JPH0728500A - Voice coder and decoder - Google Patents

Abstract

PURPOSE:To keep a narrow transmission band without deteriorating voice quality by assigning a shorter code to a code vector index of the code book of a code excitation linear prediction(CELP) coding system as the probability of generation of each code vector becomes higher. CONSTITUTION:When a conventional CELP system coder is applied, a coder 1, which has an original code book 5, is provided. Arbitrary voice signals are inputted to the coder 1 for a long time and outputted total number of vectors and the number of generation for every vector are obtained from the output codes by an all vector counter 2 and an each vector counter 3. From the total number of vectors and the number of generation for every vector, the probability of generation for every coded vector is computed. Then, the code book of each vector is obtained and a code word with the length of the book is assigned. Based on the probability of the generation of each vector, the code vector index is replaced and the coder, which has a shorter average code length, is completed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coder and a decoder, and more particularly to a highly efficient CELP system speech coder and decoder for performing information compression transmission of a speech signal.

[0002] In recent years, the development and spread of digital communication networks have been remarkable with the development of digital signal processing technology. With this, it is always required to reduce the communication cost by reducing the transmission band, and to reduce the cost. Contrary to this, there is a reduction in contradictory sound quality deterioration, and the transmission of other information by surplus bits. As a voice encoder / decoder, a more efficient transmission method needs to be studied.

[0003]

2. Description of the Related Art As a highly efficient speech coder and decoder as described above, there is a CELP (Code Excitation Linear Prediction) coding system which vector-quantizes a prediction residual signal and transmits the index of its excitation vector. Attention has been paid.

A block diagram of this CELP system is shown in FIG. 4, and in its operation, first, an input audio signal is stored in a codebook 11 for a target vector TV in which a plurality of samples are collected (vectorized). After the gain correction of each of the code vectors that have been performed, a quantized signal vector is generated by the synthesis filter 12, the difference (residual signal) is weighted by the auditory weighting filter 13, and the error evaluation is performed by the error evaluation unit 14 Done in. As a result of error evaluation performed repeatedly for all code vectors, the codebook index (code) having the smallest error is obtained and transmitted to the decoding side.

On the decoding side, the code vector indicated by the transmitted codebook index is gain-corrected,
Reproduction is performed by the synthesis filter 12 'to obtain a reproduction signal.

Further, in the CELP system, since it is necessary to notify the boundary of the code (codebook index) received by the decoder, it is necessary to transmit synchronization information between the encoder / decoder in addition to the voice code.

[0007]

In the conventional code vector, indexes of the same length (number of bits) are all assigned.

However, in the CELP type speech coder, the occurrence probability of each code vector is not uniform and there are a code vector with a high occurrence probability and a code vector with a low occurrence probability. Therefore, attention is paid to this point to improve the transmission efficiency. It is thought to be possible.

Further, as a method of transmitting the synchronization information, there are a case where the transmission band of the synchronization information is prepared outside the transmission band of the voice code and a case where it is inserted into a part of the voice code. In the latter case where information is inserted into a part of the voice code, in order to reduce the number of bits of the voice code, the search range of the code vector is narrowed to create empty bits.
This causes considerable deterioration in voice quality.

Therefore, according to the present invention, in a CELP type speech encoder and decoder, by efficiently allocating a codebook index, other information such as synchronization information can be transmitted in a reduced band without degrading speech quality. The purpose is to be able to insert.

[0011]

In the speech coder according to the present invention, a shorter code is assigned to a code vector index of a CELP type code book as the probability of occurrence of each code vector is higher. To do.

As a result, a short code (index) is used for a code vector having a high occurrence probability, and a long code (index) is used for a code vector having a low probability.
Is added, the average code length is reduced, and the transmission band can be suppressed low.

Then, by suppressing the transmission band to be low, other information such as synchronization information for synchronizing the encoder / decoder in the vacant band is transmitted, thereby affecting the voice quality. Without this, synchronization can be secured.

In this case, it is possible to improve the bit error resistance by simultaneously assigning the gray code to the code vectors having a strong correlation with each other within each code length.

Further, in the speech decoder according to the present invention,
It has a codebook to which the same index as the code vector index used in the above encoder is assigned.

[0016]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows an embodiment for generating a codebook used in a speech coder and a decoder according to the present invention, which will be described below in order.

The encoder provided with the original codebook
As an encoder to be prepared in this case, (1) bring an existing CELP type encoder, or (2) prepare a new CELP type encoder for training.

A codebook is generated for each code vector.
Find the number of times to live. This is because an arbitrary audio signal is taken out from a file prepared in advance and input to the CELP encoder 1 for a long time, the code vector selected by this encoder 1 is counted by the counter 2, and the total number of vectors is counted by the counter 3.
Is used to count the number of each vector.

This will be described more specifically with reference to FIG. 2. Gain adjustment (not shown) is performed on the target vector generated by the vectorization unit 4 by an arbitrary input speech signal and the code vector forming the codebook 5. Do)
The error determination unit 7 obtains an error from the quantized speech vector generated by the synthesis filter 6.

Then, the error determination unit 7 determines the code vector that gives the smallest of these errors. Only the index of the determined code vector is transmitted to the decoding side during normal transmission as in the CELP method shown in FIG.

However, since this is the stage before the index of each code vector is obtained, the total number of code vectors transmitted on the decoding circuit side is counted by the counter 2 and the number of times each vector is generated is counted by the counter 3. .

The total number of vectors and the number of occurrences of each vector
Therefore, the occurrence probability (p _{0 to pn} ) of each code vector
To calculate. For example, the index for each code vector in the codebook is first "00 ... 0000" to "11 ... 11".
111 ”(fixed to the length l), the number of occurrences a, b, ... N of each of these vectors is obtained and divided by the total number of occurrences (a + b + ... + n) to generate the probability p _{0 of} each vector. ~ _{Pn is} required.

Find the codebook l _i for each vector,
Assign a codeword of that length. In this case, in the field of information theory, information (code vector here) from the information source S (CELP speech coding circuit here) whose occurrence probability has a stationary distribution is (0, 1). When the binary code is used for encoding, its average code length L
When the primary entropy of the information source S is H ₁ (S), H ₁ (S) ≦ L can be satisfied and a code satisfying L <H ₁ (S) can be created.

Here, H ₁ (S) = − Σp _i log ₂ p
_i (i = 1 to M (= the number of code vectors)), and −log ₂ p _i ≦ l _i <−log ₂ p _i +1 ... The integer l _i that satisfies the formula (1) is the code of each code vector. Be long.

From this equation (1), the generation of each code vector is
Code vector index (sign) based on raw probability
Code with replacement and codebook with short average code length
Complete the vessel.

FIG. 3 shows the steps shown in and of FIG. 1 more specifically, and will be described below.

First, assuming that a codebook as shown in FIG. 3A is prepared, the number of occurrences of each index when the codebook is encoded by 1000 vectors is shown by the count number of the counter 2. It became so.

Thereafter, the probability of occurrence for each code vector (index) is calculated as shown in FIG.
It became like (b).

Then, the code length l of each code vector
_{When i} (i = 0 to n) is calculated by the above equation (1),
It becomes as shown in (c).

When the index of each code vector in FIG. 7A is modified to a new index in accordance with the code length thus obtained, as shown in FIG. 7D, the initial number of times of occurrence is the highest. The code length of the vector of index "010" becomes the minimum, and the new index becomes "0". Further, the code length of the vector of the initial index "101", which has the smallest number of occurrences, becomes the maximum, and the index becomes "101110100".

In the above case, the average code length of the original index and the average code length of the new index will be calculated.
The average code length is represented by Σ (code length of index x occurrence probability).

Average code length of original index 3 × 0.07 + 3 × 0.125 + 3 × 0.5 + 3 × 0.004 + 3 × 0.032 + 3 × 0.003
+3 x 0.25 +3 x 0.016 = 3.00

Average code length of new index 4 × 0.07 + 3 × 0.125 + 1 × 0.5 + 8 × 0.004 + 5 × 0.032 + 9 × 0.003
+2 x 0.25 +6 x 0.016 = 1.97

In this case, the code length can be reduced by about 1 bit on average.

Further, when the index generated with the above-mentioned occurrence probability is transmitted by the original index and the new index, the required transmission rate is as follows.

For example, if CELP encoding of 4 samples per code vector is performed on an input voice signal of 8 KHz, the code vector (index) generation period is 1/8 KHz × 4 samples = 0.5 ms. Each index is transmitted in a cycle.

Now, the time required to generate 1000 vectors is 0.5 ms × 1000 vectors = 500 ms.

Since each index has been transmitted each number of times within this 500 ms, the total amount is calculated. Original index 3 bits × 1000 times = 3000 bits New index 4 bits × 70 times + 3 × 125 + 1 × 500 +… +
2 x 250 + 6 x 16 = 1970 bits.

Therefore, by dividing these by 500 ms, the transmission rate can be obtained as follows. Original index 3000 bits / 500ms = 6000bps New index 1970 bits / 500ms = 3940bps

Therefore, the amount of data can be reduced by about 35%.

Further, by this means, other information such as synchronization information can be transmitted using the remaining 2060 bps.

Alternatively, since the code length of the code vector index is reduced, extra bits are prepared for another code vector and used as the index to increase the dimension (number) of the code vector. It is also possible to aim at improving the voice quality.

In order to perform the Gray coding in the above-mentioned embodiment, each code vector is classified according to the code length, the code vectors having a strong correlation with each other are found, and the codes are assigned. Bit error tolerance can be improved.

Further, the codebook obtained as described above is also prepared and used in the decoder on the receiving side, so that the signal can be reproduced.

[0045]

As described above, according to the speech encoder and decoder according to the present invention, the shorter the code is assigned to the code vector index of the codebook, the higher the probability of occurrence of each code vector. This has the effect of lowering the transmission band of the code generated from the device, and can reduce the average coding speed as a whole.

Further, by dropping the transmission band, it becomes possible to transmit other information in the vacant band. When a synchronizing signal is required between the encoder and the decoder, it is possible to eliminate unnecessary quality deterioration of the voice signal by transmitting in this vacant band.

[Brief description of drawings]

FIG. 1 is a diagram showing an embodiment of a procedure for generating a codebook used in a CELP type speech encoder and decoder according to the present invention.

FIG. 2 is a block diagram showing an example of a circuit used when executing the procedure shown in FIG.

FIG. 3 is a diagram showing in detail an embodiment of index correction.

FIG. 4 is a block diagram showing a general CELP type speech coder.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 encoder 2 total vector counter 3 each vector counter 4 target vector 5 noise codebook 6 weighting filter 7 error determination part In the figure, the same code | symbol shows the same or corresponding part.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Shuei Eguchi 3-22-8 Hakataekimae, Hakata-ku, Fukuoka City, Fukuoka City, Fujitsu Kyushu Digital Technology Co., Ltd. (72) Inventor Toshiaki Nobumoto Fukuoka City, Fukuoka Prefecture 3-22-8 Hakata Ekimae, Hakata-ku, Fujitsu Kyushu Digital Technology Co., Ltd.

Claims (4)

[Claims] 1. A CELP type speech coder, wherein a shorter code is assigned to a codebook index of a codebook as the probability of occurrence of each code vector is higher. 2. The speech coder according to claim 1, wherein other information is transmitted in a reduced transmission band. 3. A speech coder, wherein a Gray code is assigned to a code vector having a strong correlation with each other in each code length at the same time when the code is assigned. 4. The speech decoder according to claim 1, having a codebook to which the same index as the codebook index is assigned.