JPH0895599A - Encoding method and decoding method of signal and encoder and decoder using the same - Google Patents

Abstract

PURPOSE: To perform fixed bit assignment and to reduce deformation. CONSTITUTION: Power normalizing parts 31 -3n obtain the average power of input signals of plural lines CH1 -CHn every frame, generate normalized signals X1 -Xn by dividing the signals by the average power and weights V1 -Vn complied with a normalized gain. The normalized signals of plural lines are combined with predetermined arrays in an array combining part 5XY and outputted as (m) signal vectors Y1 -Ym . The array combining part 5XY converts the weights from the power normalizing parts 31 -3n to (m) weighed vectors W1 -Wm by means of the same array combination as that of the normalized signals and outputs them. Signal vectors Y1 -Ym are quantized to be weighed vectors in a vector quantization part 6 using the corresponding weighed vectors W1 -Wm and corresponding quantized indices J1 -Jm and normalized gain index are outputted as the encoded result.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of collectively encoding signals of a plurality of systems, a method of decoding the code, and an encoder and a decoder using those methods.

[0002]

2. Description of the Related Art As a signal of a plurality of systems (a plurality of channels), left and right channel audio signals, multi-channel signals, a combination of an audio signal and a video signal, a signal of a single system is distributed to a plurality of systems at regular intervals. What kind of signal sequence can be obtained by dividing the obtained multiple sequence signal or multiple signal sequence obtained by dividing a single system signal into multiple frequency bands, as long as there is a power bias between the multiple signal sequences? It can also be applied to signals.

For example, as one of the typical methods for highly efficiently encoding audio signals such as voices and musical sounds, frequency domain coefficients (frequency characteristics respectively) obtained by performing time-frequency conversion on the audio signal for each frame. There is known a transform coding method in which the sample value at the frequency point of 1) is normalized by the envelope shape (spectrum outline) of the frequency characteristic and the obtained residual coefficient is vector quantized. As another typical method, a speech signal is subjected to linear prediction analysis in the time domain, and the obtained prediction coefficient is used as a filter coefficient to synthesize speech from an excitation signal with a synthesis filter so that distortion of the synthesized speech is minimized. Code-Excited L code that encodes the excitation signal with a periodic component vector and a noise component vector
The inear Prediction Coding (CELP) method is known.

1A and 1B show an encoder 10 and a decoder 50 using a conventional transform coding method, respectively. In the encoder 10, the acoustic signal A _T such as a musical tone or a voice, which is a digital sequence from the input terminal 11, is sent to the MDCT (Modi
fied Discrete Cosine Transform: A discrete discrete cosine transform is input to the unit 23, and the discrete discrete cosine transform is performed every frame unit, for example, about 16 to 64 milliseconds for music and about 10 to 20 milliseconds for speech, and the frequency domain coefficient is applied. A _F. Further, the acoustic signal A _T from the input terminal 11 is calculated by the spectrum outline calculation unit 24 to calculate the outline (envelope) of the spectrum of the input audio signal, and the outline is quantized by the quantization unit 25 to obtain the outline index. It is output as I _E , and with this quantized outline E _q , MDC
The frequency domain coefficient _AF from the T section 23 is divided by the normalization section 26 to obtain a residual coefficient X with little fluctuation. The residual coefficient X is scalar-quantized by the scalar quantizer 27, and bit allocation is performed for each frequency band so as to adapt to the frequency characteristics of the input acoustic signal. The bit allocation is performed by the bit allocation calculation unit 28, and the allocation information B
Assignment index I _B indicating a is output from the coding unit 29, also residual coefficients X are scalar quantized in accordance with the bit allocation, the quantized residual coefficients X _q is output.

In the decoder 50, the indexes I _E and I _B input as shown in FIG. 1B are reproduced by the reproducing units 62 and 63, respectively.
To obtain the out-of-spectrum outline E _q and the bit allocation information B, respectively. The reproducing unit 64 reproduces the quantized residual coefficient X _q into the residual coefficient X ′ according to the bit allocation information B. This residual coefficient X ′ is multiplied by the reproduced outline E _q in the denormalization unit 65 and denormalized to restore the frequency domain signal. This frequency domain coefficient A _F 'is the inverse MDC
The inverse transform discrete cosine transform is performed in the T section 66 to be returned to the acoustic signal _AT ′ in the time domain and output to the output terminal 51. Figure 2
A is for example the CEL shown in US Pat. No. 5,195,137
The configuration is almost the same as the speech signal encoder using P. The audio signal supplied to the input terminal 11 is subjected to linear prediction analysis by the linear prediction analysis unit 12 for each frame of a fixed length, a linear prediction coefficient α is obtained, and the linear prediction coefficient α is supplied to the LPC synthesis filter 13 as a filter coefficient. In the adaptive codebook 14, the fixed excitation vector E of the previous frame given to the synthesis filter 13 is held, and one segment of length S is cut out from the excitation vector and becomes the frame length T. It is repeatedly connected to generate an adaptive code vector (also referred to as a periodic component vector or a pitch component vector) corresponding to the speech periodic component and outputs it. The cut-out length S can be changed to output adaptive code vectors corresponding to different periodic components. A plurality of noise code vectors each having a frame length are recorded in the noise codebook 16, and when the index C is designated, the corresponding noise code vector is read out. The adaptive code vector and the random code vector from the adaptive code book 14 and the random code book 16 are respectively multiplied by multipliers 15 and 1.
7, the weighting coefficient (gain) g from the distortion calculation / search unit 21
₀ and g ₁ are multiplied, and the results of these multiplications are added by the adder 18 and given to the synthesis filter 13 as the excitation vector E to synthesize the voice signal.

First, with g ₁ = 0, the selected cut-out length S
Output of the synthesis filter 13 when excited by the adaptive code vector generated from the segment
And the input voice signal (vector) are obtained by the subtractor 19. The error vector is given a perceptual weighting by the perceptual weighting unit 20 as necessary, and then the distortion calculation unit 21 calculates the sum of squares of the elements (inter-symbol distance) as the distortion of the synthesized speech and holds it. The distortion calculation / codebook search unit 21 changes the cutout length S and repeats the above process to determine the cutout length S that minimizes the distortion. The excitation vector E generated by this is input to the synthesis filter 13 and the synthesized sound synthesized is removed from the input signal _AT by the subtractor 19 to obtain a noise component. Next, this noise component is set as the target value of the synthetic noise when the noise code vector from the noise codebook 16 is used as the excitation vector E, and the noise code vector with the minimum distortion is selected from the codebook 16 and the corresponding index is selected. Get C. From this determined random code vector, g ₁ that minimizes the distortion is calculated. The determined weights g ₀ and g ₁ are encoded by the encoding unit 22 as weight code G = (g ₀ , g ₁ ). In this way, the linear prediction coefficient α, the cutout length S, the noise code vector index C, and the weight code G determined for each frame of the input voice are output as the code corresponding to the input voice by the encoder of FIG. 2A.

As shown in FIG. 2B, the decoder sets the given linear prediction coefficient α in the linear prediction synthesis filter 52 as a filter coefficient. In accordance with the given cut-out length S and index C, the adaptive codebook 54 and the noise codebook 56 respectively output the adaptive code vector and the noise code vector as in FIG. 2A. It is multiplied by the weights g ₀ and g ₁ . These multiplication results are added to each other by the adder 58 and given to the linear prediction synthesis filter 52 as an excitation vector, and as a result, synthesized speech is output to the terminal 51.

FIG. 2A shows an example in which the segment cut out from the excitation vector of the immediately preceding frame is repeatedly generated by the adaptive codebook to generate the pitch component vector and the pitch component vector is given to the synthesis filter 13. In CELP, for example, "CODE-EXCITED LI" is used.
NEAR PREDICTION (CELP): HIGH-FREQUENCY SPEECH AT VER
Y LOW BIT RATES ", MRSchroeder, BSAtal, IEEEICA
As shown in SSP '85, pp.937-940, the excitation signal is vector-quantized so that the distortion of the synthesized speech is minimized by using a codebook that has a fixed number of waveform vectors as excitation vectors in advance. However, it is not necessary to use the adaptively changing codebook as shown in FIG. 2A. Other CELP
For example, "HIGH-QUALITY 16KB / S SPEECH CODING W
ITH A ONE-WAY DELAY LESS THAN 2MS ", Juin-Hwey Che
n, IEEE ICASSP '90, p.453, instead of LPC analysis of the input speech signal _AT in FIG. 2A to obtain a prediction integer, prediction of synthesized speech synthesized in the past is performed by LPC analysis. The coefficient may be obtained. In this method, it is not necessary to encode the prediction coefficient and give it to the decoding side.

[0009]

By the way, for example, when encoding left and right two-channel audio signals, if the left and right channel signals are fixedly encoded into 5-bit information, quantization for each signal level is performed. The error is 1/2 ⁵ . On the other hand, when the signal powers of the left and right channels are extremely deviated, the resolutions for the levels of the left and right channels are set to be the same and, for example, 8 bits are used for the signal of the high power channel and 2 bits are used for the signal of the low power channel. assign words, the encoded even quantization error (distortion) in the same amount of information can be reduced to 1/2 ^8.

However, when two sets of the configuration shown in FIG. 2A are provided and the stereo signals of the left and right channels are encoded by a predetermined constant information amount, for example, only the half channel information amount is independently encoded for both channels. Then, it is not possible to reduce the distortion by effectively using the characteristics of a signal in which a large power deviation occurs between both channels.

As a method of forming an optimum coding according to the power bias, a method of performing adaptive bit allocation between both channels with respect to the index of the random codebook is known. Many types of allocation are required to distribute the number of bits according to the gains of both channels, and a codebook of a size corresponding to that number is required. However, since the size and processing amount of the codebook increase in proportion to the power of 2 bits, it becomes unrealistic to allocate many bits. Further, if a code error occurs in the gain information, the boundary of the index is confused and a large error occurs in all reproduction vectors.

Even when two sets of the configuration shown in FIG. 1A are provided to encode a stereo signal of two channels on the left and right, a signal having a large power bias between both channels is encoded by the same amount of information. To improve the distortion, the number of bits corresponding to the power deviation in the left and right channels may be allocated to the left and right channel quantizers 25 and the bit allocation calculator 28. Bit allocation code (index
Since it is necessary to output _IB ), if the frequency band is finely divided and detailed bit allocation is performed, the efficiency of the encoder decreases. On the other hand, when the frequency band is roughly divided, it is not possible to sufficiently cope with the bias of the frequency characteristics of the input signal, so that the quantization distortion increases and the efficiency of using the redundancy of the input signal decreases. Further, if a bit error occurs in the bit allocation index I _B , the way of dividing the bit string with respect to the quantized residual coefficient X _q is confused, and a large distortion occurs in the entire reproduction residual coefficient X ′ on the decoding side. Further, as in the case of CELP, if an error occurs in the reproduction of the bit allocation code, the reproduced voice will be greatly distorted.

Although the above description has been made with respect to the two-channel audio signal, even when the audio signal and the video signal are multiplexed and transmitted, they are usually encoded with a certain amount of information and transmitted. Even if there is a bias in information between the audio and video, it is desirable to use the property effectively. For example, voice often has a silent section, and at this time, almost no information needs to be transmitted. On the other hand, even in the case of video, when inter-frame prediction is performed for information compression, the amount of information to be transmitted becomes very small when there is no motion in the video. For applications in which the amount of information for both audio and video is constant, the overall distortion can be reduced by allocating adaptive bits between both information. However, as in the case of the stereo signal described above, there are serious problems in terms of vector quantization processing and code error resistance.

An object of the present invention is, when signals of a plurality of systems are multiplexed and encoded with a small amount of information, reduce overall distortion under a constant amount of information without bit allocation between systems. It is to provide an encoding method and a decoding method therefor, and an encoder and a decoder using the same.

[0015]

The encoding method according to the first aspect of the present invention includes the following steps: (a) For each section of a predetermined length in each of the above systems, obtain the power of the signal in that section, The weight is determined based on the power, and (b) in each of the above systems, the signal in that section is normalized by the power in each section, and a normalized signal is generated.
A gain index representing the normalized gain is output, and (c) the normalized signals of the plurality of systems are combined in a predetermined array to generate at least one series of signal vectors, and (d) the plurality of The above-mentioned weights of each system are combined in the same array as the above-mentioned normalized signal to generate at least one series of weight vectors, and (e) the above-mentioned signal vector is weighted vector-quantized with the above-mentioned weight vector, and the quantization vector is represented. The quantization index is output, and (f) the quantization index and the gain index are output as at least a part of the encoding result for the signals of the plurality of systems.

The first aspect of the encoding method of the first invention
From the viewpoint of, in the step (a), for example, (a-1) the frequency-domain coefficient is obtained by time-frequency transforming the signal in the section, and (a-2) the outline of the frequency-domain coefficient is obtained. , (A-3)
Obtaining the residual coefficient by normalizing the frequency domain coefficient with its outline, including the step of determining the weight based on the power of the residual coefficient, the step (b) is the residual coefficient of the section This is a step of normalizing with the power and obtaining a normalized residual coefficient as the normalized signal.

The time-frequency transform of step (a-1) is, for example, a modified discrete cosine transform. The step (a-2) is, for example, linear prediction analysis is performed on the signal in the section to obtain a prediction coefficient, the prediction coefficient is quantized, and the quantized prediction coefficient is Fourier transformed to obtain the outline. In the step (f), the index representing the quantization of the prediction coefficient is output as information corresponding to the outline as another part of the coding result for the signal.

Second aspect of the encoding method of the first aspect
From the viewpoint of, the mapping of a plurality of types of arrays is stored in advance in the memory means, and the step (c) performs the vector quantization in the step (e) in the case of using each of the plurality of mappings. Select the mapping that minimizes the quantization distortion required during vector quantization, and in step (f) above, the quantization index that minimizes the quantization distortion and the code that represents the selected mapping are the codes described above. Output as another part of the conversion result.

According to the third aspect of the encoding method of the first invention, the signal of one of the plurality of systems determined in advance is a video signal. According to a fourth aspect of the encoding method of the first aspect of the present invention, the method further includes the following steps: (g) dividing a single-system input signal into frames of a constant period, and each frame is stationary. It is determined whether or not there is a transient frame, (h) when it is determined to be a transient frame, the input signal of the frame is divided into sections of the constant length and distributed as signals of the multiple systems, and Perform steps (a) to (f), (i) if the frame is stationary, determine the power of the input signal of the frame, determine the weight based on that power, and (j) the frame Of the single system of the input signal is normalized by the power to obtain a normalized signal, and outputs a gain index representing the normalized gain, (k) rearranges the normalized signal of the input signal Generate at least one sequence signal vector, and The weights of the input signal are rearranged in the same manner as the normalized signal of the input signal to generate at least one series of weight vectors, and (m) the same length as the signal vector generated in step (k) above is generated. Search for a codebook having a plurality of different quantized vectors corresponding to the respective quantized indexes, and for the normalized signal of step (j),
(l) Select the quantization vector that minimizes the weighted distance by the weight vector, determine the corresponding quantization index, and (n) the gain that represents the normalized gain by the power in step (i) above. The index and the quantization index determined in step (m) are output as at least a part of the coding result for the input signal in the stationary frame.

The step (i) in the fourth aspect of the encoding method of the first invention is, for example, (i-1) time-frequency converting the input signal in the frame determined to be stationary. The frequency domain coefficient is obtained by (i-2) the spectrum outline of the input signal in the frame is obtained, and (i-3) the frequency domain coefficient is normalized by the spectrum outline, and the resulting residual coefficient is obtained. , (I-4)
Predicting the rough shape of the residual coefficient from the residual coefficient of the past frame, (i-5) normalizing the residual coefficient with the predicted rough shape to generate a fine structure coefficient, including the step, Step (j) includes the step of obtaining the average power of the fine structure coefficient and outputting the normalized fine structure coefficient obtained by normalizing the fine structure coefficient with the average power as the normalized signal. ) Outputs the gain index indicating the average power and the outline index indicating the spectrum outline as another part of the encoding result.

According to the encoding method of the second aspect of the present invention, the excitation vector having the frame length selected from the codebook for each frame of the input acoustic signal is multiplied by the gain to be applied to the linear predictive synthesis filter, and the acoustic signal synthesized by this is obtained. In the CELP coding method for determining the excitation vector of the codebook and the gain for the codebook that minimizes the distance measure between the input acoustic signal and the input acoustic signal, the codebook corresponds to each index of the frame length. The integrated noise codebook in which the integrated noise vector of the number of systems is written in advance is provided with a plurality of linear prediction synthesis filters respectively corresponding to a plurality of sequences into which a plurality of acoustic signals are input. Including the steps: (a) Obtain prediction coefficients by linear prediction analysis for each frame in the above multiple systems, and calculate the corresponding linear prediction results. It is given as the filter coefficients to the filter, given as (b) above integrated the linear prediction synthesis filter excitation component vector lines respectively corresponding portions vectors of predetermined portions each codebook the integrated vector read from,
The integrated vector is selected so that the sum of the distance measures of the plurality of systems is minimized, (c) the gain for the excitation component vector corresponding to each system of the selected integrated vector is calculated, and (d) the selection is made. An index representing the integrated vector and a code representing the gain with respect to the excitation component vectors respectively corresponding to the plurality of systems of the integrated vector are output as at least a part of the encoding result for the input acoustic signals of the plurality of systems. .

The decoding method of the third invention is a decoding method of a signal in which signals of a plurality of systems are collectively vector-quantized and coded using one integrated codebook, and (a) is input for each section of a certain length. At least one integrated vector corresponding to the at least one vector quantization index is read from the integrated codebook, and (b) the elements of the integrated vector obtained in the above step (a) are rearranged into a predetermined array and a plurality of them are arranged. Generate a normalized signal of the system, (c) reproduce the normalized gain from the gain index input for each section of constant length, and denormalize the corresponding normalized signals of the multiple systems by the normalized gain. It is output as a reproduction signal of the above-mentioned plural systems.

According to the first aspect of the decoding method of the third invention, the step (c) includes, for example, the following steps: (d) Each of the sections is input to the plurality of systems. Play the spectrum outline from the outline index,
The frequency domain coefficient is obtained by multiplying the denormalized normalized signal, and (e) the frequency domain coefficient of each system obtained in step (d) is frequency-time converted for each section. It is converted into a time domain signal and output as the reproduced signals of the plurality of systems.

According to the second aspect of the decoding method of the third aspect of the invention, the step (b) is performed, for example, by an index representing an array mapping input for each section,
One of a plurality of predetermined rearrangement mappings is selected, and the rearrangement is performed according to the selected rearrangement mapping. Third Decoding Method of the Third Invention
From the viewpoint of, the input state code is decoded to determine whether it is a stationary frame or a transient frame, and if it is a transient frame, the above steps (a) to (c ) Is performed, and the reproduced signals obtained in the above-mentioned multiple systems in each section are sequentially shifted in time to be combined into a signal of a frame having an integral multiple of the above section. Perform the following steps for: (d) reading at least one integrated vector corresponding to at least one vector quantization index input for the frame from the integrated codebook, and (e) in step (d) above. The elements of the obtained integrated vector are rearranged in a predetermined array to generate a single-system normalized signal, and (f) the normalized gain is reproduced from the gain index input for the above frame, and The normalized signal of the single system is denormalized by the normalized gain and output as a reproduced signal.

According to a fourth aspect of the decoding method of the third invention, the step (c) in the predetermined one of the plurality of systems includes, for example, the denormalized normalized signal. As a difference signal, and adding the reproduction signal of the previous section to the difference signal of the current section to obtain the reproduction signal of the current section. A decoding method according to the fourth invention is C
A method for decoding a plurality of systems of acoustic signals encoded by ELP, including the following steps: (a) The corresponding integrated vector is read by referring to the integrated codebook by the input index, and the integrated vector described above is read. The partial vectors of the parts corresponding to the multiple systems are cut out and given to the corresponding systems, respectively, and (b) the gains corresponding to the multiple systems are reproduced from the input gain codes, and the corresponding partial vectors are respectively obtained. The excitation vector is generated by multiplication, and (c) in each of the above systems, the excitation vector is given to the synthesis filter in which the prediction coefficient is given as the filter coefficient to synthesize the acoustic signal and output as the reproduced acoustic signal.

The encoder according to the fifth aspect of the present invention is an encoder that collectively encodes signals of a plurality of systems, is provided for each system, and provides a signal for each section of a predetermined length. And a power normalizing means for generating a normalized signal by normalizing the signal in the section with the power to determine a weight based on the power, and outputting a gain index representing the normalized gain. , Signal sequence combination means for combining the normalized signals from the power normalization means of each of the plurality of systems in a predetermined array to generate at least one series of signal vectors, and the signal array combination means of each of the plurality of systems. Weight array combining means for combining the weights from the power normalizing means in the same array as the normalizing signal to generate at least one series of weight vectors , A vector quantization means for quantizing the signal vector with a weight vector by the weight vector and outputting a quantization index representing the quantization vector, wherein the quantization index and the gain index are the plurality of systems. Is output as at least a part of the encoding result for the signal.

According to the first aspect of the encoder of the fifth aspect of the present invention, the frequency domain coefficient is obtained by time-frequency converting the signal in the section for each of the systems, and the general shape of the frequency domain coefficient is calculated. Residual outline calculation means for obtaining and calculating the residual coefficient by normalizing the frequency domain coefficient with its outline, and weight calculating means for determining the weight based on the power of the power normalizing residual coefficient In addition, the power normalizing means is means for normalizing the residual coefficient in the section with the power to obtain a normalized residual coefficient as the normalized signal.

According to a second aspect of the encoder of the fifth aspect of the invention, a memory means for storing a plurality of different array mappings is provided, and the signal array combining means corresponds to the plurality of different array mappings. A plurality of sets of the weight array combination means and the vector quantization means are provided, and they are respectively obtained at the time of vector quantization by the plurality of sets with respect to the normalized signals from the power normalization means of a plurality of systems. Quantization distortions are given, the quantization distortions are compared, and comparing means for outputting a code representing an array mapping corresponding to the set giving the minimum quantization distortion, and in response to the code from the comparing means, A selection means for selecting and outputting a corresponding quantization index from the plurality of sets is provided, and the quantization index in which the quantization distortion is minimum, Code representing the-option arranged Mapping is outputted as another part of the coding result.

According to a third aspect of the encoder of the fifth aspect of the invention, a prediction means for generating a difference signal by performing prediction processing on the signal for each of the sections for a predetermined one of the plurality of systems. Is provided, the power normalizing means obtains the power of the differential signal, determines the weight based on the power, normalizes the differential signal with the power, and makes the normalized signal.

According to a fourth aspect of the encoder of the fifth aspect of the present invention, further: it is judged whether the input signal of a single system is stationary or transient for each frame of a constant cycle. When it is determined that the frame is a transient frame, the input signal of the frame is divided into sections of the constant length and divided as the signals of the plurality of systems, and the sub-frame division unit is determined to be stationary. Then, the power of the input signal of the frame is determined, the weight is determined based on the power, the input signal of the single system of the frame is normalized by the power to obtain a stationary frame normalized signal, and The stationary frame power normalizing means for outputting a stationary frame gain index representing the normalized gain and the stationary frame normalized signal of the input signal of the stationary frame are rearranged. At least one series of stationary frame signal array combining means for generating at least one series of stationary frame signal vectors and at least one series of stationary signals by rearranging the weights of the input signals of the stationary frames in the same manner as the normalized signals of the input signals. A stationary frame weight array combination means for generating a frame weight vector, and a stationary frame codebook having a plurality of quantization vectors having the same length as the stationary frame signal vector and corresponding to each quantization index, For the stationary frame normalized signal, select a quantization vector that minimizes the weighting distance by the stationary frame weight vector,
Stationary frame vector quantization means for determining its corresponding stationary frame quantization index, said stationary frame gain index and said stationary frame quantization index being at least a result of encoding the input signal in a stationary frame. It is output as a part.

In the encoder according to the fourth aspect of the encoder of the fifth invention, the input signal in the frame determined to be stationary is subjected to time-frequency conversion to obtain a frequency domain coefficient, and the input in the frame is input. The spectrum outline of the signal is obtained, the frequency domain coefficient is normalized by the spectrum outline, and the residual frame coefficient is calculated by the steady frame residual outline calculating means and the residual coefficient obtained in the past. Residual rough shape prediction means for predicting from the residual coefficient of the frame, residual rough shape normalization means for normalizing the residual coefficient with the predicted rough shape to generate a fine structure coefficient, and The average power is calculated, and the normalized fine structure coefficient obtained by normalizing the fine structure coefficient with the average power is output as the steady frame normalization signal. Reduction comprises means and the, outline index representing the steady frame gain index and the stationary frame spectrum envelope representing the average power of the fine structure coefficients is outputted as another part of the coding result.

The encoder according to the sixth aspect of the present invention is an encoder for encoding input acoustic signals of a plurality of systems, which is provided for each system and is a linear prediction analysis for obtaining a prediction coefficient by a linear prediction analysis for each frame. Means, and an integrated codebook means in which a plurality of integrated vectors having a length obtained by multiplying the frame length by the number of systems are written in advance corresponding to respective indexes,
When any one of the integrated vectors is selected, a partial vector of a predetermined part of the integrated vector is given to the corresponding system, respectively, is provided in each of the plurality of systems, and the linear predictive analysis means is provided. The prediction coefficient is given as a filter coefficient, a synthesis filter means for synthesizing an acoustic signal from the excitation component vector, and a portion of each integrated vector read from the integrated codebook provided in each of the plurality of systems. Weighting means for weighting the vector with each gain and giving it to the synthesizing filter means as an excitation component vector, and synthetic sound signals from the synthesizing filter means and the input sound of the plurality of systems provided in each of the plurality of systems. From the subtraction means for obtaining the difference from the signal and the subtraction means in the plurality of systems, The distance scale between the synthetic acoustic signal and the input acoustic signal is obtained from the difference, the sum of the distance scales in each system is obtained as integrated distortion, and the integrated vector that minimizes the integrated distortion is selected from the integrated codebook. And a distortion calculation / codebook search means for respectively determining a gain for the partial vector of the selected integrated vector, an index representing the selected integrated vector, and the above-mentioned index determined for the partial vector. A code indicating the gain is output as at least a part of the coding result for each frame for the input acoustic signals of the plurality of systems.

A decoder according to a seventh aspect of the invention is a decoder for a signal in which signals of a plurality of systems are collectively vector-quantized and coded using one integrated codebook, and integrated in correspondence with respective quantization indexes. The vector is written,
An integrated codebook for reading at least one integrated vector corresponding to at least one quantization index input for each fixed-length section, and an array in which elements of the integrated vector obtained from the integrated codebook are predetermined Inverse arraying means for rearranging into a plurality of systems and outputting as a normalized signal of a plurality of systems, and a normalized gain is reproduced from a gain index input for each section of a constant length, and the normalized gain corresponds to the plurality of systems by the normalized gain. Denormalizing means for denormalizing the signal and outputting it as the reproduction signals of the plurality of systems.

According to the first aspect of the decoder of the seventh aspect of the invention, further: each of the plurality of systems is provided,
Frequency domain coefficient reproducing means for reproducing a spectrum outline from the outline index input for each of the sections, and multiplying the denormalized normalized signal to obtain a frequency domain coefficient, and the frequency domain coefficient reproducing means respectively provided in the plurality of systems. And frequency-time conversion means for frequency-time converting the frequency domain coefficient of each system for each section to convert it into a time domain signal and outputting it as the reproduction signal of the plurality of systems.

According to a second aspect of the decoder of the seventh aspect of the invention, a memory means for storing a plurality of different array mappings in advance is provided, and the inverse array means is input for each section. One of the plurality of array mappings in the memory means is selected by the index representing the array mapping and the array is performed according to the selected array mapping.

According to the third aspect of the decoder of the seventh aspect of the invention, the input state code is further decoded to determine whether it is a stationary frame or a transient frame, and The reproduction by the denormalization means of the plurality of systems is executed, and the reproduction signals obtained in the plurality of systems in each section are sequentially shifted in time and combined to form a signal of a frame having an integral multiple of the section. And a constant frame integrated codebook in which an integrated vector for a stationary frame is written corresponding to a stationary frame index for a stationary frame, and at least one input quantization index. Corresponding to the above, the elements of at least one integrated vector read from the integrated codebook are rearranged into a predetermined array to generate a single-system normalized signal. The frame reverse arrangement means and the stationary frame inverse normalization means for denormalizing the single-system normalized signal by the normalized gain reproduced from the gain index input for the stationary frame and outputting it as a reproduced signal. It is provided.

In the decoder according to the third aspect of the seventh aspect of the invention, the spectrum outline reproduced from the input outline index provided in each of the plurality of systems is inversely normalized to the normalized signal. And a frequency conversion unit which is provided in each of the plurality of systems to frequency-time-convert the frequency-domain coefficient to obtain a time-domain signal and output it as the reproduction signal. And means.

In the decoder according to the third aspect of the seventh aspect of the invention, using the spectrum outline reproduced from the outline index input in the stationary frame, the frequency is calculated from the denormalized signal. Residual coefficient reproducing means for obtaining a domain coefficient, and frequency-time converting means for frequency-time converting the frequency domain coefficient to convert it into a time domain signal and outputting it as the reproduced signal.

According to a fourth aspect of the decoder of the seventh aspect of the invention, the normalized signal denormalized in a predetermined one of the plurality of systems is a difference signal, and the normalized signal is in the current section. It includes a prediction unit that adds the reproduction signal of the previous section to the difference signal to obtain the reproduction signal of the current section. A decoder according to an eighth aspect of the present invention is a decoder for a plurality of systems of acoustic signals coded by CELP, and a gain reproducing means for reproducing respective gains for a plurality of systems from a code input for each frame, and respective decoders. Corresponding to the index, an integrated codebook in which an integrated vector having a length that is a multiple of the frame length is written, and a portion of the integrated vector read from the integrated codebook that corresponds to each of the multiple systems. Weighting means for generating an excitation vector by multiplying the partial vector by the gain, and synthesis filter means provided in each of the plurality of systems, having a linear prediction coefficient as a filter coefficient, and synthesizing an acoustic signal from the excitation vector. And, including.

According to this invention, weighted vector quantization is performed by the integrated codebook as a signal vector obtained by combining signal components of a plurality of systems with weights corresponding to those power components, among the plurality of systems. So
Even if the bit allocation is not adaptively performed, the quantization distortion can be reduced by utilizing the power imbalance between the signals of a plurality of systems.
Moreover, if the signal of one system is divided into a plurality of sub-channels and the present invention is applied, the bias of the power in the time axis direction is effectively used for the reduction of the quantization distortion, and the highly efficient encoding can be realized.

[0041]

FIG. 3A is a functional block diagram of an encoder for explaining the principle of the present invention, in which n (n is an integer of 2 or more) system signals (assumed to be digital signals).
Are input to channels CH _{1 to} CH _n , respectively, and these signals are encoded. The signals of the respective systems are divided by the frame division units 4 ₁ to 4 _n into frames of a fixed length and are given to the power normalization units 3 _{1 to} 3 _n . The power normalization units 3 _{1 to} 3 _n each obtain the average power of a desired component of a signal in a frame, normalize (divide) each signal by the average power, and generate normalized signals X _{N1 to} X _Nn . To generate. At the same time, the normalized gains G _{1 to} G _n (corresponding to the average power) and the weights V _{1 to} V _n corresponding to the average power are output.

The processing of the power normalization units 3 _{1 to} 3 _n may be processing in the time domain or may be processing in the frequency domain. One value is determined for each frame of the average powers G _{1 to} G _n . On the other hand, for example, when an acoustic signal is treated as an input signal, it is common to consider the perceptual weight, so that the weights V _{1 to} V _n each represent a vector composed of a plurality of elements. Alternatively, as will be described later, when the power normalization units 3 _{1 to} 3 _n perform the processing in the frequency domain, the spectrum deviation of each signal is contributed to the weight so that the bias of the power in the frequency domain also causes the quantization distortion. Can be used for reduction, and in that case as well, weights V ₁ ~
Each V _n is treated as a vector consisting of multiple elements.
However, in the present invention, the weights V _{1 to} V _n may be values that depend only on the gains G _{1 to} G _n . In that case, each vector
The elements in V _{1 to} V _n have the same value.

The normalized signals from the power normalizers 3 _{1 to} 3 _n are combined with each other in a predetermined array in the interleaver 5, and a predetermined number m (m is an integer of 1 or more) of a signal vector Y ₁ is combined. It is output as ~ Y _m . This array combination is performed so that each signal vector Y _{1 to} Y _m includes elements corresponding to two or more different input signals. Interleaving section 5 is the same arranged in combination with the normalized signal in the interleaving unit 5 weights V ₁ ~V _n from the power normalization unit 3 ₁ to 3 _n, is output as m pieces of the weight vector W ₁ to _W-m It Therefore, the correspondence relationship between each element of the weight vectors W _{1 to} W _m and each element of the signal vectors Y _{1 to} Y _m is maintained.

As described in the above embodiment, the vector quantizer 6 has a codebook 6CB inside, and the codebook 6CB has a quantization of the same length as each signal vector corresponding to each index. The vector has been written. In the conventional vector quantization, one codebook is used for vector quantization of a signal of one system, whereas in the present invention, each quantization vector of one codebook 6CB is different from each other. Since it is used for quantizing each signal vector Y _{1 to} Y _m including elements corresponding to the signals of the system, it is called an integrated vector, and the codebook 6CB composed of a large number of integrated vectors is called an integrated codebook below.

The vector quantizer 6 calculates each signal vector Y ₁ ...
For Y _m , the integrated codebook 6CB is searched, and the weighted distance scale is minimized (that is, the quantization distortion is the smallest) by the weight vectors W _{1 to} W _m corresponding to the quantization vector. Each is determined, and the corresponding quantization indexes J _{1 to} J _m are output. According to such vector quantization in the vector quantization unit 6, in the quantization of the signal vector, the contribution of the element corresponding to the system having a large power to the value of the quantization distortion becomes large, so that the quantization distortion is minimized. By selecting the integrated vector that becomes, the sign of the quantization distortion is automatically applied to the signal of the system having the large power.

Codes G _{1 to} G _n representing gains from the encoder of FIG. 3A
And the quantization indexes J _{1 to} J _m are output as the coding results for each frame of the signals of the channels CH _{1 to} CH _n . These signs
G _{1 to} G _n and indexes J _{1 to} J _m are given to the decoder shown in FIG. 3B, and decoding is performed as follows. The quantization index 72 is given to the vector reproducing unit 72. The vector reproduction unit 72 has the same integrated codebook 7CB as the codebook 6CB provided in the vector quantization unit 6 of the encoder of FIG. 3A,
The integrated vectors Y ₁ 'to Y _m ' corresponding to the given quantization indexes J _{1 to} J _m are read from the codebook 7CB, and the inverse array unit 7
Give to 6. The inverse array unit 76 receives the given vector Y ₁ 'to Y
_By arranging all the elements of _m ′ in the reverse order to the arrangement in the combination arrangement section 5 in FIG.
Get X to X. These normalized signals X _{1 to} X _n are given to denormalization units (multipliers) 77 _{1 to} 77 _n , and gains G ₁ to
The signals of the systems CH ₁ to CH _n are decoded by being multiplied by G _n .

FIG. 4A shows the acoustic signals of the left and right channels in FIG.
As in the case of A, it is a case of encoding in the frequency domain,
Not only the power deviation of the left and right channels, but also the time axis direction
Power deviation and power deviation in the frequency axis direction.
Is used to reduce distortion due to encoding. Along the time axis
The power imbalance is due to the input acoustic signal of each channel on the left and right channels.
Divide the ram into n subframes (n is an integer of 2 or more)
By dividing it and distributing it to n subchannels
N signal sequences (sub-sequences) are obtained, and the
It is used as a bias of work. Power deviation along the frequency axis
For each sub-channel, the acoustic signal
The bias of the power of the residual coefficient of is used. Below, in FIG. 4A
The encoder will be described in detail. This encoder is
Channel terminal 11_L, 11_RDigital input given to
Divide each frame of the acoustic signal into n subframes,
Subframe division unit 3 for distributing to n subchannels
1_L, 31_RAnd the audio signal of each sub-channel
Frequency-domain residual coefficient and spectrum outline for each
Residual / Approximate calculation part 31_L1~ 31_Ln, 31_R1~ 31_RnAnd each service
Power positive to normalize the power of the residual coefficient of the subchannel
Normalization section 33_L1~ 33_Ln, 33_R1~ 33_RnAnd for each subchannel
Multiply the spectrum outline by the power normalization gain and add
Accordingly, the result obtained by further multiplying the perceptual weighting coefficient is the weighting coefficient V
Weight calculation section 34 for outputting as (vector)_L1~ 34_Ln, 34
_R1~ 34 _RnAnd the left and right subchannel residual coefficient
Residual interleaving section 35 that sorts in the order_XYAnd the left and right sub
Sort channel weighting factors in the same order as residual coefficients
Weight interleaving unit 35_VWAnd the sorted residual coefficients
And the corresponding sorted weighting factors
A vector quantizer 36 for performing vector quantization with
It is composed of

Generally, as one method of reducing the amount of information (the number of bits) required for encoding the input acoustic signal in the frequency domain, the residual characteristic is obtained by flattening the outline of the frequency characteristic (spectrum) of the input acoustic signal. There is a method of obtaining the coefficient and encoding the spectrum outline and the residual coefficient. The following two methods can be considered as the method of obtaining the residual coefficient. (a) The input signal is converted into frequency domain coefficients, the spectrum outline of the input signal is obtained, and the frequency domain coefficients are normalized by the spectrum outline to obtain residual coefficients.

(B) The input signal is processed in the time domain by the inverse filter controlled by the linear prediction coefficient to obtain the residual signal,
The residual signal is converted into frequency domain coefficients to obtain residual coefficients. In the above method (a), the following three methods are conceivable as methods for obtaining the spectrum outline of the input signal. (c) Applying the above facts, the linear prediction coefficient of the input signal is obtained by time-frequency transform (for example, Fourier transform).

(D) The frequency domain coefficient obtained by time-frequency converting the input signal is divided into a plurality of sub-bands, and the scale factor of each sub-band (for example, the average power of the sub-bands) is obtained as a spectrum outline. . (e) The linear prediction coefficient of the time domain signal obtained by inverse transforming the absolute value of the frequency domain coefficient obtained by time-frequency transforming the input signal is obtained, and the linear prediction coefficient is Fourier transformed.

The methods (c) and (e) are based on the following facts. The linear prediction coefficient represents the frequency characteristic of the input signal. Therefore, the spectrum outline of the linear prediction coefficient corresponds to the spectrum outline of the input signal. In detail,
The spectrum amplitude obtained by Fourier transforming the linear prediction coefficient is the reciprocal of the spectrum outline of the input signal.

In the embodiment of FIG. 4A according to the present invention, the above method (a) is applied to the acoustic signal sub-sequence of each sub-frame consisting of 2N samples in each sub-channel.
Any combination of (c) or (d) or (e) may be used. Here, each residual / rough shape calculation unit 32 (32 _{L1 to} 32 _Ln , 32 _{R1 to} 32) when the combination of the above methods (a) and (c) is used.
An example of any one of _Rn ) is shown in FIG. 4B. As shown in FIG. 4B, the residual / rough shape calculation unit 32 outputs a windowing unit 32A that multiplies the sub-frame acoustic signal by a desired window function, and outputs its output by M.
MDCT unit 32 for converting into a frequency domain coefficient by DCT
B, a linear prediction analysis unit 32C that obtains a linear prediction coefficient by performing a linear prediction analysis on the output of the windowing unit 32A, a quantized prediction coefficient and a quantization index (spectrum outline Quantizer 32D that outputs I _E because it corresponds to
And a shape calculating unit 32E for obtaining the spectrum outline E from the quantized prediction coefficient, and a frequency domain coefficient from the MDCT unit 32B is normalized (divided) by the spectrum outline, and a residual coefficient X is output. And a part 32F.

The sub-frame dividing unit 31 _L of FIG. 4A is provided for each frame of the left channel acoustic signal given to the terminal 11 _L of the left channel (for example, 16 to 64 m for a tone signal).
s, 10 to 20 ms in the case of a voice signal), every N samples, the N samples and the immediately preceding N samples of the 2N sample sub-sequences are overlapped and orthogonally transformed (LOT: Lapped Orthogo).
nal Transform) processing frame, which is sequentially circulated and distributed to n sub-channels, and given to residual / rough shape calculation units 32 _{L1 to} 32 _Ln . That is, in the case of n = 4, as shown in FIG. 5, the frames F ₁ , F ₂ , ...
When (row A) is sequentially input, each frame, for example, frame F ₁ is divided into four subframes F ₁₁ , F ₁₂ , F ₁₃ , F ₁₄ (row B) each of which has N samples, and each of them is divided into four subframes F ₁₁ , F ₁₂ , F ₁₃ , F ₁₄ (row B). subframe and sub-sequence of 2N samples that consist immediately preceding subframe _{{F 04, F 11},} {F 1 1, F 12}, {F 12, F 13},
Residual / rough shape calculation unit 32 of subchannels CH _{1 to} CH ₄ corresponding to {F ₁₃ , F ₁₄ } as LOT processing frames, respectively.
Give to _{1-32 4.} The same applies to the frames F ₂ , F ₃ , ... For LOT, for example, HSMalvar, "Sign
al Processing with Lapped Transform ", Artech House
Explained.

In the residual / rough shape calculation unit 32 of each sub-channel shown in FIG. 4B, a windowing unit 32A applies a time window to each given 2N-sample sub-sequence. A Hanning window is generally used as the window shape. 2N window hanging
Each sub-sequence of samples is subjected to Nth-order modified discrete cosine transform, which is a kind of orthogonal transform, in the MDCT unit 32B, and N-sample frequency domain coefficients are obtained. On the other hand, the output of the windowing unit 32A is subjected to linear prediction analysis by the linear prediction analysis unit 32C, and P
The following prediction coefficients α ₀ , ..., α _P are obtained. This prediction coefficient α
₀ , ..., α _P is used as a LSP parameter by the quantizer 32D,
Alternatively, it is converted into a k parameter and then quantized to obtain a rough index I _E corresponding to the prediction coefficient.

The outline calculation unit 32E obtains the spectrum outline of the prediction coefficients α ₀ , ..., α _P. To obtain the spectrum outline of the prediction coefficient, for example, as shown in FIG.
After quantized prediction coefficients, 4N-length sample sequences obtained by connecting (4N-P-1) 0s are subjected to discrete Fourier transform (eg, fast Fourier transform FFT), and the 2Nth power spectrum Is calculated, the even orders of this spectrum are decimated to extract the odd orders, and the square roots of the odd orders are taken. The N-point spectrum amplitude thus obtained represents the reciprocal of the spectrum outline E.

Alternatively, as shown in FIG. 6B, FFT analysis is performed on a sample sequence of length 2N in which (2N-P-1) zeros are connected after P + 1 quantized prediction coefficients, and the result is N-th order. Calculate the power spectrum of. The reciprocal of the i-th spectrum outline coefficient starting from the 0th is i = N−
Other than 1, the square roots of the i + 1-th and i-th power spectra are averaged, that is, interpolated. The inverse of the (N-1) th spectrum shape coefficient is obtained by taking the square root of the (N-1) th power spectrum.

The normalization section 32F obtains the residual coefficient X by dividing the spectrum amplitude from the MDCT section 32B for each corresponding sample and normalizing the spectrum outline thus obtained. However, what is directly obtained by performing the Fourier transform of the quantization unit prediction coefficient as described above is the reciprocal of the spectrum outline E. Therefore, the normalization unit 32F actually outputs the output of the MDCT unit 32B and the spectrum outline calculation unit. 32
The output of E (the reciprocal of the spectrum outline E) may simply be multiplied. However, for convenience in the following description,
The spectrum outline calculation unit 32E outputs the spectrum outline E.

Returning to FIG. 4A, the residual coefficient X and the spectrum outline E from the residual / rough shape calculation units 32 _{L1 to} 32 _{Ln of} the left sub-channels CH _{1 to} CH _n are the power normalization unit. 33
_{It is given to L1 to} 33 _Ln and weight calculation units 34 _{L1 to} 34 _Ln , respectively. The power normalization units 33 _{L1 to} 33 _Ln obtain the average value of the power of the residual coefficient X for each processing frame, divide the residual coefficient X by the average value of the powers, and normalize the residual coefficient X _N (N A vector consisting of samples) is obtained and given to the residual interleaving unit 35 _XY . At the same time, the average value of the power is given as a gain G to the weight calculation units 34 _{L1 to} 34 _Ln , and the indexes I _{G1 to} I _Gn representing the gain G are output. The weight calculation units 34 _{L1 to} 34 _Ln multiply the spectrum outline E by a gain G, and output the multiplication result as a weighting coefficient V (vector consisting of N elements) to be given to the weighting interleaving unit 35 _VW . If necessary, the multiplication result is further multiplied by a perceptual weight (a vector of N elements), and the multiplication result is output as the weight coefficient V. Therefore, the weighting coefficient V thus obtained corresponds to the product of the spectrum outline and the power-normalized gain G (and the product of the perceptual weight, if necessary).

As an example of the perceptual weighting, the perceptual control is performed such that a constant of about −0.6 is raised to the power of the spectrum outline, and a small value is large and a large value is small.
As another auditory control method, the SNR (Signal to Noise Ratio) required for each sample obtained by the auditory model used in the Mpeg-audio system is non-logarithmized to obtain the above-mentioned spectrum outline shape. A method of multiplying by the reciprocal may be used. In this method, from the frequency characteristics obtained by analyzing the input signal, the minimum SNR at which noise can be detected perceptually for each frequency sample is calculated by estimating the masking amount using an auditory model. This SNR is the required SNR for each sample. The technology of auditory models in mpeg-audio is described in ISO / IEC standard IS-11172-3.
Further, the auditory sense control may be omitted, and the reciprocal of the spectrum outline may be used as the weighting signal.

Right-channel subframe division unit 31
_L , residual / rough shape calculator 32 _{L1 to} 32 _Ln , power normalizer 33 _R1
~ 33 _Rn , weight calculator 34 _{R1 to} 34 _Rn also operate in the same way as the above left channel side, and residual / rough shape calculator of each sub-channel
Quantization indexes I _{PR1 to} I _PRn are output from 32 _{R1 to} 32 _Rn . Further, the power normalization units 33 _{R1 to} 33 _Rn output the gain indexes I _{GR1 to} I _GRn and the normalized residual coefficient X, the latter being given to the residual interleaving unit 35 _XY , and the weight calculation unit 34.
_The weight coefficient V is output from _{R1 to Rn} and the weight interleave unit
Given to 35 _VW .

The residual interleaving unit 35 rearranges all 2n residual coefficients (that is, the total number of samples is 2nN) of the left and right channels thus obtained for each frame of the input acoustic signal in each frame. And output as m sequences. This interleaving is made to be mixed with each other as much as possible, that is, the deviation of the temporal power, the deviation of the power between the left and right signals of the input terminals 11 _L and 11 _R and the corresponding residual coefficients, and the deviation of the power in the frequency domain. By mixing, it becomes almost uniform in one frame.

An example of such interleaving is, for example, two acoustic signals for each of the left and right channels (that is, n =
2) Sub-channels CH _L1 , CH _L2 and CH _R1 , CH _R2 are divided into signal sub-sequences. The normalized residual coefficient vector X _L1 , X _L2 obtained in the left channel side sub-channels CH _L1 , CH _L2 is {x _1,1 , x _1,2 , ..., x _{1, N} } with their frequency components x. , {X _2,1 , x _2,2 , ..., x _{2, N} }, and the normalized residual coefficient vectors X _R1 and X _R2 obtained on the right channel side sub-channels CH _R1 and CH _R2 , respectively, _These frequency components x are represented as {x _3,1 , x _3,2 , ..., x _{3, N} }, {x _4,1 , x _4,2 , ..., x _{4, N} }. As shown in FIG. 7, these residual coefficient vectors are arranged in the horizontal direction in the order of sub-channel numbers so that the elements (frequency components) of the respective vectors are in the column direction.
First to fourth of matrix arrangement of components of one sub-channel
Starting from each component x _1,1 , x _2,1 ,…, x _4,4 in the row,
Sequentially circulate in the sub-channel number direction, sequentially extract four components at positions shifted four by four in the frequency axis direction, and obtain a sequence Y ₁ = {x _1,1 , x _2,5 , x _3,9 , x _4,13 , x _1,17 }, Y ₂ = {x _2,1 , x _3,5 , x _4,9 , x _1,13 , x _2,17 }, Y ₃ = {x _3,1 , x _4,5 , x _0,9 , x _2,13 , x _3,17 },… Y _m = {x _{1, N-16} , x _{2, N-12} , x _{3, N-8} , x _{4, N-4} , x _{1, N} } is generated to obtain _m sequences Y _{1 to} Y _m . These sequences are Y ₁ = {y ₁ ¹ , y ₂ ¹ , y ₃ ¹ , y ₄ ¹ , y ₅ ¹ }, Y ₂ = {y ₁ ² , y ₂ ² , y ₃ ² , y ₄ ² , y ₅ ² }, Y ₃ = {y ₁ ³ ,, y ₂ ³ , y ₃ ³ , y ₄ ³ , y ₅ ³ },… Y _m = {y ₁ ^m , y ₂ ^m , y ₃ ^m , y ₄ ^m , y ₅ ^m }. Therefore, for example, y ₁ ¹ = x _1,1 , y ₂ ¹ = x _2,5 ,
…, Y ₁ ² = x _2,1 , y ₂ ² = x _3,5 ,…. In the example of FIG. 7, in order to obtain a vector Y ₁ , Y ₂ , Y ₃ , ... Which has 5 elements each, the element extraction position is shifted from the first element in the 4th row by 4th column by four cyclic shifts. Again, it is at position 17-20 that the fifth element of the first 16 vectors is obtained. Similarly, starting from each element of the 21st to 24th rows, the cyclic shift is repeated 4 times, and the next 16 vectors are obtained. By performing the cyclic shift in this way, 2
16 vectors are obtained every 0 rows. For example, N = 12
In the case of 8, since 128 ÷ 20 = 6, the remainder is 8, so for the last 8 lines, for example, another extraction method is used to obtain 5 respectively.
Generate a vector with two elements, and for the last two elements, only two of them make up one vector. Therefore, in this case, the total number of vectors m is 103.

2n (n = 2 in this example) weighting coefficient vectors V _L1 = {v _1,1 , v _1,2 , ... _Which are given from the weight calculating units 34 _L1 , 34 _L2 , 34 _R1 , 34 _R2 , v _{1, N} }, V _L2 = {v _2,1 , v _2,2 , ..., v _{2, N} }, V _R1 = {v _3,1 , v _3,2 , ..., v _{3, N} } , V _R2 = {v _4,1 , v _4,2 , ..., v _{4, N} } is subjected to exactly the same sort as the residual sort in the weight interleaving unit 35, and W ₁ = {v _1,1 , v _2,5 , v _3,9 , v _4,13 , v _1,17 }, W ₂ = {v _2,1 , v _3,5 , v _4,9 , v _1,13 , v _2,17 } , W ₃ = {v _3,1 , v _4,5 , v _1,9 , v _2,13 , v _3,17 },… W _m = {v _{1, N-16} , v _{2, N-12} , _{v 3, N-8, v} 4, N-4, v 1, to generate a _N}, m-number of weight coefficient sequence W ₁ to _W-m can be obtained. These sequences are also W ₁ = {w ₁ ¹ , w ₂ ¹ , w ₃ ¹ , w ₄ ¹ , w ₅ ¹ }, W ₂ = {w ₁ ² , w ₂ ² , w ₃ ² , w ₄ ² , w ₅ ² }, W ₃ = {w ₁ ³ ,, w ₂ ³ , w ₃ ³ , w ₄ ³ , w ₅ ³ },… W _m = {w ₁ ^m , w ₂ ^m , w ₃ ^m , w ₄ ^m , w ₅ ^m }. These are also provided to the vector quantizer 36.

Another example of rearrangement is shown in FIG. This example is based on the array of elements of the normalized residual coefficient vector X _L1 = {x _1,1 , x _1,2 , ..., x _{1, N} } of the left sub-channel CH _{L1 in} the case of FIG. As shown in FIG. 8, the array of N elements of each of the vectorized residual coefficient vectors X _L2 , X _R1 , and X _R2 is cyclically shifted in the frequency axis direction by the number of elements 1, 2, and 3, respectively, and arranged in a matrix. In the same manner as in the case of 7, repeating the procedure starting from each component of the 1st to 4th rows of the array while sequentially circulating four subchannels and extracting five components at positions shifted by 4 downwards. Sequence Y ₁ = {x _1,1 , x _2,6 , x _3,11 , x _4,16 , x _1,17 }, each having 5 components Y ₂ = {x _2,2 , x _{3 , 7, x 4,12, x 1,13} , x 2,18}, Y 3 = {x 3,3, x 4,8, x 1,9, x 2,14, x 3,19}, ... get the m-number of series Y ₁ ~Y _m rearranged with. The same applies to the weighting factor.

The reordering of the residual and weight coefficients as shown in the example of FIG. 7A or 8A is preferably such that the coefficients are intermixed between subchannels and in the frequency domain, as described above, and thus overlap. It is better not to do it randomly, and it is not necessary to do it regularly. For example, for each rearranged sequence Y _{1 to} Y _m , the first element of which channel should bring which element, and for each element what channel of which channel should bring. It is possible to prepare a table indicating whether or not to do so and perform the rearrangement according to this table. You may determine a rearrangement position of each element by a comparatively simple operation.

The respective sequences Y _{1 to} Y _m from the residual interleaving unit 35 _XY thus obtained are respectively weighted by the corresponding weight sequences W _{1 to} W _m in the vector quantizing unit 36. To be done. FIG. 9 shows an example of the configuration of the vector quantization unit 36. The k-th residual sequence after interleaving (the number of elements is 5 as in FIG. 7A or 8A) Y _k = {y ₁ ^k , y ₂ ^k , …, Y ₅ ^k } = {y _i ^k | i = 1 to 5} and the corresponding interleaved k-th weight sequence W _k = {w ₁ ^k , w ₂ ^k , ..., w ₅ ^k } = The case where the weighted vector quantization is performed by {w _i ^k | i = 1 to 5} is shown. In the integrated codebook 36A, various expected integrated vectors having a certain length are stored in advance corresponding to the respective indexes. The normalized residual sequence Y _k is an integrated vector read from the index j of the integrated codebook 36A.
Representing the i-th element of C (j) as c _i (j), the vectors Y _k and C
The difference y _i ^k -c _i (j) between the corresponding elements of (j) is obtained by the subtractor 36B for i = 1 to 5, and the difference is squared by the squarer 36C, and the inner product calculation unit 36E. On the other hand, each component w _i ^k of the weight vector W _k is a squarer 36D.
Squared with the difference of 2 of the difference given to the inner product calculation unit 36E.
Weighted distance measure d ^k when the dot product with the power is vector quantization
It is obtained as (j) and given to the optimum code searching unit 36F. That is, the weighted distance measure is expressed by the following equation.

D ^k (j) = Σ [w _i ^k {y _i ^k −c _i (j)} ² where Σ is an addition operator for i = 1 to 5. The codebook search unit 36F uses the index j of the code vector that minimizes the distance measure d ^k (j) obtained as described above for the code vector C (j) read from the integrated codebook 36A for all the indexes j. Is output and is output as the determined vector quantization index J _k . Similarly, weighted vector quantization is performed on all residual sequences Y _{1 to} Y _m of k = ₁ to _m , and the determined m vector quantization indexes J _{1 to} J _m are the terminals 37 of FIG. 4A. Is output to.

In FIG. 4A, each residual / rough shape calculation unit 32 _L1 ...
32 _Ln and 32 _{R1 to} 32 _Rn are not limited to the configuration example of FIG.
As shown in FIG.
The frequency domain coefficient obtained by transforming the T processing frame by the MDCT unit 32B is branched, and each sample (spectrum) thereof is divided.
The absolute value of is calculated by the absolute value calculation unit 32G, the absolute value output is inverse Fourier transformed by the inverse Fourier transform unit 32H to obtain the autocorrelation coefficient, and the time domain signal which is the autocorrelation coefficient is subjected to linear prediction analysis. linear prediction analysis in part 32C, with then outputs the prediction coefficients in the same quantization unit 32D and described for Figure 4B an index representing the quantized prediction coefficients are quantized as envelope index I _E, quantised prediction The coefficient may be given to the rough shape calculation unit 32E to calculate the spectrum rough shape V, and may be given to the normalization unit 32F to obtain the residual coefficient. Alternatively representative value SF ₁ - SF sub-band coefficients each divided by the band representative value calculating unit 32J ₁ ~32J _P frequency-domain coefficients is divided into several small band from the MDCT unit 32B as shown in FIG. 10B _P a (scale factor) is calculated respectively, and outputs the representative value SF ₁ - SF _P indexes representing the quantized scale factors are quantized by the quantization unit 32K as envelope index I _E, the quantization representative The value may be given to the normalization unit 32F as the outline V.
As each representative value, for example, the root mean square value of the coefficients in the small band can be used.

FIG. 11 shows an embodiment of a decoder for the encoder shown in FIG. 4A. The vector quantization indexes J _{1 to} J _m are input to the vector reproducing unit 72 from the input terminal 71, and the m vectors Y ₁ 'to Y _m ' corresponding to the respective indexes are reproduced, and the general shape is input from the input terminal 73. Index I _EL1 ~
I _ELn and I _{ER1 to} I _ERn are input to the outline reproduction unit 62, and the outlines E _{L1 to} E _Ln and E _{R1 to} E _Rn are reproduced respectively. That is, the prediction coefficient α for each sub-channel is reproduced for each frame by the reproducing unit 62A provided corresponding to each sub-channel,
From the reproduced prediction coefficient α, the spectrum outline calculation unit 3 in the residual / shape calculation unit 32 of the encoder shown in FIG. 4B.
The approximate shape of the frequency characteristic is calculated for each frame of each sub-channel by the approximate shape calculation unit 62B that performs the same calculation as in 2E,
The approximate shapes E _{L1 to} E _Ln and E _{R1 to} E _Rn are obtained. The gain indexes I _{GL1 to} I _GLn and I _{GR1 to} I _GRn are input to the gain reproducing unit 75 from the input terminal 74 and the normalized gains (average power) G _{L1 to} G _Ln and G _R1 G _Rn is played respectively.

The m vectors reproduced by the vector reproducing unit 71 are
Cutle Y₁'~ Y_m'Is the reverse interleave section 76_YXAnd the mark in Figure 4A
Interleaving performed by the interleaving unit 35 of the encoder
Reverse left interleaving to find the n left channel side residuals
Number X_L1~ X_LnAnd n residual coefficient X on the right channel side_R1~ X_Rn
And get. These residual coefficient vectors X_L1~ X_Ln， X_R1~ X _Rn
Denormalization unit 77_L1~ 77_Ln, 77_R1~ 77_RnCorrespond with
Playback gain G_L1~ G_Ln， G_R1~ G_RnIs multiplied, that is, the inverse normal
And each of its multiplication outputs is multiplied by a multiplier 65. _L1~ 6
Five_Ln， 65_R1~ 65_RnShape E given and reproduced in_L1~
E_Ln， E_R1~ E_RnMultiplying with gives an outline,
Each frequency domain coefficient is reproduced. These frequency ranges
MDCT section 66_L1~ 66_Ln, 66_R1~ 66_Rnso
2N samples (subframe
Twice the length: processed frame)
A time window is applied from the
Sete 78_L1~ 78_Ln, 78_R1~ 78_RnSo each processing frame
Time domain signal on each subchannel
The N samples in the latter half and the first half of the immediately preceding processing frame
N samples are overlaid and each of N samples
Corresponding left and right channel frames as sub-frame signals
Lame combiner 79_L, 79_RGiven to. For each frame
Left channel frame synthesizer 79_L Then left subcha
Subframe signal of N samples each
Sequence is sequentially shifted by N cycles (N samples).
1 frame left channel signal is reproduced and output end
Child 51_L To the right channel frame synthesis as well
Part 79_R Now, for each N subsample of the right subchannel,
The subframe signals are sequentially shifted by subframe periods.
Then, the right channel signal of one frame is played back and reproduced.
Output terminal 51_R Is output to.

In the embodiment of FIG. 4A, the residual interleaving unit 35
_XYAnd weight interleaving section 35_VWAlways the same predetermined
If you want to sort according to interleave mapping
As shown, prepare multiple sorting mappings in advance.
Select the mapping that minimizes the distortion of vector quantization.
It may be configured to do so. An example thereof is shown in FIG. 12A.
In this example, the residual interleaving part and weight
2 sets of interleave part and vector quantization part 35_XY1， 35
_VW1, 36₁ And 35_XY2， 35_VW2, 36₂ Provided these
Interleave section 35_XY1， 35_VW1And 35_XY2， 35_VW2To give to
Store different interleaved mappings
Mapping table TB₁, TB₂And two vectors
Quantizer 36₁, 36₂The magnitude of quantization distortion in
Comparator 38 and two vector quantizers 36₁, 36₂
Based on the comparison result of the comparator 38,
And a selector 39 for selecting and outputting the other output
There is. In addition, in order to simplify the figure, the subframe in FIG.
Lame division unit 31_L, 31_R, Residual / rough shape calculator 32_L1~ 32
_Ln, 32_R1~ 32_RnIs not shown, and the power normalization unit 33_L1~ 3
3 _Ln, 33_R1~ 33_RnAnd weight calculation unit 34_L1~ 34_Ln, 34_R1~ 34
_Rn33 for each_L, 33_R, 34_L, 34_RIs represented by
It

Power normalization unit 33_L, 33_RNormalization from
The residual coefficient X is calculated by the residual interleaving unit 35. _XY1, 35_XY2And that's it
Mapping table TB₁, TB₂Sorted by
Sort differently according to the mapping. same
The weight calculator 34_L, 34_RIs a weighting coefficient V from
Leave section 35_VW1, 35_VW2Each mapping table TB
₁, TB₂According to the interleave mapping shown in
Sort each. Table TB₁ Arranged according to
Replaced residual coefficient sequence and corresponding weight coefficient system
The column is the vector quantizer 36.₁ The residual coefficient series given to
Weighted vector quantization is performed. Table as well
Residual coefficient sequence sorted according to TB and corresponding
The weighting coefficient sequence to be used is the vector quantization unit 36.₂ Given to
And the weighted vector amount of the residual coefficient sequence described in FIG.
Childhood is performed. These quantizers 36₁, 36₂In
Minimum distortion obtained in vector quantization
(Minimum distance) d_i ^k Are compared with each other in the comparator 38 and selected
The unit 39 performs vector quantization according to the comparison result of the comparator 38.
Part 36₁, 36₂The vector with the smaller distortion calculated
Output index J of the quantizer₁~ J_mSelect any
Information that indicates whether the table in (I) is used (comparison result) I_STogether with
To the terminal 37.

As shown in FIG. 12A, the decoder for the encoder which selects the interleave mapping so that the quantization distortion becomes small is the decoder shown in FIG.
12A, the mapping tables TB ₁ and TB _{2 of} FIG.
Corresponding to, the mapping tables ITB ₁ and ITB ₂ for performing the rearrangement to be restored are provided, and the index I _S representing the selected mapping table is input to the terminal 71 together with the vector quantization indexes J _{1 to} J _m. By selector 8
1 is controlled, and corresponding ones of the mapping tables ITB ₁ and ITB ₂ are selected and used for rearrangement in the deinterleaving unit 76. The subsequent configuration is the same as that of FIG. 11,
Not shown. In the outline reproduction section 62 of FIG.
If only it reproduced and outputted representative value of each subband according to its envelope index I _E If envelope index I _E indicating a representative value SF ₁ - SF _P sub-band shown is input to 0B Good.

In the above description, the left and right stereo channels are input, and each channel is divided into a plurality of subframes for each frame to form a plurality of subchannels. However, a stereo four-channel or five-channel signal is used. May be input, and each may be divided into a plurality of subframes to form a multichannel signal. One input signal is separated into a low-frequency side signal and a high-frequency side signal, and the input terminal 1 of FIG.
You may input to 1 _L and 11 _R. In the above _description , the signals from the plurality of input terminals 11 _L and 11 _R are divided into a plurality of subframes for each frame, but the frame of the signal input to the plurality of terminals 11 _L and 11 _R is not divided into subframes. Residual coefficients for each are sorted by interleaving section 35 _XY and m
Alternatively, the weighting vector quantization may be performed for each of the m sequences by similarly rearranging the weighting factors generated from the spectrum outline and power for each frame in the interleaving unit 35 _VW . Alternatively, one input signal, that is, one moral signal may be input, divided into a plurality of subframes for each frame, each residual signal may be obtained, and weighted vector quantization may be performed.

In the above, conversion to the frequency domain signal is performed by MDC.
Although it is performed at T, it may be transformed into a frequency domain signal by other orthogonal transformation means except when the frame is divided into subframes. On the decoding side, the conversion is the reverse of the encoding conversion. As described above, according to the embodiments of FIGS. 4A and 11, when there is a power imbalance between channels or a power imbalance on the time axis in each frame, quantization is performed without adaptively changing bit allocation. The distortion can be reduced and no operation is required for bit allocation. Moreover, since the bit allocation information is not transmitted, it is resistant to code errors. Therefore, the decoding method of the present invention can correctly decode the decoding by the above coding method.

In the embodiment of the encoder shown in FIG. 4A, the input acoustic signal is always divided into subframes for each frame and distributed to a plurality of subchannels. Since the amount of information required for encoding is reduced by using the power bias (in other words, the encoding distortion is reduced), there is no effect when the power of the input signal is stationary. Nevertheless, the vector quantization indexes J _{1 to} J _{m of} all subchannels, the rough index I
_{EL1 to} I _ELn , I _{ER1 to} I _ERn , Normalized gain index I _GL1
It is _possible that the output of ~ I _GLn and I _GR1 ~ I _GRn always requires a larger amount of coded information than the case of coding without dividing into subchannels. Further, the processing amount for encoding may be larger than that when not divided into subchannels. An embodiment of an encoder in which this point is improved is shown in FIG. 13 for the case of the encoder for a one-channel input acoustic signal.

In FIG. 13, the subframe division unit 3
1. Residual / rough shape calculation unit 32₁~ 32_n, Power normalization unit 3
Three₁~ 33_n, Weight calculator 34₁~ 34_n, Residual interleaving
Part 35 _XY, Weight interleaving section 35_VW, Vector quantizer
In the encoder shown in FIG.
It is the same as the one with the configuration related to the tunnel side removed.
The work is the same. In this embodiment, the input acoustic signal is
Frame division unit 41 for dividing each frame, for each frame
State determination unit 42 for determining the state of the signal of the stationary frame
Residual / Approximation meter for finding residual coefficient and spectrum outline of
Arithmetic unit 32₀ , Normalize the power of residual coefficients of stationary frames
Power normalizing unit 33₀ , Spectra of stationary frame
Weight calculation unit 3 for obtaining a weight coefficient from the outline shape and the normalized gain
Four₀ , The residual error that rearranges the normalized residual coefficient into multiple series.
Interleave section 43_XY, The weighting coefficient is the same as the residual coefficient
Weight interleaving section 43 for rearranging in columns_VW, Residual subseries
Vector-quantize with a corresponding weight subsequence
Vector quantizer 44, vector quantizers 36 and 44
Selector 4 that selectively outputs the output according to the state of the frame
5A, stationary and transient frame shape indices
I_E0 And I_E1~ I_EnOutput depending on the state of the frame
Selector 45B, normalization of steady and transient frames
Gain index I_G0 And I_G1~ I_GnAccording to the condition of the frame
A selector 45C for selectively outputting the same is also provided. each
Residual / rough shape calculation unit 32₀ , 32₁~ 32_nThe configuration is the same
Yes, using whichever one shown in Figure 4B, 10A or 10B
Good.

Input digital audio signal from the input terminal 11
The signal sequence is input to the frame division unit 41, and every M samples.
And the residual / rough shape calculation unit 32 is divided into frames₀When,
It is supplied to the state determination unit 42 and the subframe division unit 31.
Be done. In the state determination unit 42, each frame for each of the above M samples
It is determined whether the signal of the system is stationary or transient. In other words
If the time variation in the frame is large, that is, the power variation or
If the spectral envelope changes suddenly, it is judged as a transient frame.
Set. So, for example, each frame is 4 minutes in time
Average power or average of each divided part
To obtain the general spectrum outline, and the power of these four parts
Change rate or spectrum change rate, and
If the rate of change of
If it is less than or equal to the value, it is determined to be stationary. In the state determination unit 42
Judgment is for each frame F₁， F₂, ..., but is done every time, for example
As shown in FIG. 14, a unit for performing MDCT (processing frame
2M service of each frame and the frame immediately before it.
Whether the sample signal sequence is stationary or transient, M samples
It is performed while shifting by (one frame). For example
Arm F₂And that frame and the previous frame F₁2M consisting of
When the processing frame of a sample is determined to be stationary, its
The window function WF is applied to the 2M sample, and the residual /
Outline calculator 32 ₀ At MDCT is performed. Also an example
Frame F_FourAnd that frame and the previous frame F₃From
2M sample processing frames are determined to be transient
And the M samples in the center of the 2M samples
n (n = 4 in this example) subframes SF₄₁~SCIENCE FICTION₄₄To
It is divided, and each of those subframes is the residual / rough shape calculation unit.
32₁~ 32_FourSubframe SF immediately before₃₄,SCIENCE FICTION
₄₁,SCIENCE FICTION₄₂,SCIENCE FICTION₄₃2M / n samples combined with
MDCT in units of 2 subframes and its window function
WS_S As shown in FIG.
It is something that is worth. Also, before and after dividing into subframes,
The boundaries between stationary and transient frames are
As you can see, the window function WF as shown in FIG._A , That is,
On the transient side, the window function WF_S Each half, and the steady side is WF_S of
Maximum value. These window functions WF_S， WF_AIs an example,
Other methods may be used.

If it is determined to be stationary, the residual / rough shape calculation unit 42 ₀ calculates and outputs the spectrum rough shape E _{0 of the} frame and the residual coefficient X ₀ in the frequency domain and outputs the spectrum rough shape. The rough index I _E0 calculated at the time of calculation is output. That is, for example, as the residual / rough shape calculation unit 42 ₀ shown in FIG.
When the one shown in FIG. 2 is used, the windowing unit 32A applies a time window to the processing frame of 2M samples obtained by adding the immediately preceding M samples for each input M sample, and the MDCT is performed.
The unit 32B converts the frequency domain coefficient. In addition, the linear prediction analysis unit 32C, the prediction coefficient quantization unit 32D, and the spectrum outline calculation unit 32E determine the spectrum outline E and output the outline index I _E0 .

Residual / rough shape calculation unit 32₀The residual coefficient from is
Power normalization unit 33₀ With the average power in the processing frame is positive
Normalized and its normalized residual coefficient X_N0 Residual interleaving
Part 43 _XYGive to. Its normalized gain G₀Is the weight calculator 34₀ 
Given to the spectrum outline E₀Is multiplied by the weighting factor
V₀And the weight interleaving unit 43_VWGiven to. Residual error
Coefficient X_N0 Is the residual interleave section 43_XYAs mentioned above
Are sorted into multiple (for example, h = 4) small series,
It is provided to the Kuttor quantizer 44. Outline of each sub-series
It is desirable to rearrange so that they are almost the same. Well
Weight calculator 34₀ Weighting factor from V₀Also weight interleaving
Part 43_VWBy the same sort as the residual sort by
To the vector quantization unit 44.
To be The vector quantizer 44 uses h residual small series as
Each weighted vector using the corresponding weight subsequence
Quantize h quantization indices J₁~ J_FourAnd output
It is given to one input of the selector 45A.

When the state determination unit 42 determines that the frame is a transient frame, the sub-frame division unit 31 is used as in the case of FIG. 4A.
Then, each frame of M samples is equally divided into n sub-frames, and the sub-frames corresponding to the _n sub-channels CH _{1 to} CH _n and 2 M / h samples (processing frames) of the immediately preceding sub-frame are respectively generated. To be distributed. From the distributed n processing frames, the residual /
Residual coefficients X ₁ to X _n and the spectrum envelope E ₁ to E _n is generated by the envelope calculation section 32 ₁ to 32 _n. Power normalization unit 33 ₁ ~
33 _n normalizes the residual coefficients X _{1 to} X _n with the average power of each subframe, and the normalized residual coefficients X _{N1 to} X _Nn.
To the residual interleaving unit 35 _XY , the normalized gains G _{1 to} G _n to the weight calculating units 34 _{1 to} 34 _n , and outputs the indexes I _{G1 to} I _Gn representing those gains. Weight calculator 34
_{1 to} 34 _n are spectrum gains E _{1 to} E _n and normalized gains G _{1 to} G _n
Are respectively multiplied to generate weighting factors V _{1 to} V _n, which are applied to the weighting interleaving unit 35 _VW .

Sub-channel CH thus obtained
The n normalized residual sub-sequences X _{N1 to} X _Nn of _{1 to} CH _n are input to the residual interleaving unit 35 and all sub-sequence components are rearranged in the same manner as in the case of FIG. It is output as the series Y _{1 to} Y _m . The same weighting interleaving unit 35 performs the same interleaving on the _n weighting coefficient sub-sequences V _{1 to} V _{n of the} sub-channels CH _{1 to} CH _n , and outputs them as _m sequences W _{1 to} W _m . The thus-obtained residual series (vectors) Y _{1 to} Y _m after interleaving are weighted in the vector quantizer 36 using the corresponding weighted series (vector) W _{1 to} W _m after interleaving. The vector is quantized, and the quantized indexes J _{1 to} J _m are output.

A 1-bit code Id indicating a stationary frame or a transient frame, vector quantization indexes J _{01 to} J _0h in the stationary frame, and an approximate shape index I _E0
And the normalized gain index I _G0 are output, and the vector quantization indices J _{1 to} J _m , the rough index I _{E1 to} I _En, and the normalized gain index I _G0 are output in the transient frame.
_{G1 to} I _Gn are output. These vector quantization index, outline index, and normalized gain index are selected and output by the selectors 45A, 45B, and 45C according to the state determination code Id.

FIG. 15 shows an example of a decoder corresponding to the embodiment of the encoder shown in FIG. 13, and parts corresponding to those in FIG. 12 are given the same reference numerals. Roughness index I _E0 or I _E1 ~ I
_En is input and is supplied to the outline reproduction section 62 ₀ or 62 by the selector 80B according to the input state determination code Id. Also, the vector quantization index J _{01 to} J _0h or J ₁
~ J _m is input to the selector 80A, and the vector reproducing unit 72 ₀ or 7 is selected by the selector 80A according to the code Id.
2 and the normalized gain index I _G0 or I _{G1 to} I
_{One of Gn} is input to the selector 80C and is supplied to the gain reproducing section 75 ₀ or 75 depending on the code I _S. In the general shape reproduction units 62 ₀ and 62, the reproduction unit 62 A reproduces the prediction coefficient corresponding to the input index,
The approximate shape of the frequency characteristic is calculated by the approximate shape calculation unit 62B using the reproduced predicted number.

When the code Id indicates a stationary signal frame, vector quantization indexes J _{01 to} J _0h are given to the vector reproducing unit 81, each index is inverse vector quantized, and h small series Y ₀₁ 'to Y _0h ' will be played.
These small sequences will be integrated into the deinterleaving portion 82 normalized residual coefficients X _N0 by residual interleave part 45 opposite sort in the encoder of the original 1 series in FIG. 13 _YX, to the residual coefficients X _N0 The multiplier 77 ₀ multiplies the reproduction gain G ₀ from the reproduction unit 75. The multiplication output is the reproduction outline coefficient E ₀ from the outline reproduction unit 62 ₀ in the power denormalization unit 65 ₀ .
Are multiplied, that is, denormalized, to recover the frequency domain coefficients. Its frequency domain coefficients are reversed modified discrete cosine transformed time domain signal by the inverse MDCT section 66 ₀ similarly to the decoder of Figure 12, 2M since the windowed optionally in mating portion 78 ₀ overlaid further frame The first half of the sample frame and the second half of the immediately preceding 2M sample frame are overlapped and output to the terminal 51 as a reproduced sound signal of 1 frame of M samples.

When the code Id indicates a transient frame, the approximate shape reproducing section 62 reproduces the spectrum outline of each sub-frame from the input outline indexes I _{E 1 to} I _En, and the outline coefficients are obtained. E ₁ to E _n is outputted. The normalized gain index I _{G 1} ~I _Gn that is input are respectively reproduced by the gain reproducing unit 75 as the gain G ₁ ~G _n. Further, the inputted vector quantization indexes J _{1 to} J _m are the vector reproducing unit 7
2 will be played respectively. Vector Y ₁ to Y _m of m that is they play in reverse interleave part 76 _YX, n sub-sequence by performing reordering encoder sorting performed in interleaving section 35 _XY opposite 13 The normalized residual coefficients of X _{N1 to} X _Nn . These residual coefficient sub-sequences X _{N1 to} X _Nn are multiplied by the corresponding gains G _{1 to} G _n reproduced by the multipliers 77 _{1 to} 77 _n , respectively. The respective multiplication outputs are respectively subjected to the denormalization unit 6
In 5 ₁ to 65 _n, is the inverse normalization is multiplied by the envelope E ₁ to E _n reproduced, respectively the frequency domain coefficients are reproduced. These frequency domain coefficients are respectively used in the inverse MDCT units 66 _{1 to} 66 _n.
Are inversely MDCT-converted into a time domain signal, and these time domain signals are further windowed by the frame superimposing sections 78 _{1 to} 78 _n as needed, so that the first half and the second half of adjacent subframes (adjacent subchannels) are separated. The superposed portions are superposed, and the superposed portions are temporally sequentially combined by the frame synthesizing unit 79 to reproduce the acoustic signal in the transient frame and output it to the output terminal 51.

In the case where the input signal is a stereo signal, the encoder in FIG. 16 corresponds to FIG. 13 and has the same reference numeral for the left channel side and the same reference numeral for the right channel side. The suffix "R" is attached to each. That is, the left channel signal and the right channel signal input from the input terminals 11 _L and 11 _R are processed in the same manner as the processing for the stationary frame shown in FIG. The residual coefficients of the divided subframes of the frame and the right channel signal frame are mixed with each other and weighted vector quantization is performed. Others are the same.

A decoder for decoding the encoded output of the stereo signal of FIG. 16 has the same reference numerals as those in FIG. 15 corresponding to those of FIG. 15 with the subscript “L” for the left channel side and the subscript for the right channel side. "R" is attached. In this case, in the stationary frame, both the left channel signal and the right channel signal are decoded as in the case of FIG. 15, but in the transient frame, m reproduced by the vector reproducing unit 72.
All components of these vectors are deinterleaved by 76 _YX in Fig. 1.
6. The rearrangement by the residual interleaving unit 35 _XY of 6 is restored, and the decoded signals of the left channel subframes are combined by the frame combining unit 79 _L , and the terminal 51
Is output to the _L, the decoded output of the right channel subframe is output after being synthesized by the frame synthesizing portion 79 _R to the terminal 51 _R.

In FIGS. 13, 15, 16 and 17, interleave mapping and deinterleave mapping are determined in advance, but a plurality of interleave mapping methods are prepared and vector quantization is performed using each of them, and the most distorted distortion is obtained at that time. Choose a mapping that reduces
The selection information may be transmitted. Spectral outlines and auxiliary information of normalized gain (power) (that is, outline index and gain index) are transmitted for each frame in a normal (stationary) frame. In the above-described embodiment, in the transitional frame, the spectrum outline and gain information are transmitted independently for each subframe, but vector quantization may be performed collectively for each frame. In a system that transmits a fixed amount of information in each frame, if the same number of bits as in a normal frame is assigned to auxiliary information in each subframe, the proportion of auxiliary information in the total amount of information increases and the overall distortion increases. Because there is a possibility that It is also possible to collectively perform vector quantization on a plurality of subframes. Alternatively, the spectrum outline of each subframe and the gain may be commonly used for each subframe.

In the embodiments of FIGS. 13 and 16, since the power bias cannot be used for the reduction of the coded information amount in the steady state, the steady state frame and the transient state frame are discriminated, and the steady state frame is subframed. The case where the principle of the present invention is applied to a frame in a transient state is divided into subframes and applied as a plurality of series of signals is shown. By the way, particularly in an acoustic signal including a pitch component such as a voice or a musical sound, the correlation between the frames becomes high when the steady state frames are continuous,
Utilizing this fact, the amount of encoded information can be reduced by modifying a part of the embodiments of FIGS. 13 and 16 as follows.

That is, as shown in FIG. 18, the residual shape calculating unit 91 predicts the rough shape E _R of the residual coefficient of the present frame from the residual coefficient of the past stationary frame, and calculates the residual error in FIG. / The residual coefficient X ₀ from the normalization unit 32F in the approximate shape calculation unit 32 ₀ is normalized by being divided by the predicted residual outline E _R in the residual approximate normalization unit 92. . In the case of voice or musical sound, the residual coefficient in the frequency domain for each frame usually has a pitch component, and this pitch component often continues over a plurality of frames. However, in this embodiment, the residual shape E _R and the residual coefficient X ₀ in a stationary frame are used.
Such a pitch component can be suppressed by normalizing, and as a result, the flattened fine structure coefficient Z
It is said. That is, the modified embodiment of FIG. 18 realizes predictive coding using inter-frame correlation by vector quantizing the fine structure coefficient Z instead of vector quantizing the residual coefficient.

The spectrum outline E ₀ from the spectrum outline calculating unit 32E and the residual coefficient outline E _R from the residual outline calculating unit 91 are multiplied by the weight calculating unit 93 for each corresponding sample and weighted. The coefficient is V ₀ . The power normalization unit 33 ₀ normalizes the fine structure coefficient Z by its power and outputs a normalized gain index I _G0 . The power-normalized fine structure coefficient Z _N is weighted by the vector quantization unit 44 by the weighting vector V ₀ from the weighting calculation unit 93. In this case, as in the case of FIG.
_The fine structure coefficient Z _N and the weighting coefficient V ₀ are rearranged by _ZY and 43 _VW into h small series Y _{01 to} Y _0h and W _{01 to} W _0h , respectively, and the vector quantization unit 44 sets each fine structure coefficient small series. Weighting vector quantization is performed on Y _{01 to} Y _0h with the corresponding small weight series W _{01 to} W _0h . The quantization indexes J _{01 to} J _0h are given to the selector 45A in FIG. Quantization index J ₀₁
Respectively through J _0h corresponding vector _{C (J 01) ~C (J} 0h) is being denormalized by residual envelope E _R from the residual envelope calculation section 91 in the inverse normalization unit 94, the residual The coefficient X _q is restored, and the residual shape calculating unit 91 predicts the residual shape of the next frame based on the residual coefficient X _q . It is obvious that the modification of FIG. 18 can be similarly applied to the embodiment of FIG.

Decoding in the case where the fine structure coefficient Z as shown in FIG. 18 is vector-quantized, the configuration between the vector reproducing unit 81 and the denormalization unit 65 ₀ in the decoder of FIG. 15 is shown in FIG.
It may be modified as shown in. That is, the vector quantization indexes J _{01 to} J _0h are restored by the vector reproducing unit 81 to the vectors Y ₀₁ 'to Y _0h ' of the small series, that is, the vectors C (J ₀₁ ) to C (J _0h ), and the deinterleave is performed. Further in FIG. 18 in part 82 _YZ
Rearrangement by the interleaving unit 43 of
It is integrated into a series of normalized fine structure coefficients Z _N '. Part 1
The sequence output is the multiplier 77 ₀ and the power normalization unit 33 of FIG.
Gain G ₀ of reproducing normalized gain index I _G0 from ₀
Is multiplied (gain reproducing section 78 in FIG. 15) to reproduce the fine structure coefficient Z ′. The fine structure coefficient Z is denormalized by the residual shape inverse normalization unit 83 by the residual shape E _R 'from the residual shape calculation unit 84, and the residual coefficient X ₀ ' is reproduced. The residual coefficient is denormalized by the inverse normalization unit 65 ₀ by the spectrum outline E ₀ from the outline calculation unit 62B in FIG. 15 to be a frequency domain coefficient, and the inverse MDCT unit 66 in FIG.
Given to ₀ . The subsequent processing is the same as in the case of FIG. The residual rough shape calculating unit 84 is configured in the same manner as the residual rough shape calculating unit 91 (FIG. 18) of the encoder, and calculates the residual error of the current frame from the input residual error coefficient X ₀ 'of the past frame. Approximate coefficient E _R 'is predicted. The modification of FIG. 19 can be similarly applied to the decoder of FIG.

When encoding and decoding as such a fine structure, for example, in a transient frame, encoding is performed without using inter-frame correlation (prediction encoding),
When returning to a normal frame, the past state of interframe correlation is reset. That is, it is only necessary to reset the inside of each of the residual shape calculating units 91 and 84 in FIGS. Alternatively, a subframe in a transient frame and one frame of a stationary frame have different time lengths,
Assuming that the power normalization gain and the spectrum outline parameter are common, the prediction based on the interframe correlation may be used as they are. That is, for example, the residual shape calculating units 91 and 84 may be operated by using the spectral outline parameter or gain in two subframes as the spectral outline parameter or gain two frames before.

Stereo 4 channels or 5 channels
Input a signal, each of which is
It may be divided into subframes to form a multi-channel signal. 1
Figure shows two input signals separated into a low-frequency side signal and a high-frequency side signal
16 input terminals 11_L , 11 _R You may enter in. this
In the decoder, the interleaving unit 76_XYThen multiple groups
Residual coefficients are created or multiple residuals corresponding to each output port are created.
The difference coefficient is created or the residual coefficient of one group is created 1
It is played back in one time domain signal.

As described above, in the embodiments of the encoders shown in FIGS. 13 and 16, only the frame in which the power fluctuation on the time axis is large is divided into sub-frame units and the coefficients in the frequency domain are obtained by MDCT. By rearranging, the deviation of the power in the frequency domain and the deviation of the power in the subframes, that is, the deviation in the time domain can both be reflected in the variation of the weighting coefficient in the vector.
Then, the deviation of the average weighting coefficient (power) between the small series can be reduced. If small series are created in order without rearrangement, the variation of the weighting coefficient within the small series is small, and the variation of the average weighting coefficient between the small series is large. In this case, the distortion cannot be reduced unless adaptive bit allocation is performed between the small sequences. According to the present invention, since the variation between the small sequences is small, even if the bit allocation is fixed, the distortion reduction effect is hardly impaired. This is because, in the case of the present invention, the control of the quantization distortion for reducing the distortion is weighted, and the control of the vector quantization distortion is performed by the weighted vector quantization.

In the embodiments of FIGS. 4A, 11 and 16, the case where the same type of signal, for example, an acoustic signal, is coded as a plurality of input signals to be coded has been described. It can be applied to a plurality of signal sequences of different types as long as the deviation of B occurs.
As an example, an encoder to which the encoding method of the present invention is applied when two sequences of a video signal and an audio signal are multiplexed and transmitted with a constant amount of information will be described with reference to FIG. Here, an example in which transform coding is used for an audio signal and both interframe prediction and transform coding are used for a video signal is shown. What kind of coding method is applied to each signal is essential to the present invention. That is not the case.

In FIG. 20, digital video signals and digital audio signals are input to terminals 11a and 11b, respectively. The processing for the audio signal is a case where the same processing as the processing in one sub-channel of FIG. 4A is performed for each same frame as the video signal, and the configuration thereof is shown in a simplified manner. The audio signal is time-frequency converted for each frame by the conversion unit 32Bb. The frequency domain coefficient obtained as a result is given to the outline calculation unit 32Eb and the normalization unit 32Fb. The rough shape calculation unit 32Eb calculates the spectrum rough shape Eb from the frequency domain coefficient, and the normalization unit 32Fb and the weight calculation unit 34b.
And the index I _Eb representing the outline is output. The normalization unit 32Fb obtains the residual coefficient by dividing the corresponding components of the spectrum outline Eb of the corresponding frequency domain coefficients, and further normalizes (divisions) by the average power of the frame to obtain the normalized residual. The difference coefficient X _Nb is output, and the index I _Gb representing the power normalization gain _Gb is output. The weight calculation unit 34b calculates the spectrum outline Eb and the gain.
Gb is multiplied and, if necessary, the perceptual weight is also multiplied to output the weighting coefficient Vb. These residual coefficient X _Nb and weighting coefficient Vb are respectively the residual interleaving section 35 _XY and the weighting interleaving section.
Given to 35 _VW .

On the other hand, the video signal is interframe predictor 95.
At, the signal predicted from the previous frame is subtracted, and the difference signal D is given to the conversion unit 32Ba. The conversion unit 32Ba performs time-frequency conversion on the difference signal D to generate frequency domain coefficients,
It is given to the rough shape calculation unit 32Ea and the normalization unit 32Fa. Outline calculator 32
The operations of Ea, the normalization unit 32Fa, and the weight calculation unit 34a are the same as the operations of the portions 32Eb, 32Fb, and 34b for the acoustic signal,
The outline calculation unit 32Ea outputs the spectrum outline Ea and the index I _Ea representing the spectrum outline Ea, and the normalization unit 32Fa outputs the normalized residual coefficient.
X _Na , a power normalized gain Ga, and a gain index I _Ga are output, and the weight calculation unit 34a outputs a weight coefficient Va. The residual coefficient X _Na and the weighting coefficient Va are respectively the residual interleaving unit 35
_XY and weight are given to interleave section 35 _VW . The residual interleaving unit 35 _XY rearranges the components of the residual coefficients X _Na and X _Nb , m
The residual small series (vectors) Y _{1 to} Y _m are output and given to the vector quantizer 36. The weight interleaving unit 35 _VW rearranges the components of the weighting factors Va and Vb in the same manner as the residual rearrangement, outputs _m weighting subsequences W _{1 to} W _m, and supplies them to the vector quantizing unit 36. The vector quantizer 36 performs weighted vector quantization on the residual small sequences Y _{1 to} Y _m using the corresponding small weight sequences W _{1 to} W _m , and outputs quantization indexes J _{1 to} J _m .
After all, the encoder of FIG. 20 encodes the quantization indexes J _{1 to} J _m , the outline indexes I _{E a} and I _Eb , and the gain indexes I _Ga and I _Gb for each frame with respect to the input video signal and the input audio signal. Output as a result.

FIG. 21 shows an example of the structure of the inter-frame prediction section 95 shown in FIG. The subtractor 95F subtracts the prediction signal obtained in the previous frame from the predictor 95E from the input video signal of the current frame and outputs the difference signal D. Quantization index J ₁ of the current frame from the vector quantization unit 36
When J _m is given to the reproducing unit 95A, the reproducing unit 95A reads m vectors corresponding to the indexes J _{1 to} J _m from the internal codebook (inverse quantization), and performs the residual rearrangement and the reverse arrangement. The residual series X _Na and X _Nb corresponding to the residual series X _Na and X _Nb are reproduced to reproduce the residual series X related to the video signal.
a'is given to the denormalization unit 95B. The inverse normalization unit 95B
The residual series Xa 'is multiplied (denormalized) by the spectrum outline Ea to obtain a frequency domain coefficient, which is given to the inverse transform unit 95C. The inverse transformation unit 95C frequency-time transforms the given frequency domain coefficient to generate a time domain signal (corresponding to the reproduction difference signal D '), and supplies it to the adder 95D. The adder 95D adds the prediction signal used in the current frame from the predictor 95E to the reproduction difference signal D'and updates the prediction signal to the predictor 9D.
Give to 5E. The predictor 95E holds the updated prediction signal and supplies it to the subtractor 95F as a prediction signal for the input video signal of the next frame.

FIG. 22 shows a simplified example of a decoder for the encoder shown in FIG. The vector reproducing unit 72 calculates _m vectors Y ₁ ′ to Y from the given quantization indexes J _{1 to} J _m.
It gives the deinterleaving unit 76 reads out the _m 'from within the codebook. The inverse interleaving section 76 rearranges the components of the vectors Y ₁ 'to Y _m ' in the reverse of the residual rearrangement in FIG. 20, and outputs the residual series Xa 'and Xb' corresponding to the differential video signal and the audio signal, The denormalization units 65a and 65b are supplied to the respective denormalization units 65a and 65b. On the other hand, the reproducing unit 62a uses the given outline index I _Ea.
To reproduce the spectrum outline of the differential video signal and further multiply by the gain Ga designated by the gain index I _{G a} to give it to the denormalization unit 65a as a spectrum outline coefficient. The inverse normalization unit 65a multiplies each component of the residual coefficient Xa 'by the corresponding spectrum outline coefficient of the differential video signal (inverse normalization), obtains the frequency domain coefficient of the differential video signal, and supplies it to the inverse transform unit 66a. give. The inverse transform unit 66a frequency-time transforms the given frequency domain coefficient to perform differential video signal D '.
To the adder 67, add to the decoded video signal of the previous frame held in the predictor 68, output the addition result to the terminal 51a as the decoded video signal of the current frame, and for the next frame. It is held in the predictor 68.

Similarly, the reproducing unit 62b uses the indexes I _Eb , I
The spectrum outline coefficient of the acoustic signal is reproduced from _Gb and given to the denormalization unit 65b. The inverse normalization unit 65b uses the residual coefficient X
Each component of b'is multiplied by the corresponding coefficient of the spectrum outline of the acoustic signal to obtain the frequency domain coefficient of the acoustic signal, which is given to the inverse transform unit 66b. The inverse transform unit 66b frequency-time transforms the given frequency domain coefficient and outputs the decoded acoustic signal to the terminal 51b.

As described above, in the embodiment shown in FIG. 20, in the frame in which the power is biased between the video signal and the audio signal, the weighted vector quantization results in the coding in which the higher power is more important. Will be done. Therefore,
Quantization distortion can be reduced to that extent. In the embodiment of FIG. 20, since the inter-frame prediction processing is further performed on the video signal, the level of the differential video signal D becomes small when the change of the image is small over a plurality of frames. The power deviation can be increased, and the quantization distortion due to the coding of the video signal and the audio signal can be reduced accordingly.

FIG. 23A shows the left and right channels according to the CELP method.
The present invention is suitable for an encoder that encodes a channel stereo signal.
An example of the case of application is shown, and parts corresponding to FIG.
Same reference number with L and R indicating left and right channels
I have. Left and right channel input terminals 11
_L, 11_RCorresponding to the LPC having the same configuration as that of FIG. 2A.
Analysis part 12_L, 12_R, LPC synthesis filter 13_L, 13_R,
Adaptive codebook 14_L , 14_R , Multiplier 15_L, 15_R, Addition
Bowl 18_L, 18_R, Subtractor 19_L, 19_R, Hearing weighting unit
20_L, 20_RIs provided. However integrated noise
Codebook 16, distortion calculation / codebook search unit 21, weight coding unit
22 is provided in common for the left and right channels.
Each operates in the same manner as described in FIG. 2A. Left and right channel
Input audio signal for each frame, as in the case of FIG. 2A.
LPC analysis, linear prediction coefficient P _L, P_RIs required.
This is a characteristic feature of this embodiment to which the present invention is applied.
Corresponds to each index C in the integrated noise codebook 16
The number of frames equal to the number of signal systems (number of channels)
Sano's noise code vector (called integrated noise code vector)
Have been written and each integrated noise code vector
A predetermined division part (partial vector)
Used as the random code vector for each channel
Is to be done. In this example, each index in the codebook is
2 frame length integrated noise code vector
Are stored, and the first half and the rear of each joint random code vector are stored.
Half are the noise code vector for the left channel and the right channel, respectively.
It is used as a random code vector for the channel. Integrated noise
For the codebook 16, one integrated noise code vector index
Box C, the corresponding specific pair of left and right
Channel noise code vector is read out and multiplied
Bowl 17_L, 17_RGiven to.

Similar to that described with reference to FIG. 2A, first, g _L1 = g _R1
= 0, the cutout lengths S _L and S _R that give the adaptive code vector that minimizes the distortion of the synthesized sound with respect to the left and right input audio signals for each frame are determined by the distortion calculation / codebook search unit 21, and are further cut out. From the adaptive code vector (pitch component vector) generated by the lengths S _L and S _R , the gains g _L0 and g _R0 that minimize the distortion of each system are calculated. With S _L , S _R , g _L0 , and g _R0 determined in this way, the pitch component synthesized sound output from the synthesis filters 13 _L and 13 _R is removed from the input acoustic signal by subtractors 19 _L and 19 _R. The noise component vector obtained as a result is next targeted as the noise component synthesized sound by the synthesis filters 13 _L and 13 _R when the noise code vector is used as the excitation vector. This target noise component synthesized sound is represented by R _L = [R _L1 , ..., R _Ln ] ^t , R _R = [R _R1 ,
, R _Rn ] ^t , the impulse response matrices of the synthesis filters 13 _L and 13 _R are represented as H _L and H _R, and the partial vectors corresponding to the left and right channels of the read integrated noise code vector are C _Lj = [ C _Lj1 , ..., C _Ljn ] ^t , C _Rj = [C _Rj1 , ..., C _Ljn ] ^t , the integrated distortion d is expressed by the following equation: d = ‖R _L −g _{L1 HL} C _Lj ‖ ² + ‖R It can be represented by _R −g _R1 H _R C _Rj ‖ ² . In order to determine the integrated noise vector that minimizes the integrated distortion, it is tentatively g
_L1 = g _R1 = 1 and putting the minimum value of d _min is d _min = ‖R _L ‖ ² + ‖R _R ‖ ^{_{2 - (R L t H L}} C Lj) 2 / ‖H L C Lj ‖ ² - (R _R ^t H _R C _Rj ) ² / ‖H _R C _Rj ‖ ² Since the target values R _L and R _R are constants, in order to minimize the integrated distortion d, the following equation D = (R _L ^t H _L C _Lj ) ² / ‖H _L C _Lj ‖ ² + (R _R ^t H _R C _Rj ) ² / ‖H _R C _Rj ‖ ^{2 The} maximum integrated vector C _j = {C _Lj , C _Rj } may be selected from the integrated noise codebook 16. Next, g _L1 and g _R1 are determined so that the distortion of each of the left and right channels is minimized.

The ideal gain is the following expression g _L1 = R _L ^t H _L C _Lj / ‖H _L C _Lj ‖ ² g _R1 = R _R ^t H _R C _Rj / ‖H _R C _Rj ‖ ² Therefore, the above D is represented by the following equation D = g _L1 ² ‖H _L C _Lj ‖ ² + g _R1 ² ‖H _R C _Rj ‖ ² . That is, in the embodiment using this CELP, the optimum gain that minimizes the distortion in the process of vector quantizing the excitation signal using the integrated noise codebook 16 is performed.
This means that g _L1 and g _R1 are automatically determined.

The gains g _L0 , g _R0 , g determined in this way
_The code G obtained by, for example, vector-coding _L1 and g _R1 in the coding unit 22 and the above-mentioned codes S _L , S _R , C, P _L , and P _R are shown in FIG.
It is output as the encoding result for each frame for the left and right channel acoustic signals by the A encoder. Note that, in FIG. 23A described above, the case where the prediction analysis and the encoding are performed for each frame is described so that the operation of the CELP to which the present invention is applied can be easily understood. It is obvious that each frame may be divided into a plurality of subframes and the above-mentioned codes S _L , S _R , C and G may be obtained for each subframe.

Here, for example, if the powers of the input acoustic signals of the left and right channels are almost the same, the gains g _L0 and g _R0 are about the same, and the gains g _L1 and g _R1 are also about the same. Therefore, the pair of left channel noise vector and right channel noise vector, which are the first half and the second half of the integrated noise vector read from the integrated noise codebook 16, are given similar gains and are added to the periodic component vector, respectively. Excitation vectors E _L and E _R , the pair of left-channel noise vector and right-channel noise vector corresponding to the index selected to minimize the integrated distortion contribute to the integrated distortion to the same degree. ing. This means that almost the same amount of information is assigned to the encoding of the left and right channel acoustic signals.

If the power of the left channel acoustic signal is significantly larger than the power of the right channel acoustic signal, the gain g for the adaptive code vector determined so as to minimize the distortion of the synthesized sound for each channel.
The relationship between _L0 , g _R0 and the gains g _L1 , g _R1 for the noise code vector is also g _L0 >> g _R0 and g _L1 >> g _R1 . In the latter relationship, in the process of selecting the integrated noise code vector from the integrated noise codebook 16 so that the integrated distortion is minimized, the contribution of the noise code vector of the left channel to the integrated distortion is significantly larger than the contribution of the right channel noise code vector. Therefore, it is equivalent to allocating a large amount of information to the coding by placing emphasis on the left channel acoustic signal having high power. However, in this embodiment, since the deviation of the power in the time direction is not taken into consideration, the total amount of information used for encoding is constant for each frame.
On the contrary, when the power of the input audio signal of the right channel is larger than that of the left channel, the noise vector of the right channel is considered to be more important, which is equivalent to allocating more encoded information amount to the right channel audio signal. Become.

As described above, according to the embodiment of the encoder of FIG. 23A, the coded information amount is automatically and flexibly distributed without bit allocation according to the bias of the powers of the acoustic signals of the left and right channels. However, the random codebook does not need to be separately provided for the left and right channels, and only one integrated random codebook is required. Further, even if the determined gains g _L1 and g _R1 are incorrect, the entire information is not confused. FIG. 23B is shown in FIG.
Codes S _L , S _R obtained by the 3A encoder for each frame,
An embodiment of a decoder for decoding left and right channel acoustic signals from P _L , P _R , G, and C is shown. Parts corresponding to those in FIG. 2B are shown with the same reference numerals with L and R representing left and right channels. is there. This decoder uses the synthesis filter 5 in FIG. 2B.
2, the multipliers 55 and 57, and the adder 58 are respectively provided on the left and right channels, but one weight decoding unit 53 and one integrated noise codebook 56 are provided, and the integrated noise codebook 56 is the integrated noise in FIG. 23A. It is the same as the codebook 16. Codes S _L and S _R are given to adaptive code books 54 _L and 54 _R ,
Specify the cutout length for the excitation vector of the immediately preceding frame held there. Weight decoding unit 53
From the code G, the gains g _L0 , g _L1 , g _R0 and g _R1 are decoded and given to the multipliers 55 _L , 57 _L , 55 _R and 57 _R , respectively.

The adaptive codebooks 54 _L and 54 _R cut out a partial vector of a length designated by the codes S _L and S _R from the excitation vector of the previous frame, and repeatedly connect their copies to adapt one frame length. Generate a code vector and multiply by 55
Give to _L and 55 _R. Multipliers 55 _L and 55 _R multiply the adaptive code vectors by gains g _L0 and g _R0 , respectively, and give them to adders 58 _L and 58 _R. Codes P _L and P _R that represent linear prediction coefficients
Is given as a filter coefficient to the linear prediction synthesis filters 52 _L and 52 _R. The integrated noise code vector specified by the index code C is read from the integrated noise code length 56, and the first half and the latter half thereof are given to the multipliers 57 _L and 57 _R , respectively, and are multiplied by the gains g _L0 and g _R0 . The multiplication result is added to the adaptive code vector by the adders 58 _L and 58 _R to generate excitation vectors E _L and E _R , which are given to the linear prediction synthesis filters 52 _L and 52 _{R, and} are supplied to the left and right channels. The voice is synthesized and output to the terminals 51 _L and 51 _R. When the similarity between the left and right channel signals L and R is high, the power can be concentrated on the L + R signal by converting them into L + R and LR signals and inputting them respectively. Further, the distortion can be reduced.

In the embodiment of FIG. 23A, adaptive codebook 1
4 _L and 14 _R may be omitted, and an integrated codebook having an excitation vector as an integrated vector may be used instead of the integrated noise codebook 16, as in the case described with reference to FIG. 2A. In addition, the linear prediction analysis units 12 _L and 12 _R in FIG.
The predictive coefficients P _L and P _R may be determined by LPC analysis of the output synthesized sounds of the synthesis filters 13 _L and 13 _R in the past frame. Corresponding to these modifications, the adaptive codebooks 54 _L and 54 _R are omitted also in the decoder of FIG. 23B, and an integrated codebook having a large number of excitation vectors corresponding to indexes is used instead of the integrated noise codebook 56. May be. Further, the linear prediction coefficients P _L and P _R given to the synthesis filters 52 _L and 52 _R are the same as the synthesis filters 52 _L instead of using those received from the outside.
The past output synthetic sounds of _L and 52 _R may be obtained by LPC analysis.

[Brief description of drawings]

FIG. 1A is a block diagram of an encoder that performs encoding by conventional adaptive bit allocation scalar quantization, and B is a block diagram showing a decoder for the encoder of A. FIG.

2A is a block diagram of a conventional encoder using CELP, and B is a block diagram of a decoder for the encoder of A. FIG.

FIG. 3 is a block diagram for explaining an encoder according to the principle of the present invention, and B is a block diagram for explaining a decoder according to the principle of the present invention.

4A is a block diagram showing an example of an encoder when the present invention is applied to a transform coding method, and FIG. 4B is a block diagram showing a configuration example of a residual / rough shape calculation unit in A. FIG.

FIG. 5 is a time chart showing division of an input signal into subframes for each frame and frames for LOT processing in each subchannel.

6A is a diagram showing a procedure for obtaining a spectrum outline from a prediction coefficient in the outline calculation unit 32E of FIG. 4A, and FIG.
The figure which shows the modification procedure of A.

FIG. 7 is a vector element array diagram showing the elements of a residual series vector when each frame of the left channel and the right channel is divided into two, and the interleaving method for them.

FIG. 8 is a vector element array diagram showing a modification of the interleaving method of FIG.

FIG. 9 is a block diagram showing the configuration of a vector quantization unit.

10A is a block diagram showing another example of each residual / rough shape calculation unit in FIG. 4A, and B is a block diagram showing still another example of each residual / rough shape calculation unit in FIG. 4A.

FIG. 11 is a block diagram illustrating an example of a decoder for the encoder of FIG. 4A.

12A is a block diagram showing a configuration example of selecting a plurality of interleaving methods in the encoder of FIG. 4A, and B is FIG.
FIG. 3 is a block diagram showing a configuration example for switching a rearrangement method on the decoding side for the A encoder.

FIG. 13 is a block diagram showing a functional configuration of an encoder configured so that only a frame whose signal is in a transient state is divided into subframes and encoded.

14 is a diagram showing an example of frame division in the encoder of FIG. 13, division of a transient frame into subframes, and each window function in each MDCT.

15 is a block diagram showing a functional configuration example of a decoder corresponding to the encoder shown in FIG.

16 is a block diagram showing a functional configuration example of an encoder in which the encoding method of FIG. 13 is applied to encoding a stereo signal.

17 is a block diagram showing a functional configuration example of a decoder corresponding to the encoder of FIG.

FIG. 18 is a block diagram showing a configuration of a deforming unit in the case where the encoder of FIGS. 13 and 16 is modified so as to encode the fine structure coefficient instead of the residual coefficient.

FIG. 19 is a block diagram showing the configuration of a modification unit corresponding to the modification of FIG. 18 in the decoder of FIGS. 15 and 17;

FIG. 20 is a functional block diagram showing the configuration of an encoder to which the present invention is applied when a video signal and an audio signal are multiplexed and encoded.

21 is a block diagram showing a configuration example of an inter-frame prediction unit 95 in FIG.

22 is a block diagram showing a configuration of a decoder corresponding to the encoder shown in FIG.

23A is a block diagram showing an example of the encoder of the present invention when applied to CELP, and B is a block diagram of a decoder for the encoder of A. FIG.

Claims (58)

[Claims] 1. A coding method for collectively coding signals of a plurality of systems, including the following steps: (a) For each section of a predetermined length in each of the above systems, the power of the signal in that section Then, the weight is determined based on the power, and (b) in each of the above systems, a signal in that section is normalized with the power to generate a normalized signal for each section, and
A gain index representing the normalized gain is output, and (c) the normalized signals of the plurality of systems are combined in a predetermined array to generate at least one series of signal vectors, and (d) the plurality of The above-mentioned weights of each system are combined in the same array as the above-mentioned normalized signal to generate at least one series of weight vectors, and (e) the above-mentioned signal vector is weighted vector-quantized with the above-mentioned weight vector, and the quantization vector is represented. The quantization index is output, and (f) the quantization index and the gain index are output as at least a part of the encoding result for the signals of the plurality of systems. 2. The encoding method according to claim 1, wherein the step (a) comprises: (a-1) time-frequency converting the signal in the section to obtain a frequency domain coefficient; and (a-2) the frequency. Obtaining the outline of the region coefficient, (a-3) including the step of normalizing the frequency domain coefficient with the outline to obtain the residual coefficient, and determining the weight based on the power of the residual coefficient, Step (b) is a step of normalizing the residual coefficient in the section with the power to obtain a normalized residual coefficient as the normalized signal. 3. The encoding method according to claim 2, wherein the time-frequency transform in step (a-1) is a modified discrete cosine transform. 4. The encoding method according to claim 2 or 3, wherein the step (a-3) determines the weight based on at least the power and the outline. 5. The encoding method according to claim 2, 3 or 4, wherein the step (a-3) further determines a weight based on a perceptual weight. 6. The encoding method according to claim 1, wherein a signal having a length of the section is extracted from an input signal of a single system at regular intervals and sequentially circulated and distributed to the plurality of systems. Generating the signals of the plurality of systems is included. 7. The encoding method according to claim 6, wherein the length of the section in each system is 2 times the length of the cycle.
In step (a), (a-1) the signal in the section is transformed into a frequency domain coefficient by a modified discrete cosine transform, and (a-2) the outline of the frequency domain coefficient is obtained. -3) Normalizing the frequency domain coefficient with its outline to obtain a residual coefficient, and determining the weight based on the power of the residual coefficient.The step (b) includes the residual of the section. It is a step of normalizing the difference coefficient with the power and obtaining a normalized residual coefficient as the normalized signal. 8. The encoding method according to claim 2, 3, 4, 5, or 7, wherein in step (a-2), a prediction coefficient is obtained by performing linear prediction analysis on the signal in the section, and the prediction coefficient is calculated. , And each step of Fourier transforming the quantized prediction coefficient to obtain the above-mentioned outline, the step (f) includes an index representing the quantization of the above-mentioned prediction coefficient and information corresponding to the above-mentioned outline. Is output as another part of the encoding result for the signal. 9. The encoding method according to claim 2, 3, 4, 5 or 7, wherein the step (a-2) is a time domain signal by inverse Fourier transforming the absolute value of the frequency domain coefficient in the section. And, the prediction coefficient is obtained by linear prediction analysis of the time domain signal, the prediction coefficient is quantized, and the quantized prediction coefficient is Fourier-transformed to obtain the outline, and the step (f) is performed. Outputs an index representing the quantization of the prediction coefficient as another part of the coding result for the signal as information corresponding to the outline. 10. The encoding method according to claim 2, 3, 4, 5, or 7, wherein said step (a-2) divides said frequency domain coefficient into a plurality of sub-bands. Determining a scale factor representative of frequency domain coefficients, quantizing the scale factor of each of the sub-bands to obtain a quantized scale factor as the outline, the step (f) is the quantization. An index representing the scale factor is output as information corresponding to the above outline as another part of the encoding result for the above signal. 11. The encoding method according to claim 1, wherein the predetermined array in steps (c) and (d) is the residual coefficient from each system for each section. All the elements of the above are rearranged according to the mapping of the array stored in advance in the memory means, and are generated as a predetermined number of the above-mentioned signal vectors, and all the elements of the above-mentioned weights from the respective systems are set into the above-mentioned memory means. The weighting vectors are rearranged according to the mapping of the same array stored in, and are generated as a predetermined number of the weight vectors. 12. The encoding method according to any one of claims 1 to 10, wherein mappings of a plurality of types of arrays are stored in advance in the memory means, and the step (c) includes each of the plurality of mappings. When used, the vector quantization in step (e) above is performed and the mapping that minimizes the quantization distortion required during the vector quantization is selected, and in step (f) above, the quantization distortion is minimized. And the code representing the selected mapping is output as another part of the encoding result. 13. The encoding method according to claim 1, wherein the signal of the one predetermined system is a video signal. 14. The encoding method according to claim 13, wherein the step (a) in the one predetermined system of the plurality of systems generates a differential signal by performing prediction processing on the signal for each section. Then, the power of the differential signal is obtained, and the weight is determined based on the power. The step (b) in the one predetermined system normalizes the differential signal with the power, This is a step of converting to a digitized signal. 15. The encoding method according to claim 13 or 14, wherein the other predetermined signal of the system is an acoustic signal. 16. The encoding method according to claim 1, further comprising the following steps: (g) dividing a single-system input signal into frames of a constant period, and dividing each frame. Is a stationary or transient, and (h) when it is determined to be a transient frame, the input signal of that frame is divided into sections of the above-mentioned fixed length to obtain the signals of the above multiple systems. Distributing, performing steps (a) to (f) above, (i) if the frame is stationary, determine the power of the input signal of the frame, determine the weight based on that power, and j) The input signal of the single system of the frame is normalized with the power to obtain a normalized signal, and a gain index representing the normalized gain is output, and (k) the normalized signal of the input signal. To rearrange at least one sequence of signal vectors (L) The weights of the input signal are rearranged in the same manner as the normalized signal of the input signal to generate at least one series of weight vectors, and (m) the weight vector is generated in the step (k). Search for a codebook having a plurality of different quantization vectors of the same length as the signal vector corresponding to each quantization index, for the normalized signal of step (j), the step
(l) Select the quantization vector that minimizes the weighted distance by the weight vector, determine the corresponding quantization index, and (n) the gain that represents the normalized gain by the power in step (i) above. The index and the quantization index determined in step (m) are output as at least a part of the coding result for the input signal in the stationary frame. 17. The encoding method according to claim 16, wherein the step in each of the plurality of systems
(a) is (a-1) Time-frequency conversion is performed on the signal in the section to obtain frequency domain coefficients, (a-2) Spectrum outline of the signal in the section is obtained, and (a-3) The frequency domain coefficient is normalized by the spectrum outline to obtain the residual coefficient, and the weight is determined based on the power of the residual coefficient.The step (f) includes the outline representing the spectrum outline. The index is output as another part of the encoded result. 18. The encoding method according to claim 17, wherein the step (f) for each section in the transient frame is at least one of a set of approximate indexes and a set of gain indexes of the plurality of systems. One of them is quantized together and the corresponding quantized code is output as a part of the above-mentioned coding result. 19. The encoding method according to claim 16, 17 or 18, wherein the step (i) predictively encodes the input signal of a frame determined to be stationary, and the resulting code is stationary. It is output as at least a part of the encoding result of the input signal in the frame. 20. The encoding method according to claim 16, 17 or 18, wherein said step (i) comprises (i-1) time-frequency converting said input signal of said frame determined to be stationary. Obtain the frequency domain coefficient, (i-2) obtain the spectrum outline and average power of the input signal in the frame, (i-3) normalize the frequency domain coefficient with the spectrum outline and the average power, the Outputting the resulting normalized residual coefficient as the normalized signal, wherein step (f) includes the gain index representing the average power and the outline index representing the spectrum outline as the stationary frame. It is output as a part of the encoding result for the above input signal of. 21. The encoding method according to claim 16, 17 or 18, wherein said step (i) comprises: (i-1) time-frequency converting said input signal in said frame determined to be stationary. Obtain the frequency domain coefficient, (i-2) obtain the spectrum outline of the input signal in the above frame, (i-3) normalize the frequency domain coefficient with the spectrum outline, and obtain the residual coefficient obtained as a result. (I-4) Predict the outline of the residual coefficient from the residual coefficient of the past frame, and (i-5) Normalize the residual coefficient with the predicted outline to generate the fine structure coefficient. In the step (j), the average power of the fine structure coefficient is obtained, and the normalized fine structure coefficient obtained by normalizing the fine structure coefficient with the average power is output as the normalized signal. Including the above step (f) The envelope index representing the gain index and the spectrum envelope representing the average power output as another part of the coding result. 22. An excitation vector having a frame length selected from a codebook for each frame of an input acoustic signal is multiplied by a gain to be applied to a linear prediction synthesis filter, and between the acoustic signal synthesized thereby and the input acoustic signal. In the CELP coding method for determining the excitation vector of the codebook with the minimum distance measure and the gain for the codebook, the codebook is preliminarily written with an integrated vector of the number of systems of the frame length corresponding to each index. And a plurality of linear prediction synthesis filters respectively corresponding to a plurality of sequences to which a plurality of acoustic signals are input, and the method includes the following steps: (a) the plurality of systems , The prediction coefficient is obtained for each frame by linear prediction analysis, and given as a filter coefficient to each corresponding linear prediction synthesis filter, (b) The partial vector of each predetermined part of each integrated vector read from the integrated codebook is given as the excitation component vector of the linear prediction synthesis filter of the corresponding system, and the sum of the distance scales of the plurality of systems is the minimum. And (c) calculate the gain for the excitation component vector corresponding to each system of the selected integrated vector, and (d) the index representing the selected integrated vector and the integrated vector. And a code representing the gain with respect to the excitation component vectors respectively corresponding to the plurality of systems, are output as at least a part of the encoding result for the input acoustic signals of the plurality of systems. 23. A method of decoding a signal in which a plurality of systems of signals are collectively vector-quantized and coded using one integrated codebook, wherein: (a) at least one vector quantum input for each section of constant length. At least one integrated vector corresponding to the generalization index is read from the integrated codebook, and (b) the elements of the integrated vector obtained in step (a) above are rearranged in a predetermined array to generate a normalized signal of multiple systems. Then, (c) the normalized gain is reproduced from the gain index input for each constant length section, and the normalized signals corresponding to the plurality of systems are denormalized by the normalized gain to obtain the reproduced signals of the plurality of systems. Output. 24. The decoding method according to claim 23,
The above step (c) includes the following steps: (d) Reproducing a spectrum outline from the outline index input for each of the above intervals for each of the plurality of lines,
The denormalized signal is multiplied to obtain the frequency domain coefficient, and (e) the frequency domain coefficient of each system obtained in step (d) is frequency-time converted for each section. It is converted into a time domain signal and output as the reproduced signals of the plurality of systems. 25. The decoding method according to claim 24,
The frequency-time transformation in step (e) above is the inverse modified discrete cosine transformation. 26. The decoding method according to claim 24 or 25, wherein the step (e) is a step of sequentially combining the time domain signals of the plurality of systems in a predetermined time order to form one integrated reproduction signal. including. 27. The decoding method according to claim 24,
The step (e) includes a step of combining the plurality of time domain signals into a plurality of integrated reproduction signals having a smaller number. 28. The decoding method according to claim 23, wherein the step (b) is performed by using an index representing an array mapping input for each section,
One of a plurality of predetermined rearrangement mappings is selected, and the rearrangement is performed according to the selected rearrangement mapping. 29. The decoding method according to claim 23, wherein the input state code is decoded to determine whether it is a stationary frame or a transient frame, and if it is a transient frame. For example, the steps (a) to (c) are executed for the frame, and the reproduction signals obtained in the plurality of systems in each section are sequentially shifted in time and combined to produce an integer multiple of the section. As a signal of a frame, if it is a stationary frame, perform the following steps for that frame: (d) combine at least one integrated vector corresponding to at least one vector quantization index input for the above frame Read from the codebook, (e) rearrange the elements of the integrated vector obtained in step (d) above into a predetermined array to generate a single-system normalized signal, and (f) enter it into the above frame. The normalized gain is reproduced from the applied gain index, and the normalized signal of the single system is denormalized by the normalized gain and output as a reproduced signal. 30. The decoding method according to claim 29,
The step (c) includes the following steps: (c-1) The spectrum outline is reconstructed from the outline indexes respectively input to the plurality of lines, and the spectrum outline is denormalized as described above. The normalized signal is multiplied to obtain the frequency domain coefficient, and (c-2) the frequency domain coefficient of each system obtained in step (c-1) above is frequency-time transformed to a time domain signal. , And output as the reproduction signals of the plurality of systems. 31. The decoding method according to claim 29 or 30, wherein when the frame is a stationary frame, the step (f) includes the following steps: (f-1) an approximate input to the frame. The spectrum shape is reproduced from the shape index, the frequency domain coefficient is obtained from the denormalized signal by using the spectrum shape, and (f-2) is obtained in step (f-1). The frequency domain coefficient is subjected to frequency-time conversion to be converted into a time domain signal, which is output as the reproduction signal. 32. The decoding method according to claim 31, wherein
In the step (f-1), the denormalized normalization signal is reproduced as a fine structure coefficient, and the fine structure coefficient is denormalized by the residual outline predicted from the past frame to obtain the residual coefficient. , The step of multiplying the residual coefficient by the spectrum outline to obtain the frequency domain coefficient. 33. The decoding method according to claim 32,
The residual outline is obtained as an outline from the residual coefficient obtained in step (f-1) in the past stationary frame. 34. The decoding method according to claim 32,
The residual shape is obtained from the spectrum shape reproduced from the shape index in one section in the past transient frame. 35. The decoding method according to claim 30, wherein
When the status code indicates a transient frame, the spectrum outline of one section reproduced in step (c-1) in each system is commonly used for all sections in the frame. 36. The decoding method according to any one of claims 23 to 33, wherein in the predetermined one of the plurality of systems, the step (c) is a step of subtracting the denormalized signal from the denormalized signal. And a reproduction signal of the previous section is added to the difference signal of the current section to obtain a reproduction signal of the current section. 37. A method of decoding a plurality of audio signals encoded by CELP, comprising the steps of: (a) reading a corresponding integrated noise vector by referring to an integrated codebook according to an input index. , The partial vector of the part of the integrated vector that corresponds to each of the above-mentioned multiple systems is cut out and given to each corresponding system, and (b) the gains for the partial vectors corresponding to the above-mentioned multiple systems are respectively reproduced from the input gain code. Then, the excitation vectors are obtained by multiplying the corresponding partial vectors respectively, (c) in each of the above systems, the excitation vector is given to the synthesis filter in which the prediction coefficient is given as the filter coefficient, and the acoustic signal is synthesized and reproduced. Output as an acoustic signal. 38. An encoder for collectively encoding signals of a plurality of systems, which is provided for each of the systems and obtains the power of the signal of the section for each section of a predetermined length, and the power thereof The weight is determined based on, the signal in the section is normalized with the power to generate a normalized signal, and a power normalizing means for outputting a gain index representing the normalized gain, and each of the plurality of systems. Signal sequence combining means for combining the normalized signals from the power normalizing means in a predetermined array to generate at least one series of signal vectors, and the power normalizing means for each of the plurality of systems. Combine the weights in the same array as the normalized signal above,
A weight array combination means for generating at least one series of weight vectors, and a vector quantization means for quantizing the signal vector with the weight vector, and outputting a quantization index representing the quantization vector; The quantization index and the gain index are output as at least a part of the encoding result for the signals of the plurality of systems. 39. The encoder according to claim 38, wherein the signal of the section is subjected to time-frequency conversion for each of the systems to obtain frequency domain coefficients, and an approximate shape of the frequency domain coefficients is obtained. Residual rough shape calculating means for obtaining a residual coefficient by normalizing the coefficient with its rough shape, and weight calculating means for determining the weight based on the power of the residual coefficient,
The power normalizing means is means for normalizing the residual coefficient in the section with the power to obtain a normalized residual coefficient as the normalized signal. 40. The encoder according to claim 39, wherein the time-frequency conversion performed by the residual outline calculation means is a modified discrete cosine transform. 41. The encoder according to claim 39, wherein said residual outline calculation means of each said system obtains a prediction coefficient by performing linear prediction analysis on the signal of said section, and said prediction coefficient. Including a coefficient quantizing means for quantizing and predicting the approximate shape by Fourier transforming the quantized predictive coefficient, and quantizing the predictive coefficient. The index indicating is output as another part of the encoding result for the signal as information corresponding to the outline. 42. The encoder according to claim 39, wherein the residual outline calculation means of each system calculates the absolute value of the frequency domain coefficient in the section and the inverse Fourier transform of the absolute value. Inverse Fourier transforming means for transforming into a time domain signal, linear predictive analyzing means for linearly predicting and analyzing the time domain signal to obtain a prediction coefficient, and outputting an index representing the quantized prediction coefficient by quantizing the prediction coefficient. Coefficient quantizing means and Fourier transform of the quantized prediction coefficient to obtain the rough shape calculating means, and the index representing the quantization of the predictive coefficient is the signal as information corresponding to the rough shape. It is output as another part of the encoding result for. 43. The encoder according to claim 39, wherein the residual shape calculating means of each system divides the frequency domain coefficient into a plurality of small bands, and represents the frequency domain coefficient of each of the small bands. Scale factor calculating means for determining the scale factor to be obtained, and scale factor quantizing means for quantizing the scale factor of each of the small bands to obtain a quantized scale factor as the outline. Is output as another part of the encoding result for the signal as information corresponding to the outline. 44. The encoder according to claim 38, further comprising a memory means for storing a plurality of different array mappings, said signal array combining means corresponding to said plurality of different array mappings. , A plurality of sets of the weight array combining means and the vector quantizing means are provided, and the normalization signals from the power normalizing means of a plurality of systems are respectively obtained at the time of vector quantization by the plurality of sets. Quantized distortions are given, the quantized distortions are compared with each other, and the quantized distortions are compared with each other, and the quantized distortions are compared with each other. Means for selectively outputting the corresponding quantization index from the plurality of sets, and the quantization index that minimizes the quantization distortion. And a code representing the selected array mapping are output as another part of the above coding result. 45. The encoder according to any one of claims 38 to 44, wherein prediction is performed on the signal for each predetermined interval of the plurality of systems for each interval to generate a differential signal. Means is provided, and the power normalizing means is means for obtaining the power of the differential signal, determining the weight based on the power, normalizing the differential signal with the power, and making the normalized signal. 46. The encoder according to claim 38, further comprising: determining whether the input signal of a single system is stationary or transient for each frame of a constant cycle. State determination means for determining, a sub-frame dividing means for dividing the input signal of the frame into intervals of the constant length and distributing as signals of the plurality of systems when it is determined to be a transient frame, and determining as stationary The power of the input signal of the frame is determined, the weight is determined based on the power, the input signal of the single system of the frame is normalized by the power to obtain a stationary frame normalized signal, A stationary frame power normalizing means for outputting a stationary frame gain index representing the normalized gain and the stationary frame normalized signal of the input signal of the stationary frame are arranged. Stationary frame signal array combining means for generating at least one series of stationary frame signal vectors and at least one series by rearranging the weights of the input signals of the stationary frames in the same manner as the normalized signals of the input signals. And a stationary frame codebook having a plurality of quantization vectors having the same length as that of the stationary frame signal vector and corresponding to each quantization index. , The stationary frame vector quantization for searching the stationary frame normalized signal, selecting the quantization vector having the smallest weighting distance by the stationary frame weight vector, and determining the corresponding stationary frame quantization index. Means are provided, and the stationary frame gain in Box and the stationary frame quantization index is output as at least part of the coding result for said input signal in the stationary frame. 47. The encoder according to claim 46, wherein the input signal in the frame determined to be stationary is subjected to time-frequency conversion to obtain a frequency domain coefficient, and a spectrum outline of the input signal in the frame is calculated. Then, the frequency domain coefficient is normalized by the spectrum outline and the residual coefficient obtained as a result is calculated as a steady frame residual outline calculating means, and the outline of the residual coefficient is calculated from the residual coefficients of past frames. Residual rough shape predicting means for predicting, residual residual normalizing means for normalizing the residual coefficient with the predicted rough shape to generate a fine structure coefficient, and obtaining an average power of the fine structure coefficient, Stationary frame power normalizing means for outputting the normalized fine structure coefficient obtained by normalizing the fine structure coefficient with the average power as the stationary frame normalization signal. Envelope index indicating a steady frame gain index and the stationary frame spectrum envelope representing the average power of the fine structure coefficients is outputted as another part of the coding result. 48. An encoder for encoding input acoustic signals of a plurality of systems, wherein the encoder is provided for each system, and a linear prediction analysis unit for obtaining a prediction coefficient by a linear prediction analysis for each frame, and a frame length are set as the above. An integrated codebook means in which a plurality of integrated vectors having a length multiplied by the number of systems is written in advance corresponding to each index, and when any one of the integrated vectors is selected, the integrated vector is determined in advance. The partial vector of each part is provided to each corresponding multiple system, is provided in each of the multiple systems, and the partial vector of each part of the integrated vector read from the integrated codebook is multiplied by the gain to excite. Weighting means for generating each vector, and the prediction coefficient from the linear prediction analysis means provided in each of the plurality of systems, Of the composite acoustic signal from the excitation filter and the composite acoustic signal from the synthetic filter means and the input acoustic signals of the plurality of systems. A subtracting means for obtaining a difference, and a distance measure between the synthetic acoustic signal and the input acoustic signal is obtained from the difference from the subtracting means, a sum of these distance measures is obtained as an integrated distortion, and the integrated distortion with the minimum integrated distortion is obtained. Distortion calculation / codebook search means for selecting a vector from the integrated codebook and calculating the gain when the integrated distortion is minimized for each of the partial vectors of the selected integrated vector. A frame for the input audio signals of the plurality of systems, including a code representing the determined gain and an index representing the selected integrated vector. And outputs as at least part of the coding result. 49. A decoder of a signal in which signals of a plurality of systems are collectively vector-quantized and coded using one integrated codebook, and integrated vectors are written corresponding to respective quantization indexes, An integrated codebook for reading at least one integrated vector corresponding to at least one quantization index input for each fixed-length section, and an array in which elements of the integrated vector obtained from the integrated codebook are predetermined Inverse arraying means for rearranging into a plurality of systems and outputting as a normalized signal of a plurality of systems, and a normalized gain is reproduced from a gain index input for each section of a certain length, and the normalized gain corresponding to the plurality of systems is reproduced by the normalized gain. Denormalizing means for denormalizing the signal and outputting it as the reproduction signals of the plurality of systems. 50. The decoder according to claim 49, further comprising: demultiplexing the spectrum outlines from the outline indexes respectively provided in the plurality of systems and input in each of the sections, and denormalized. Frequency domain coefficient reproducing means for multiplying the normalized signal to obtain a frequency domain coefficient, and frequency domain coefficients for each of the systems, which are respectively provided in the plurality of systems and are frequency-time converted to a time domain signal. And a frequency-time converting means for outputting the reproduction signals of the plurality of systems. 51. The decoder according to claim 49 or 50, further comprising frame synthesizing means for sequentially combining the time domain signals of the plurality of systems in a predetermined time order to form at least one integrated reproduction signal. There is. 52. The decoder according to claim 49, 50 or 51, further comprising memory means for storing a plurality of different array mappings in advance, and said inverse array means is an array input for each section. One of the plurality of array mappings in the memory means is selected by the index representing the mapping, and the array is performed according to the selected array mapping. 53. The decoder according to claim 49, further decoding the input state code to determine whether it is a stationary frame or a transient frame, On the other hand, the reproduction by the denormalization means of the plurality of systems is executed, and the reproduction signals obtained in the plurality of systems in each section are sequentially shifted in time and combined to produce a signal of a frame having an integral multiple of the section. A stationary frame integrated codebook in which an integrated vector for a stationary frame is written corresponding to a stationary frame index for a stationary frame, and at least one quantization A constant for generating a single-system normalized signal by rearranging at least one integrated vector element read from the integrated codebook corresponding to the index into a predetermined array. An ordinary frame inverse arrangement means, and an ordinary frame inverse normalization means for denormalizing the single-system normalized signal by the normalized gain reproduced from the gain index input to the stationary frame and outputting it as a reproduced signal. Is provided. 54. The decoder according to claim 53, wherein the denormalized signal is multiplied by the spectrum outline reproduced from the input outline index provided in each of the plurality of systems. A multiplication means for obtaining a frequency domain coefficient and a frequency-time conversion means provided in each of the plurality of systems for frequency-time converting the frequency domain coefficient to obtain a time domain signal and outputting the time domain signal as the reproduction signal. Including. 55. A decoder according to claim 53 or 54, wherein the spectrum domain recovered from the profile index input in the stationary frame is used to generate the frequency domain from the denormalized signal. Residual coefficient reproducing means for obtaining a coefficient, and frequency-time converting means for frequency-time converting the frequency domain coefficient to convert it into a time domain signal and outputting it as the reproduced signal. 56. The decoder according to claim 55, wherein the residual coefficient reproducing means is the denormalized normalized signal from the stationary frame reverse rearranging means in the stationary frame is a fine structure coefficient, Residual outline prediction means for predicting the residual shape of the fine structure coefficient from the residual coefficient of the past frame, and the residual of the current frame by denormalizing the fine structure coefficient with the predicted residual shape. Residual outline denormalization means for obtaining a coefficient, and multiplication means for multiplying the residual coefficient by the spectrum outline to obtain the frequency domain coefficient. 57. The decoder according to any one of claims 49 to 56, wherein the normalized signal denormalized in a predetermined one of the plurality of systems is a difference signal, and the current section Prediction means for obtaining the reproduction signal of the current section by adding the reproduction signal of the previous section to the above difference signal. 58. A decoder of acoustic signals of a plurality of systems encoded by CELP, which corresponds to gain reproducing means for reproducing a gain corresponding to each of a plurality of systems from a code input for each frame and each index. Then, an integrated codebook in which an integrated vector having a length that is a multiple of the frame length is written, and the plurality of integrated vectors read from the integrated codebook corresponding to the index input for each frame. Weighting means for generating an excitation vector by multiplying the partial vector corresponding to each system by the gain, and a linear prediction coefficient provided as a filter coefficient provided in each of the plurality of systems, and an acoustic signal is generated from the excitation vector. And a synthesizing filter means for synthesizing.