging them in sequence one by one to form a sequence of m values and a step of arithmetically encoding the sequence of m values are provided, and the occurrence probability of each value of the m value is lower than that value. The arithmetic coding is performed as a conditional probability calculated based on the occurrence of a sequence and the occurrence of a homologous sequence and a low-order sequence of each value sampled in the past.

(E) Operation According to the above configuration, since the sampling value m ⁿ is converted into the n-th order m-ary number, the division of the interval and the calculation of the probability need not be m ⁿ ways but m × n ways, and can be performed at high speed. Can be encoded. Furthermore, when encoding a number, the conditional probability calculated based on the fact that a sequence of a lower order than that value and the occurrence of a sequence of lower and higher orders of each value sampled in the past are calculated. Since the encoding is performed, it is possible to prevent the compression rate from decreasing.

(F) Embodiment Prior to the description of the embodiments of the present invention, arithmetic coding will be described. The outline of arithmetic coding is as follows.

Assuming that 0 and 1 occur with a probability of 1/4 and 3/4, respectively,
A case of encoding a binary information bit string (x ₁ , x ₂ , ... X _n ) having a length n will be described. For encoding, first, the interval [0,1) on the given number line is divided into partial intervals [0,1 / 4] and [1 / 4,1] that are proportional to the probability of occurrence of the bit string. ). If the first bit x ₁ is 0, the section [0, 1/4) corresponding to 1/4 is selected, and if the first bit x ₁ is 1, the section [1/4, 1) corresponding to 3/4 is selected. Select.

If the first bit x ₁ is 1 and the interval [1 / 4,1) is selected, then the interval [1 / 4,1) is also a partial interval [1 / 4,7 / 16] that is also proportional to the occurrence probability. ) And [7 / 16,1). If the second bit x ₂ is 0, the section [1/4, 7/16) corresponding to 1/4 is selected, and if 1 is 1, the section corresponding to 3/4 [7/16, 1) Select.

In this way, the same operation is repeated for the selected section, and among the numerical values included in the finally selected section, a numerical value having the smallest number of digits when displayed in binary is selected, The bit string below the decimal point is the code word.

Therefore, in FIG. 4, for example, considering the code word of the information bit string (1,0,1, ...), the first 3 bits are 101, so the section a [x, y) is selected. X = 1/4 + 3/4 × 1/4 × 1/4 = 19/64 = 0.29685 …… 0.01011 (binary number display) y = 1/4 + 3/4 × 1/4 = 28/64 = 0.4375 …… 0.0111 (binary number display). Since it is 0.01 as long as it is included in the section a, the beginning of the code word may be encoded as 01 from the first 3 bits. In this way, it is possible to proceed with the encoding sequentially without knowing all the information bit strings.

Now, consider the case where a signal having a temporal correlation such as a voice signal is encoded. If an audio signal is sampled and the sampled value is converted into a binary bit string and arithmetically encoded, for example, even if the original audio signal has a correlation, if this is converted into a binary bit string,
Since there is no correlation between the arranged 1 and 0, the compression rate decreases. That is, in the case of binary encoding, the compression rate for one sampling is (−log ₂ P (0 | S (t)) * P
(0 | P (t)) + (-log ₂ P (1 | S (t))) * P (1 | S
It is known that the value is close to (t)), and P is 0.5.
The compression rate is the lowest when, and is the best when it is close to 0 or 1.

Therefore, in the present invention, the sampling value x (t) of the audio signal (its range is m ⁿ ) is converted into a sequence of m values (hereinafter, m = 2 for simplification of description). That is, x (t) = b ₀ (t) × 2 ⁰ + b ₁ (t) × 2 ¹ + ... + b _n (t)
As × 2 ⁿ , b _n (t) (n = 0, 1, ... N) of this sequence is arranged to form a sequence b ′ (t ′). This is shown in FIG.

Specifically, assuming that the range of x (t) is 2 ³ = 8, x
When (t) is 5,6,1, the sequence shown in FIG. 3 is created.

Then, b ′ (t ′) is the conditional probability P (0 | S (t ′)), P (1 | S
(T ')) and arithmetically coded. In this case, S (t ') is based on all signals before b' (t ') to be encoded S (t')) = b '(t'-1) b'(t'-2) .. can be ideally compressed if determined, but in this embodiment, S (t ') = b'(t'-1) b '(t '-2) b'
(T'-n-1) = b '(t'-n-2) b'(t'-n-
3) This means that in FIG. 2, for example, when calculating the probability of the bit indicated by the arrow, the calculation is based on the bit indicated by the diagonal line, and the calculation is performed by the lower bit of the bit to be encoded. is there. This is done because this lower bit contains a lot of information with respect to the bit to be coded, and the conditional probability largely deviates to 1 or 0, and the compression rate can be increased. Is.

Conditional probabilities P (0 | S (t ')), P (1 | S (t'))
The calculation method of is as follows. A memory of a predetermined size is prepared for all S (t ') states, and if b' (t ') at a certain S (t') is encoded, then b'at that time. (T ') is left in the memory for the state S (t') at that time, and the oldest memory is discarded. Then, P (0 | S (t ')), P (1 | S
(T ′)) is a value obtained by dividing the number of 0s or 1s in the memory for S (t ′) by the size of the memory. However,
The initial state of the memory is 0,1, 0,1 ... According to this method, when encoding and decoding, exactly the same probability can be obtained without changing the time and transmitting information to the probability.

For example, S (t ') = b'(t'-1) b '(t'
-2) and the size of the memory is 3. And
For example, the previous generation state of 1 and 0 for each state of S (t ′) is as follows.

The previous state of b '(t') is 0,0 and b '(t') = 0
, The conditional probability of 0 is 2/3 and the occurrence probability of 1 is 1/3. Therefore, divide the interval into 2/3 and 1/3,
The section 2/3 corresponding to 0 is selected and encoded. Furthermore, S
The top (old) 1 of the memory in the state 00 of (t ') is discarded and 0 is added. Therefore, the new memory table for the 00 state of S (t ') will be 000.

In the case of this example, since the size of the memory is as small as 3, a state occurs in which the contents of the memory table are all 0. In such a case, a predetermined probability is set for the probabilities of 0 and 1.
Assign each.

FIG. 1 is a block diagram for carrying out an encoding method and a decoding method according to the present invention. A digitized audio signal (m ⁿ ) is input to the terminal 1. As described above, the signal converter 2 converts this signal into a sequence of n-th order m-ary numbers. In circuit 3, the conditional probability is calculated. next,
The encoder 4 performs interval division on the number to be encoded with a probability based on the above calculation result, selects a predetermined interval, and encodes it.

The above coding is the same as normal arithmetic coding except that it uses conditional probabilities. For more information about arithmetic coding, see "Arithmetic coding" IBM J DEVELOP VOL 23.
NO2 P149-162.

For decoding, except using conditional probabilities,
It is performed in the same manner as the normal arithmetic code decoding. That is, the probability calculated by the circuit 6 which performs the same probability calculation as the circuit 3 is supplied to the decoder 5 to perform decoding. afterwards,
The signal converter 7 restores the bit string to the original signal and outputs it to the output terminal 8.

(G) Effect of the Invention According to the present invention described above, since the arithmetic coding is performed, the original digital audio signal can be reproduced without distortion. Also, since the probability is sequentially calculated while encoding, there is no time delay, and it is possible to cope with the time change of the probability, and highly efficient compression is possible.

Furthermore, since the number of types of encoding is reduced, the encoder / decoder becomes very simple and fast, and the reduction in compression efficiency at the expense of higher speed is small.

[Brief description of drawings]

FIG. 1 is a block diagram for carrying out the method of the present invention, FIG. 2 is a diagram showing a state in which a signal is converted into a sequence, FIG. 3 is a diagram showing a concrete example thereof, and FIG. 4 is a diagram of arithmetic coding. FIG. 2 is a signal converter, 3 is a probability calculator, 4 is an encoder

Claims (1)

[Claims] 1. A step of sampling an audio signal, and a step of converting a sampled value range into m ⁿ , converting the sampled value into an n-th order m-ary number, and sequentially arranging them in a line to form an m-valued sequence. And the step of arithmetically encoding the m-valued sequence, wherein the occurrence probability of each m-valued value is the occurrence fact of a sequence of a degree lower than that value and the homogeneity of each sampled value in the past. And a method of encoding a voice signal, characterized in that the arithmetic coding is performed as a conditional probability calculated based on the fact that a low-order sequence is generated.