JPH11305798A - Voice compressing and encoding device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To decrease the amount of arithmetic operations in the encoding of a noise excitation source by encoding (quantizing) a target signal itself for noise source information extraction instead of the conventional noise excitation source code vector search in the process for encoding of the CELP(code excited linear prediction coding system). SOLUTION: A noise source extraction part 206 consists of a target signal constitution part 301, a discrete cosine (DCT) transformation part 302, and a coefficient conversion part 303, wherein the part 303 is equipped with a coefficient selection part 304 for selecting a DCT transformation coefficient, an intensity quantization part 305 for quantizing the intensity of the DCT transformation coefficient, a bit string output part 306 which receives the processing result and outputs a bit sequence of specific length. Here, the intensity quantization part 305 encodes the positions and values of coefficients by each specific number of coefficients and the coefficient value are encoded by using only the codes of the respective coefficients and the intensity representing all the coefficients.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice compression encoding apparatus applied to an answering machine, a voice response system, a voice mail, and the like. More particularly, the present invention relates to an analog voice waveform input and converted into a digital voice signal. Thereafter, the present invention relates to an audio compression encoding apparatus which simultaneously reduces the storage memory cost and the processing cost by performing high-efficiency encoding of the digital audio signal using a predetermined low bit rate and low operation amount encoding method.

[0002]

2. Description of the Related Art In recent years, there has been a demand for practical low bit rate speech coding technology due to the expansion of channel capacity in mobile communications such as automobile telephones and the necessity of storing and transmitting enormous information in multimedia communications. Is growing.

In addition, a facsimile modem, a data
As an additional function of the modem, a voice encoding / decoding function for an answering machine is required, and it is desired to develop a low bit rate voice compression / encoding method for the encoding / decoding. It is rare.

At present, the mainstream of low-bit-rate speech compression encoding systems of 10 kbps or less is CELP (Code E).
xcited Linear Prediction
coding system). This CEL
The P method uses AR (Auto-R) for speech based on linear prediction.
This is a model-based compression encoding method based on an egressive (egressive) model.

More specifically, on the encoding side, speech is divided into units called frames, and LPCs (Linear Presets) representing a spectral envelope for each unit.
A parameter corresponding to each of a prediction coding (linear prediction) coefficient, pitch lag information representing the pitch information, noise (source) information as sound source information, and gain information corresponding to each of the pitch lag information and the noise source information is extracted. , Encoding (quantization), and storing or transmitting. The processing of the pitch lag information, the noise source information, and the gain information corresponding to each of the bitch lag information and the noise source information (such as parameter extraction and encoding) is performed on a unit called a subframe obtained by further dividing the frame. Sometimes it is done.

On the decoding side, the coded information is restored, an excitation source signal is generated by adding pitch information and gain information to noise source information, and this excitation source signal is composed of LPC coefficients. Through a linear prediction synthesis filter,
This is to obtain synthesized speech.

[0007]

However, the above-mentioned conventional CELP system has an advantage that a good voice can be obtained at a low bit rate of 10 kbps, but on the other hand, the amount of calculation in the encoding process of each parameter is reduced. There is a problem in that a large number hinders realization of real-time processing.

In particular, for encoding pitch lag information and noise source information, a synthetic speech is generated by passing each excitation source code vector stored in an excitation source codebook through a linear prediction synthesis filter. It is performed by selecting the one closest to the original sound compared to the sound,
Since many operations are required for the filter operation, it is necessary to perform a large amount of calculations to compare all excitation source code vectors stored in the excitation source codebook through a filter. It was very difficult to realize the processing.

For this reason, various improvements have been made to reduce the amount of computation. For example, instead of performing a filter operation on all the excitation source code vectors and comparing them with the original speech, the excitation source code can be approximately compared with the original speech using parameters with a relatively small amount of computation. One of the preselection methods is to narrow down the vectors to a small number.

In general, the excitation source codebook stores a number of excitation source code vectors represented by a given number of bits. However, by devising the configuration, the amount of calculation is reduced. Methods to reduce it have also been proposed. VSEL, which has the number of excitation source code vectors by the number of bits and represents the number of excitation source code vectors represented by the number of bits by the sum and difference thereof, drastically reduces the number of filter calculations.
P (Vector Sum Excited Line)
The ar Prediction coding method is one example.

The present invention has been made in view of the above. In the CELP coding process, the coding (quantization) of noise source information performed by searching for a noise excitation source code vector is performed by a noise source. It is an object of the present invention to provide a speech compression encoding apparatus in which a target signal itself for information extraction is encoded (quantized) to reduce a calculation amount in encoding a noise excitation source.

[0012]

In order to achieve the above object, a speech compression encoding apparatus according to the first aspect of the present invention comprises an A / D converter for digitizing an analog speech waveform into a digital speech signal.
D conversion means, voice coding means for coding the digital voice signal by a predetermined coding method, storage means for storing the coded digital voice signal, and taking out the stored digital voice signal An audio compression encoding apparatus comprising: audio decoding means for decoding; and D / A conversion means for converting the decoded digital audio signal into an analog audio signal.
Frame dividing means for dividing the digital audio signal into processing units called frames; spectrum envelope encoding means for extracting and encoding spectrum envelope information representing a spectrum envelope for the divided frames; Sub-frame forming means forming a processing unit called a frame; pitch information extracting means for extracting and encoding pitch information of the sub-frame; and gain information corresponding to the pitch information of the sub-frame by extracting and encoding Gain information extracting means for converting the encoded spectral envelope into noise information, and noise source information extracting means for extracting and encoding the noise source information which is the sound source information of the subframe. Spectrum envelope information decoding means for decoding information, and decoding the encoded noise source information Sound source information decoding means, pitch information decoding means for decoding the encoded pitch information, gain information decoding means for decoding the encoded gain information, the decoded noise source information, pitch information and gain Excitation source signal generation means for generating an excitation source signal from the information;
Combined signal generation means for generating a combined signal from the excitation source signal and the decoded spectral envelope information,
The noise source information extracting means for extracting a target signal for noise source information extraction, a discrete cosine transform means for transforming the extracted target signal into a discrete cosine transform coefficient sequence, and the discrete cosine transform means Coefficient conversion means for converting the discrete cosine transform coefficient sequence obtained in the above into a bit string of a predetermined length, further comprising a coefficient selecting means for selecting a discrete cosine transform coefficient from the discrete cosine transform coefficient sequence. Means, intensity quantizing means for quantizing the intensity of the discrete cosine transform coefficient selected by the coefficient selecting means, and a bit string output for outputting a bit string of a predetermined length in response to processing results of the coefficient selecting means and the intensity quantizing means. Means, and the intensity quantizing means further comprises: a discrete cosine transform coefficient selected by the coefficient selecting means.
The position of the coefficient and the coefficient value are encoded by a predetermined number, and the encoding of the coefficient value is performed using only the sign of each coefficient and the intensity representative of all the coefficients.

According to a second aspect of the present invention, in the audio compression encoding apparatus according to the first aspect, the coefficient selecting means divides the discrete cosine transform coefficient into a predetermined number of blocks. , A predetermined number of coefficients are selected from each block.

[0014]

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram showing an embodiment of a voice compression encoding apparatus according to the present invention.

FIG. 1 is a schematic configuration diagram of a speech compression encoding apparatus 100 according to the present embodiment. Speech compression encoding device 10
0 denotes an A / D conversion unit 101 as A / D conversion means for digitizing an analog audio waveform into a digital audio signal.
And a digital audio signal from the A / D conversion unit 101, and a voice encoding unit 102 as a voice encoding unit for encoding the digital voice signal by a predetermined encoding method. Storage section 103 as storage means for storing the encoded digital audio signal (encoded spectrum envelope information, pitch information, gain information, and noise source information), and audio for extracting and decoding the stored digital audio signal Speech decoding unit 1 as decoding means
And a D / A conversion unit 105 as D / A conversion means for converting the decoded digital audio signal into an analog audio signal. The A / D converter 101
Examples thereof include an A / D converter, a PC sound board, and the like. Further, examples of the D / A converter 105 include a D / A converter, a PC sound board, and the like.

FIG. 2 shows a block diagram of the speech encoding unit 102. The voice coding unit 102 converts the input digital voice signal into a predetermined number of samples (for example, 24
(0 sample), a frame composing unit 201 for outputting a frame signal and extracting the frame envelope information representing the spectrum envelope in frame units from the frame (frame signal) divided by the frame composing unit 201 Envelope extracting section 20 for encoding
2, a subframe forming unit 203 that further divides the frame divided by the frame forming unit 201 into subframe units of a predetermined number of samples (for example, 60 samples) and outputs a subframe signal, and a spectrum envelope extracting unit A pitch information extraction unit 204 that extracts and encodes pitch information (pitch lag information) from the subframes divided by the subframe configuration unit 203 using the spectrum envelope information extracted in 202, and corresponds to the pitch information of the subframe. A gain extraction unit 205 for extracting and encoding gain information;
A noise source extraction unit 206 that extracts and encodes noise source information as excitation information of the subframe from the spectrum envelope information, the subframe, the pitch information, and the gain information.

FIG. 3 is a block diagram of the noise source extracting unit 206. The noise extraction unit 206 extracts the spectrum envelope information extracted by the spectrum envelope extraction unit 202, the subframe signal output from the subframe construction unit 203, the pitch information extracted by the pitch information extraction unit 204, and the gain extraction unit 205. Using the gain information obtained,
Target signal forming unit 3 for extracting a target signal for noise source information extraction
01 and a discrete cosine transform (Discret)
e Cosine Transform: DCT
DCT transform unit 3 for obtaining a DCT coefficient sequence
02, and a coefficient transforming unit 303 that transforms a DCT coefficient sequence obtained by the DCT transform by the DCT transforming unit 302 into a bit sequence of a predetermined length.

Further, a coefficient conversion unit 303 includes a coefficient selection unit 304 for selecting a DCT coefficient from the DCT coefficient sequence, an intensity quantization unit 305 for quantizing the intensity of the DCT coefficient selected by the coefficient selection unit 304, and an intensity quantization unit 305. And a bit string output section 306 that receives the processing result of the conversion section 305 and outputs a bit string of a predetermined length.

FIG. 4 is a block diagram of the speech decoding unit 104. The audio decoding unit 104 stores
3, a spectrum envelope decoding unit 401 for receiving the digital audio signal (encoded spectrum envelope information, pitch information, gain information, and noise source information) and decoding the encoded spectrum envelope information. A pitch information decoding unit 402 for decoding the encoded pitch information, a noise source decoding unit 403 for decoding the noise source information from the encoded noise source information, a gain decoding unit 404 for decoding the encoded gain information, A speech synthesizer 4 that generates a synthesized speech from the decoded spectrum envelope information and the excitation source signal generated from the decoded pitch information, gain information, and noise source information.
05.

The operation of the above configuration will be described with reference to a schematic flowchart of the speech compression encoding apparatus 100 of this embodiment in FIG. 5 and a flowchart showing the operation procedure of the speech encoding unit in FIG. In FIG. 1, an analog audio signal (analog audio waveform) input from an analog audio input device (not shown) is converted into a digital audio signal by the A / D converter 101 (S501). Here, examples of the analog audio input device include a microphone, a CD player, a cassette deck, and the like.

Subsequently, the voice coding unit 102, which has received the digital voice signal, codes the digital voice signal by a predetermined coding method (S502). Here, the speech encoding process by the speech encoding unit 102 will be described in detail with reference to a schematic flowchart of the speech encoding unit 102 in FIG.

First, the digital audio signal is divided into a unit called a frame having a predetermined number of samples (for example, 240 samples) by the frame composing process of the frame composing unit 201 (S601). This frame is used as a frame signal in the spectral envelope extraction unit 20.
2 and output to the subframe configuration section 203.

Next, spectrum envelope information is extracted and encoded (quantized) from the frame signal by a spectrum envelope extraction process of a spectrum envelope extraction unit 202, and output to a pitch information extraction unit 204 and a noise source extraction unit 206. (S602). As the spectrum envelope information, for example,
Examples include a linear prediction coefficient based on a linear prediction analysis, a PARCOR coefficient, and an LSP coefficient. The encoding (quantization) of the spectral envelope information includes vector quantization, scalar quantization, split vector quantization, multistage vector quantization,
Alternatively, a combination of a plurality of quantizations may be used.

On the other hand, when the frame signal is input from the frame forming unit 201, the subframe forming unit 203 executes a subframe forming process to divide the frame signal into a predetermined number of samples (for example, 60 samples). Then, it is output as a subframe signal (S603). Note that the subframe may coincide with the frame depending on conditions.

The pitch information of each subframe signal is extracted and encoded by the pitch information extraction processing of the pitch information extraction section 204 using the spectrum envelope information extracted by the spectrum envelope extraction section 202 (S60).
4). Note that the pitch information may be extracted by an adaptive codebook search in the CELP method, or a method of obtaining from pitch information such as Fourier transform or wavelet transform. In the case of an adaptive codebook search, an auditory weighting filter may be used. The auditory weighting filter can be composed of linear prediction coefficients.

The pitch information extracted by the pitch information extraction unit 204 is input to the gain extraction unit 205, and gain information (gain component) is extracted and encoded by a gain extraction process (S605).

The noise source extracting section 206 includes a target signal forming section 301, a DCT transforming section 302, and a coefficient transforming section 303.
Performs a noise source extraction process (S606). More specifically, first, as shown in FIG. 3, a target signal composing unit 301 for extracting a target signal for noise source information extraction inputs a subframe signal, spectrum envelope information, pitch information, and gain information, and Construct the target signal for the extraction. At this time, a pitch component residual signal is composed of the residual signal up to the previous subframe, the pitch information extracted by the pitch information extracting section 204, and the gain information extracted by the gain extracting section 205, and further composed. After constructing a pitch component signal from the pitch component residual signal and the spectrum envelope information, a target signal for noise source information extraction can be obtained by subtracting the pitch component signal from the subframe signal. In order to obtain a pitch component signal from the pitch component residual signal and the spectrum envelope information, a method of passing the residual signal through a synthesis filter obtained from the spectrum envelope information can be used. The target signal thus configured is output to DCT transform section 302.

Next, the DCT transform unit 302 receives the target signal, performs DCT transform, and outputs a plurality of DCT coefficients (that is, a DCT coefficient sequence) obtained by the DCT transform to the coefficient transform unit 303.

In the coefficient conversion unit 303, the coefficient selection unit 304
Receives a plurality of DCT coefficients, selects a DCT coefficient from a DCT coefficient sequence, and outputs a selection result. In addition,
As a method of selecting a coefficient in the coefficient selection unit 304, a method of selecting a predetermined number from the coefficient having the largest absolute value (that is, intensity) of the amplitude of the coefficient can be cited. In addition, DC
The T coefficient sequence is divided into blocks evenly or unequally,
There is also a method in which a selection is made from each block and the sum of the numbers of the selected coefficients is set to a predetermined number.

The intensity quantization unit 305 includes a coefficient selection unit 304
And encodes the position and coefficient value (sign and strength) of the DCT coefficient selected by. At the time of encoding, position and sign are coded for each coefficient, and strength is not coded for each coefficient, but one or more values representative of all coefficients are calculated. Encode. As a method of calculating a representative value of one intensity, an input subframe signal is defined as a vector (x) having a dimension of a subframe length, a predetermined intensity is given to selected coefficients, and a zero value is given to unselected coefficients. When the signal sequence obtained by performing the inverse DCT transform on the coefficient sequence given with
There is a method of obtaining g so that x-ga | is minimized.

The bit string output unit 306 is provided for the intensity quantization unit 3
The quantization result of step 05 is input and a bit string having a predetermined bit length is output.

Returning to FIG. 5, the quantized signal (encoded digital audio signal) output from audio encoding section 102
Are stored by the storage unit 103 (S503).

Next, the quantized signal (encoded digital audio signal) stored in the storage section 103 is read and decoded by the audio decoding section 104 as necessary (S604). In speech decoding section 104, as shown in FIG. 4, spectrum envelope information is decoded by spectrum envelope decoding section 401, pitch information is decoded by pitch information decoding section 402, and noise source information is decoded by noise source decoding section 403. The gain information is decoded by the gain decoding unit 404. Here, the decoded pitch information, noise source information, and gain information constitute a residual signal (excitation source signal). The speech synthesis unit 405 generates a decoded speech (synthesized speech) as a digital speech signal from the decoded spectrum envelope information and the residual signal, and outputs the decoded speech to the D / A conversion unit 105.

Subsequently, the digital audio signal output from the audio synthesizer 405 (ie, the audio decoder 104) is converted into an analog audio signal (analog audio waveform) by the D / A converter 105 as shown in FIG. (S50
5).

The coding of the selected DCT coefficient may be performed by coding the coefficient position (ie, frequency component) and the coefficient value (code and strength) to a predetermined bit length. Since the target signal for noise source information extraction is a phase component excluding components such as the speech spectral envelope and pitch, the characteristic frequency component is more accurate than the strength of a specific frequency. If the emphasis is placed on the topological characteristics, the audible sound quality is improved. For this reason, the position and sign of the coefficient are coded for each coefficient, and for the strength, a value representative of all the coefficients is calculated and coded, thereby reducing the number of bits required for coding the strength and selecting as much. The number of coefficients can be increased.

Here, a specific description will be given with reference to FIGS. 7 (a) and 7 (b). In the figure, (a) shows a method of coding for each coefficient, and (b) shows a method of coding one value representing all coefficients. As is clear from the figure, (a) shows (a + b + c) × n bits are required, and (b) requires (a + b) × n + d bits. in this case,
Since (b) can be reduced by c × n−d bits (d can be considered to be substantially the same as c), the reduced bits are allocated to the increase in the number (n) of coefficients. Thus, it is possible to encode the phase component of the target signal with higher accuracy, and it is possible to expect improvement in sound quality as compared with (a) at the same bit rate.

When the coefficient selection unit 304 divides the DCT coefficient into a predetermined number of blocks and selects a predetermined number of coefficients from each block, as in the speech compression encoding apparatus according to the second aspect, for example, By analyzing in advance the frequency at which each coefficient is selected based on the learning data and dividing the coefficient appropriately according to the frequency, it is possible to greatly reduce the number of bits to be selected without deteriorating sound quality. .

As described above, the present embodiment uses the CELP
This is an audio compression encoding device using an audio compression encoding method belonging to audio encoding.

In the conventional CELP system, a codebook serving as an excitation source is provided, and each code vector belonging to the codebook is passed through a linear prediction filter representing a speech spectral envelope, and the result is used as a target signal for noise source information encoding. By comparison, a code that gives the closest synthesized signal is obtained. Incidentally, an auditory weighting filter can be used in this search. However, although the CELP method is a high-quality, low-bit-rate voice compression coding technique, it has a problem in that the amount of calculation in coding the noise source information is large.

On the other hand, according to the speech compression encoding apparatus of the present embodiment, in encoding the target signal for the noise source information extraction, the target signal is subjected to discrete cosine transform regardless of the codebook or filter calculation. (DCT) and the resulting D
The encoding is performed by converting the CT coefficient into a bit string of a predetermined length. As described above, the DCT coefficients are sent to the coefficient conversion unit 303, selected by the coefficient selection unit 304, passed through the intensity quantization unit 305, and output to the bit string output unit 306.
Is converted into a bit string of a predetermined length. In other words, since it does not have a codebook and does not perform codebook search using filter calculation, it is possible to maintain a high quality and a low bit rate as compared with the conventional CELP system, and to reduce the amount of operation of a speech compression code. Device becomes possible.

In the above-described embodiment, the input analog audio waveform is digitized into a digital audio signal, encoded (compressed) and stored, and the stored digital audio signal (the encoded digital audio signal) is stored. ) Is decoded, decoded, and further converted to an analog audio signal and output. However, the essence of the present invention lies in the encoding and decoding means in the audio compression encoding apparatus. Even if the coded signal is transmitted by the device of the present invention through a transmission means such as a network or a communication device and then decoded by the device of the present invention, the signal naturally belongs to the scope of the present invention. Is clear.

[0042]

As described above, according to the speech compression encoding apparatus of the present invention (claim 1), the noise source information extracting means includes a target signal extracting means for extracting a target signal, and a target signal extracting means for extracting the extracted target signal. Discrete cosine transform means for transforming into a discrete cosine transform coefficient sequence, and coefficient transforming means for transforming the discrete cosine transform coefficient sequence obtained by the discrete cosine transform means into a bit string of a predetermined length, further comprising: Coefficient selecting means for selecting a discrete cosine transform coefficient from a sequence of discrete cosine transform coefficients, intensity quantizing means for quantizing the intensity of the discrete cosine transform coefficient selected by the coefficient selecting means, and coefficient selecting means and intensity quantizing means. Bit string output means for outputting a bit string of a predetermined length in response to the processing result, and further comprising an intensity quantization means for converting the discrete cosine transform coefficient selected by the coefficient selection means to Then, the position of the coefficient and the coefficient value are coded by a predetermined number, and the coding of the coefficient value is performed using only the sign of each coefficient and only the intensity representative of all the coefficients. In the encoding process, the encoding (quantization) of the noise source information, which was performed by searching for the noise excitation source code vector, is performed by encoding (quantizing) the target signal itself for extracting the noise source information. It is possible to provide a speech compression encoding device in which the amount of calculation in encoding the excitation source is reduced.

According to a second aspect of the present invention, in the first aspect of the present invention, the coefficient selecting means divides the discrete cosine transform coefficients into a predetermined number of blocks. In order to select a predetermined number of coefficients, for example, the frequency at which each coefficient is selected based on learning data is analyzed in advance, and the coefficients are appropriately divided into blocks according to the frequency, so that the number of bits to be selected without deteriorating sound quality Can also be significantly reduced.

[Brief description of the drawings]

FIG. 1 is a schematic configuration diagram of a speech compression encoding device according to the present embodiment.

FIG. 2 is a block diagram illustrating a configuration of a speech encoding unit according to the present embodiment.

FIG. 3 is a schematic block diagram of a noise source extraction unit according to the present embodiment.

FIG. 4 is a block diagram illustrating a partial configuration of a speech decoding unit according to the present embodiment.

FIG. 5 is a schematic flowchart of the audio compression encoding apparatus according to the present embodiment.

FIG. 6 is a flowchart showing an operation procedure of a speech encoding unit according to the present embodiment.

FIG. 7 is an explanatory diagram showing the effect of reducing the number of bits required for encoding the strength and increasing the number of coefficients to be selected accordingly in the present embodiment.

[Explanation of symbols]

 REFERENCE SIGNS LIST 100 audio compression encoding apparatus 101 A / D conversion section 102 audio encoding section 103 storage section 104 audio decoding section 105 D / A conversion section 201 frame configuration section 202 spectrum envelope extraction section 203 subframe configuration section 204 pitch information extraction section 205 gain extraction unit 206 noise source extraction unit 301 target signal configuration unit 302 DCT conversion unit 303 coefficient conversion unit 304 coefficient selection unit 305 intensity quantization unit 306 bit sequence output unit 401 spectrum envelope decoding unit 402 pitch information decoding unit 403 noise source decoding unit 404 Gain decoding unit 405 Voice synthesis unit

Claims (2)

[Claims] 1. A / D conversion means for digitizing an analog audio waveform into a digital audio signal, audio encoding means for encoding the digital audio signal by a predetermined encoding method, and Storage means for storing the signal; voice decoding means for extracting and decoding the stored digital voice signal; and D / D for converting the decoded digital voice signal into an analog voice signal.
And A conversion means.
Frame encoding means for dividing the digital audio signal into processing units called frames, and spectral envelope encoding means for extracting and encoding spectrum envelope information representing a spectrum envelope for the divided frames. A subframe constructing unit that constitutes a processing unit called a subframe from the divided frames, a pitch information extracting unit that extracts and encodes pitch information of the subframe, and a subframe corresponding to the pitch information of the subframe. Gain information extraction means for extracting and encoding gain information; noise source information extraction means for extracting and encoding noise source information as excitation information of the subframe;
Wherein the speech decoding means comprises: spectrum envelope information decoding means for decoding the encoded spectrum envelope information; noise source information decoding means for decoding the encoded noise source information; Pitch information decoding means for decoding the pitch information, gain information decoding means for decoding the encoded gain information, and the decoded noise source information;
An excitation source signal generation unit configured to generate an excitation source signal from pitch information and gain information; and a synthesized signal generation unit configured to generate a synthesized signal from the excitation source signal and the decoded spectrum envelope information. Information extracting means for extracting a target signal for extracting noise source information, discrete cosine transform means for transforming the extracted target signal into a discrete cosine transform coefficient sequence, and discrete cosine transform means; Coefficient conversion means for converting the obtained discrete cosine transform coefficient sequence into a bit string of a predetermined length, further comprising: coefficient selection means for selecting a discrete cosine transform coefficient from the discrete cosine transform coefficient sequence Intensity quantizing means for quantizing the intensity of the discrete cosine transform coefficient selected by the coefficient selecting means; And a bit string output means for outputting a bit string of a predetermined length in response to the result of the processing described above, and further comprising: the intensity quantizing means, for the discrete cosine transform coefficients selected by the coefficient selecting means, Is encoded by a predetermined number, and the encoding of the coefficient values is performed using only the sign of each coefficient and only the intensity representative of all the coefficients. 2. A speech compression code according to claim 1, wherein said coefficient selecting means divides said discrete cosine transform coefficient into a predetermined number of blocks, and selects a predetermined number of coefficients from each block. Device.