JPH0773249B2 - Speech encoding / decoding transmission method - Google Patents

Description

DETAILED DESCRIPTION [Overview] A codeword, which is a result of information amount compression by an encoder such as ADPCM in a transmitter, is further entropy coded by an entropy encoder to compress the information amount and then transmitted. Regarding the voice encoding / decoding transmission method in which decoding is performed on the receiving side in reverse, it is possible to encode and transmit the input voice signal in the best condition without passing through the buffer and without adversely affecting the encoder. For the purpose of providing a possible audio signal encoding / decoding transmission system, a plurality of encoders having different number of quantization levels or different quantization characteristics and number of quantization levels, and each of the plurality of encoders Corresponding to, a plurality of entropy encoders with different probability distributions are provided, and the encoder and the entropy encoder are operated in parallel, and for each frame of the input speech signal,
For the output bit rates of all of the plurality of entropy encoders, select a bit rate equal to or lower than the transmission bit rate,
Evaluate the quality of the locally decoded signal from all encoders corresponding to the selected entropy encoder and select and transmit the codeword via the best encoder and entropy encoder combination To configure. The decoding is performed in the reverse order.

[Industrial application field]

The present invention relates to a coding transmission device for compressing the information amount of a voice signal, and in particular, an entropy coder for a code word which is a result of the information amount compression of the voice signal by a coder such as ADPCM or ATC on the transmission side. The present invention relates to a voice encoding / decoding transmission system in which an information amount is further compressed and transmitted by entropy code word conversion in (1) to reversely decode at the receiving side to reproduce a voice signal.

With the spread of digital lines in recent years, there is a demand for a method of compressing the information amount of a voice signal and encoding it with high efficiency in order to effectively use the line.

In addition, when storing voice information in a voice storage or response system, etc., it is important to reduce the capacity of the storage memory, but a method of compressing the amount of information and encoding with high efficiency requires such a requirement. Can be answered.

[Conventional technology]

Conventionally, as a device for encoding a voice signal, an encoder for compressing and encoding a voice signal such as adaptive differential pulse coding (ADPCM) and adaptive transform coding (ATC), and statistics of quantized values There has been a coding transmission device that codes a voice signal with high efficiency by combining with an entropy coder that removes static redundancy.

FIG. 5 shows a conventional speech coding transmission system, in which 71 is an ADPCM encoder and 72 is an entropy encoder.

As shown in FIG. 5, in the encoding / transmission device for highly efficiently encoding the voice signal, the non-stationarity of the voice signal causes
Since the statistical properties of the output of the encoder change, the code length (bit rate) of the entropy coding result is not constant,
At times, there was a problem that the transmission bit rate was exceeded and transmission became impossible.

Conventionally, in order to solve such a problem, as shown in FIG. 5, the entropy coding result is temporarily stored in the buffer 73 and then transmitted to the line. At that time, according to the amount of data stored in the buffer 73, A buffer control method has been used in which the entropy coding result is controlled by modifying the quantization characteristic of the quantizer (encoder). This conventional method is described in the literature "M. Copperi,
“On the information rate control in entropy−code
d speech transmission systems ", CSELT Rapportitechic
i Vol. X-No. 6-DECEMBER, 1982 pp.435-439 ".

[Problems to be Solved by the Invention]

As described above, in the above-mentioned conventional technique, the buffer is used to control the entropy coding result.
As a practical matter, overflow or underflow may occur in the buffer depending on the nature of the input speech and the nature of the encoder and entropy encoder.

In the buffer control method, the shape of the quantizer in the encoder is modified to control the statistical properties of the output codeword so that the entropy becomes smaller, that is, it concentrates on a certain small number of values. Because
There is a drawback that the characteristics of the quantizer itself are not always optimum, which causes deterioration of the characteristics of the encoder.

Furthermore, if there is a transmission error, error propagation will occur in entropy decoding, so once a transmission error occurs, the previous error will also be superimposed on the subsequent decoding results. There was a problem that it affected over many samples.

SUMMARY OF THE INVENTION The present invention is intended to solve such a problem of the conventional art, and encodes and transmits an input voice signal in the best state without passing through a buffer and adversely affecting an encoder. Voice coding
It is intended to provide a decoding transmission method.

[Means for Solving the Problems]

 FIG. 1 shows the basic configuration of the present invention.

In FIG. 1 , on the transmitting side 1 , A1 to An are encoders and A11.
~ Anm is an entropy encoder, 3 is a bit rate of each output of the entropy encoder (whether or not the code length is less than or equal to the transmission bit rate, and compares the quality of the corresponding encoder to determine the optimum Coding characteristic / code length evaluation unit that selects a combination of an encoder and an entropy encoder, 4
Is a selection multiplexing unit that multiplexes and transmits the corresponding codeword and selection signal according to the selection signal from the coding characteristic / code length evaluation unit 3.

In the present invention, in order to achieve the above-mentioned object, the transmitting unit
1 is provided with a plurality of encoders A1 to An each having a different number of quantization levels or different quantization characteristics and number of quantization levels,
A plurality of entropy encoders A11 to Anm having different probability distributions are provided for each of the plurality of encoders A1 to An, and they are operated in parallel. Then, by the encoding characteristic / code length evaluation unit 3, for each frame of the input speech signal,
For all outputs of the multiple entropy encoders A11 to Anm, select those whose bit rate is less than or equal to the transmission bit rate, and all encoders corresponding to the entropy encoder that generates the selected output. A coded signal having the best quality is selected by evaluating the quality of the locally decoded signal from, and a selection signal indicating from which combination of the encoder and the entropy coder this codeword was selected. It is generated and given to the selective multiplexing unit 4. The selection multiplexing unit 4 multiplexes the selected codeword and this selection signal and sends them to the receiving side 2 via the transmission path.

On the receiving side 2 , 5 is a separation unit that separates the reception multiplexed signal, 7 is a switching unit that switches the output based on the selection signal, and B11 to Bnm are entropy encoders A11 to Anm, respectively.
, Entropy decoders with different probability distributions, B1 to Bn are the number of quantization levels corresponding to the encoders A1 to An, or a plurality of decoders with different quantization characteristics and number of quantization levels. Is.

The demultiplexing unit 5 demultiplexes the multiplexed signal received from the transmission side via the transmission path into a code word and a selection signal. The switching unit 7
Based on the selection signal from the separation unit 5, the entropy decoders B11 to Bnm and the decoders B1 to Bnm corresponding to the combination of the encoder and the entropy encoder specified on the transmission side.
Switching is performed so that the received codeword is transferred to the path formed by the combination with Bn. The entropy decoder selected in this way performs entropy decoding according to its probability distribution, and the selected decoder decodes according to the number of quantization levels or the quantization characteristic and the number of quantization levels. The audio signal is reproduced by performing the conversion.

[Action]

Since the voice signal is a non-stationary signal, the statistical properties of the code word output from the encoder are different for each frame. Therefore, the code length as a result of entropy coding differs for each frame depending on the number of quantization levels, or the quantization characteristics and the number of quantization levels, and the characteristics of the entropy encoder. Therefore, a plurality of entropy encoders having different probability distributions are provided for the plurality of encoders, and the codeword encoded by the best one of the combinations is transmitted.

At the transmitting side, the input voice signal is
Are input in parallel to the encoders A1 to An having different numbers of quantization levels or different quantization characteristics and numbers of quantization levels. The n encoders A1 to An simultaneously perform encoding, respectively, and output code words having different quantization levels or different quantization characteristics and quantization levels, and a locally decoded signal, respectively.

Since codewords are statistically biased, there is redundancy due to statistical bias. To eliminate this, a plurality of entropy encoders A11 with different probability distributions are used.
Entropy coding at ~ Anm.

Then, the coding characteristic / code length evaluation unit 3 evaluates the code length of the entropy coding result and the quality of the locally decoded signal for each frame, and evaluates the entropy coding result equal to or lower than the transmission bit rate allowed on the transmission path. Among code words having a code length, a combination of an encoder and an entropy encoder with the best characteristic quality of a locally decoded signal is selected for each frame and a selection signal indicating the combination is generated to perform selective multiplexing. It is multiplexed with the code word selected through the unit 4 and transmitted to the receiving side.

On the receiving side, a plurality of entropy decoders B11 to Bnm and decoders B1 to Bn are selected based on the selection signal indicating which encoder and entropy encoder combination has been selected.
From the above, one is selected, the entropy decoding and the decoding are performed, and the reproduced voice is output.

Therefore, unlike the prior art, without using the buffer for accumulating the entropy coding result for keeping the transmission rate constant, it is possible to always reproduce the voice with good characteristics on the receiving side. Further, since it is possible to select the optimum quantizer in each encoder, there is no deterioration of characteristics due to the deformation of the quantizer. Further, when a transmission path error occurs, the error propagation is propagated only in the frame in which the error occurs.

Further, from a combination of a plurality of encoders that differ only in the number of quantization levels and an entropy encoder with different probability distributions, the encoder and entropy encoder with the best local decoder characteristics are selected and encoded. In that case, the statistical property of the quantized signal is used only in the entropy encoder and is not considered in the encoder side, so the quantization characteristic of the encoder is not the best, and the encoding transmission device The deterioration of the characteristics will remain.

On the other hand, because the quantizer of the encoder and the entropy encoder are combined, the statistical characteristics of the quantized signal are used to make the quantization characteristic of the encoder not optimal. It is possible to prevent deterioration.

〔Example〕

2A and 2B show an embodiment of the present invention,
FIG. 2A shows a configuration example on the transmitting side, and FIG. 2B shows a configuration example on the receiving side.

In this embodiment, the transmission bit rate of the transmission line is 4 bits / sample. Therefore, as an encoder,
An encoder corresponding to 4, 5, and 6 bits of ADPCM is provided. The known Huffman (Human
ffman) encoder (hereinafter simply referred to as H encoder).

In FIG. 2A, the 5-bit ADPCM encoder 22 has
The 5-bit H encoders (A to C) 24 to 26 correspond to them, and among them, A, B, and C have different probability distributions. Similarly, the 6-bit ADPCM encoder 23 also corresponds to the 6-bit H encoders (A to C) 27 to 29, of which A,
B and C indicate that the probability distributions are different.

In the H coder, it is necessary to create a probability code in entropy coding in advance from the statistical properties of codewords. It is said that the distribution of the prediction error signal in the differential encoding such as ADPCM is well approximated by the Laplace distribution. (Reference: Hirohisa Yamaguchi, “Optimal Quantization of Laplace Distribution Signal”, IEICE Transactions, 84/2 VO1.J67-B No.2) Therefore, this distribution is assumed in this embodiment and An H coder with a near-entropy coding probability distribution is used.

The output of the 4-bit encoder 21 is directly input to the selection multiplexer 4 without passing through the H encoder. This means that if the assumed codeword probability distribution is significantly different from the actual distribution, the codewords passed through the H encoders 24 to 29 may exceed the transmission bit rate of 4 bits / sample. This is because the codeword is secured even in such a case because it is obtained. That is, the 4-bit ADPCM code word output can always be transmitted.

Further, the ADPCM encoders 21 to 23 have a decoding unit inside, and the locally decoded signal e, f, which is decoded by this decoding unit,
Input g to the coding characteristic / code length evaluation unit 3.

The coding characteristic / code length evaluation unit 3 includes a signal-to-noise ratio (S / N) calculation unit 31, and the input signal a and the locally decoded signal are included.
From e, f, g and the locally decoded signal of each ADPCM encoder 21-23
Calculate S / N. It also has a code length calculation unit 34, measures the output code length of each H encoder 24-29, and compares each output code length with the transmission bit rate in a code length comparison unit 33,
Those with a codeword output with a bit rate less than or equal to the transmission bit rate are selected.

Furthermore, it has a selection unit 32, and the S / N calculation result of the S / N calculation unit 31,
Based on the bit rate comparison result of the code length comparison unit 33, the combination of the ADPCM encoder and the H encoder having the best quality at the transmission bit rate or less is selected, and the selection signal indicating this combination is output.

In the present embodiment, since the number of quantization bits changes for each frame, in order to maintain consistency with the decoding side, the number of quantization bits and the quantization characteristics selected in the frame immediately before the first frame are Copy the internal coefficients of the ADPCM encoder. The frame length is selected so that the characteristics of the ADPCM encoder become the best when the number of quantization bits and the quantization characteristics change.

In this embodiment, it is set even when entropy coding is not performed.

In FIG. 2B, the demultiplexing unit 40 demultiplexes the code word and the selection signal from the multiplexed signal transmitted via the transmission path. The switching unit 41 responds to the selection signal output from the separating unit 40,
Switch so as to transfer the received codeword to the path consisting of the combination of the H decoder and the ADPCM decoder, which corresponds to the combination of the encoder and the H encoder selected on the transmission side. To do. Thus, the selected H decoder performs entropy decoding according to the probability distribution, and the selected decoder performs decoding according to the number of quantization levels to reproduce a voice signal.

FIG. 3 is a flow chart for explaining the operation of the present invention, and particularly shows the flow of control of the coding characteristic / code length evaluation unit 3. The operation of this embodiment will be described in detail below with reference to FIG.

(1) The audio signal a for one frame is input on the transmitting side. (S1) (2) ADPCM encoders 21 to 23 having a plurality of quantization levels
In parallel, and codewords b to d are output. Further, the code words b to d are decoded by the decoders respectively included in the ADPCM encoders 21 to 23, and the locally decoded signals e to
Output g. (S2) (3) For each of the code words b to d, the code word b is not entropy coded, and the code word c is entropy coded via the H encoders 24 to 26 having different probability distributions corresponding to 5 bits, The code word d is entropy-encoded via H encoders 27 to 29 having different probability distributions corresponding to 6 bits to generate code words b and hm. The code words b, h to m are input to the selective multiplexing unit 4 and the coding characteristic / code length evaluation unit 3. (S3) (4) The code length calculation unit 34 calculates the code length of one frame for each of the code words b, h to m. (S4) (5) The code length comparison unit 33 compares the code length calculation result of the code length calculation unit 34 with the transmission bit rate (here, 4 bits / sample) of the transmission path used in this system. Here, a line to the selective multiplexing unit 4 corresponding to a code word having a bit rate of 4 bits / sample or less is set. (S5) (6) In the S / N calculator 32, the locally decoded signals e to g from the ADPCM encoders 21 to 23 are compared with the input signal a to the ADPCM encoder to calculate their characteristics. To be done. The characteristic in this case is the signal-to-noise ratio (S / N), and the larger this value, the better the characteristic.

As a result, in the selection unit 32, among the ADPCM encoders corresponding to the ones whose lines are set in step S5, S / N
The best of is selected. The selection unit 32 displays the corresponding H
The selection signal n for selecting the signal corresponding to the encoder is output to the selection multiplexing unit 4. (S6) (7) The selection multiplexer 4 selects only the code word corresponding to the selection signal n from the input code words and multiplexes this code word and the selection signal n. (S7) (8) The multiplexed signal generated in step S7 is transmitted.
(S8) (9) On the receiving side, the transmitted codeword and the selection signal are separated by the separation unit 40, and the selection signal is switched by the switching unit 41.
Is input as a control signal. (S9) (10) The switching unit 41 inputs the codeword to the H decoder specified by the selection signal. The H-decoder entropy-decodes the code word, and the entropy-decoded signal is further decoded in the designated ADPCM decoder to obtain a reproduced voice signal. (S10) In the above embodiments, S / N is given as an example for evaluating the coding characteristic, but as another evaluation means, the sum of absolute values of error signals and the absolute value of error signals are used. It is possible to use a method such as a peak value, a cepstrum distance, a sum of squares of error signals, or a combination thereof. As an entropy encoder, H
Although the encoder has been illustrated, an arithmetic code, a run length code, a Ziv-lempel code, or the like may be used instead, and the present invention is not limited to this embodiment.

FIG. 4 shows another embodiment of the present invention. FIG. 4A shows a configuration example on the transmitting side, and FIG. 4B shows a configuration example on the receiving side. In each drawing, the same parts as those in FIGS. 2A and 2B are indicated by the same numbers.

In this embodiment, it is assumed that the transmission bit rate is about 4 bits / sample and an ADPCM encoder is used as the encoder.

A 4-bit or 5-bit ADPCM encoder is used as the encoder, and the quantization characteristic of the ADPCM encoder is assumed to be Laplace distribution or Gaussian distribution, and Lloy that minimizes the sum of squares of the error signal.
The one designed by the ed-Max method (reference: Kazuo Nakata "High-efficiency coding of speech", Morikita Publishing 1986 pp.22-23) is used.

In FIG. 4A, 51 is a 4-bit ADPCM encoder having a Laplace distribution, and 52 is a 4-bit A having a Gaussian distribution.
DPCM encoder, 53 is a 5-bit AD with Laplace distribution
PCM encoder, 54 is 5-bit ADPCM with Gaussian distribution
It is an encoder.

A Huffman encoder is used as the entropy encoder. In the Huffman encoder, since it is necessary to create a code in entropy encoding from the probability distribution of the codeword in advance, when the Laplace distribution and the Gaussian distribution are specified, the quantization characteristic of the ADPCM encoder A Huffman encoder code is generated from the probability distribution of the quantization result.

In FIG. 4A, 55 is a 5-bit Huffman encoder corresponding to the Laplace distribution, and 56 is a 5-bit Huffman encoder corresponding to the Gaussian distribution.

The S / N for each frame is used for the evaluation of the coding characteristic in the coding characteristic / code length evaluation unit 3, and the information indicating which combination of the encoder and the entropy encoder is selected as the best is assisted. The information is multiplexed and transmitted together with the codeword. In this embodiment, this information has 2 bits.

Arithmetic code, run length code, Ziv-lempel code, etc. can be used for entropy coding, and the sum of absolute value of error signal, peak value of absolute value of error signal, As in the case of the embodiment shown in FIGS. 2A and 2B, the cepstrum distance, the sum of squares of error signals, and combinations thereof can be used.

In the present embodiment, since the number of quantization bits changes for each frame, in order to maintain consistency with the decoding side, the number of quantization bits and the quantization characteristics selected in the frame immediately before the first frame are Copy the internal coefficients of the ADPCM encoder. The frame length is selected so that the characteristics of the ADPCM encoder become the best when the number of quantization bits and the quantization characteristics change.

In this embodiment, it is set even when entropy coding is not performed. This corresponds to the case where the ADPCM encoders 51 and 52 are used in FIG. 4A. This is because the average code length of entropy coding may exceed the code length of the codeword before entropy coding when the assumed codeword probability distribution and the actual probability distribution are significantly different. hand,
In such a case, the case where entropy coding is not performed is selected. This corresponds to a combination of entropy encoders with uniform distribution characteristics.

In FIG. 4B, the separation unit 40 separates the code word and the selection signal, and the switching unit 41 corresponds to the combination of the encoder and the Huffman encoder selected on the transmission side according to the selection signal. Switching is performed so that the received codeword is transferred to the path formed by the combination of the decoder and the ADPCM decoder.

In FIG. 4B, 61 is a 5-bit Laplace distribution Huffman decoder, 62 is a 5-bit Gaussian distribution Huffman decoder, 63 is a 4-bit Laplace distribution ADPCM decoder, and 64 is a 4-bit Gaussian distribution ADPCM. Decoder, 65 is a 5-bit Laplace distribution ADPCM decoder, and 66 is a 5-bit Gaussian distribution ADPCM decoder.

The selected Huffman decoder performs entropy decoding according to its probability distribution, and the selected ADPCM decoder performs decoding according to its quantization characteristic and the number of quantization levels, and reproduces a voice signal. .

It should be noted that the receiving side is also configured to allow a combination in which entropy decoding is not performed, corresponding to the case where entropy coding is not performed on the transmitting side. AD
This is equivalent to the case where the PCM decoders 63 and 64 are used.

As still another embodiment, an encoder that performs adaptive transform coding (ATC) can be adopted.

In ATC, a speech signal is windowed and cut out, and then orthogonal transform [discrete cosine transform (DCT), discrete Fourier transform (D
FT) etc.]. The transformed component is divided into subblocks, the amount of information to be assigned to each subblock is determined, and quantization is performed.

In this quantization, quantizers with multiple quantization levels and multiple quantization characteristics are operated in parallel, and the result of this quantization is entropy coded by an entropy encoder with multiple probability distributions. Select the combination that gives the shortest spectral distance from the quantized signal for each sub-block, whose length is less than the amount of information allowed for the sub-block, and multiplex and transmit the information indicating which combination was selected. .

The characteristics of the quantizer, the probability distribution of the entropy encoder, etc. may be designed in the same manner as in the case of the ADPCM encoder.

〔The invention's effect〕

As described above, according to the present invention, the number of quantization levels,
Alternatively, a plurality of encoders having different quantization characteristics and quantization levels and a plurality of entropy encoders having different probability distributions are provided on the transmission side, and the output bit rate is equal to or lower than the transmission bit rate. Since the combination of the encoder and the entropy encoder that is included is selected, the codeword by this combination is selected and transmitted, and the receiving side decodes with the same combination, the non-stationarity of the voice signal It is possible to select a combination of an encoder and an entropy encoder having the optimum number of quantum levels or quantization characteristics and the number of quantization levels such that the amount of information is less than or equal to the transmission bit rate according to It is possible to perform coding with the best decoding characteristics while removing redundancy due to statistical properties of codewords. Furthermore, when a transmission line error occurs, error propagation propagates only in the frame in which the error has occurred, so the resistance to the transmission line error is improved. Therefore, according to the present invention, a highly reliable voice encoding / decoding transmission system can be realized.

[Brief description of drawings]

FIG. 1 is a diagram showing a basic configuration of the present invention, FIGS. 2A and 2B are diagrams showing an embodiment of the present invention, FIG. 3 is a flow chart for explaining the operation of the present invention, and FIGS. 4A and 4B. Is a diagram showing another embodiment of the present invention, and FIG. 5 is a diagram showing a conventional voice coding transmission system. 1 is a transmitting side, 2 is a receiving side, 3 is an encoding characteristic / code length evaluation unit, 4 is a selective multiplexing unit, 5 is a demultiplexing unit, 7 is a switching unit, and A1 to
An is an encoder, A11 ~ Anm is an entropy encoder, B1 ~
Bn is a decoder corresponding to the encoders A1 to An, and B11 to Bnm is a decoder corresponding to the entropy encoders A11 to Anm.

Claims (6)

[Claims] 1. A voice coding transmission system for compressing an information amount of an input voice signal through an encoder and further an entropy encoder and transmitting the compressed information at a predetermined transmission bit rate. A plurality of encoders (A1 to An) having different characteristics and number of quantization levels, and a plurality of entropy encoders (A11 to Anm) having different probability distributions corresponding to each of the plurality of encoders. A plurality of encoders and a plurality of entropy encoders are operated in parallel, and the output bit rate of all of the plurality of entropy encoders is less than or equal to the transmission bit rate for each frame of an input speech signal. And the quality of the locally decoded signals from all the encoders corresponding to the extracted entropy encoder is evaluated to obtain the locally decoded signal having the best characteristics. Selecting a combination of the corresponding encoder and the entropy encoder, and multiplexing and transmitting the codeword and the signal indicating the combination through the combination of the selected encoder and the entropy encoder. Characterized voice coding transmission method. 2. A speech decoding method corresponding to a speech encoding unit for compressing an information amount of an input speech signal via an encoder and further an entropy encoder and transmitting the compressed information at a predetermined transmission bit rate. A plurality of decoders (B1 to Bn) having different levels or different quantization characteristics and quantization levels, and a plurality of entropy decoders having different probability distributions corresponding to the plurality of decoders (B1 to Bn). B11 to Bnm) and the transmitted codeword is transmitted by the corresponding entropy decoder and decoder combination according to the signal indicating the best transmitted encoder and entropy encoder combination. A voice decoding method characterized by decoding. 3. A voice encoding / transmission device adopting the voice encoding transmission system according to claim 1 and a voice decoding device adopting the voice encoding system according to claim 2. Audio encoding / decoding transmission device. 4. A voice coding transmission system for compressing an information amount of an input voice signal through an encoder and an entropy encoder and transmitting the compressed information at a predetermined transmission bit rate, which corresponds to the predetermined transmission bit rate. An encoder having the number of bits to be used, a number of quantization levels having a number of bits larger than the predetermined transmission bit rate, or a plurality of encoders having different quantization characteristics and quantization levels, and a plurality of encoders of the plurality of encoders. Providing a plurality of entropy encoders with different probability distributions corresponding to each, the encoder of the number of bits corresponding to the predetermined transmission bit rate, the plurality of encoders and a plurality of entropy encoder Outputs of all encoders that operate in parallel and have the number of bits corresponding to the predetermined transmission bit rate for each frame of the input audio signal and the plurality of entropy encoders Bit rate that is less than or equal to the transmission bit rate, and evaluates the quality of locally decoded signals from all encoders corresponding to the extracted encoders and entropy encoders to obtain the best characteristics. An encoder or a combination of an encoder and an entropy encoder corresponding to the locally decoded signal having is selected, and a codeword through the selected encoder or the combination of the encoder and the entropy encoder A speech coding transmission system characterized by multiplexing and transmitting with a signal indicating the encoder or a combination. 5. A speech decoding method corresponding to a speech coding unit for compressing an information amount of an input speech signal via an encoder and further an entropy encoder and transmitting at a predetermined transmission bit rate, A decoder having the number of bits corresponding to the transmission bit rate of, and a plurality of decoders having different numbers of quantization levels or more quantization characteristics and quantization levels than the predetermined transmission bit rate, And a plurality of entropy decoders having different probability distributions corresponding to each of the plurality of decoders, according to a signal indicating a transmitted encoder or a combination of the best encoder and the entropy encoder. A speech decoding method characterized in that a transmitted codeword is decoded by a corresponding decoder or a combination of an entropy decoder and a decoder. 6. A voice encoding / transmission device adopting the voice encoding transmission system according to claim 4, and a voice decoding device adopting the voice encoding system according to claim 5. Audio encoding / decoding transmission device.