JPH09127987A - Signal coding method and device therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform a plurality of decoding operations differed in bit rate, enhance the coding efficiency, and minimize the deterioration of a regenerated sound even at a low bit rate. SOLUTION: From the input signal from a terminal 101, a low-pass side signal is taken out by a LPF(low pass filter) 102, LPC-analyzed by a LPC analyzing and quantizing part 130, a LPC residual which is a short-range predictive residual is determined by a LPC reverse filter 111, a pitch is then determined by a pitch analyzing circuit 115, and a long-range prediction is performed by a pitch reverse filter 112 to determine a pitch residual. This pitch residual is MDCT-processed by a MDCT(modified discrete cosine transformation) circuit 113, and vector-quantized by a VQ(vector quantizing) circuit 114, and the quantized index is transmitted together with pitch lag and pitch gain. As the parameter for expressing LPC coefficient, a SLP (line spectrum pair) is transmitted.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a signal coding method and apparatus for coding an input signal such as a wideband voice signal, and particularly to a telephone band and a band other than the telephone band in which sufficient intelligibility as voice is obtained. The present invention relates to a signal coding method and apparatus that can be divided into two and realize an independent codec for a telephone band.

[0002]

2. Description of the Related Art Various coding methods are known in which signal compression is performed by utilizing the statistical properties of audio signals (including voice signals and acoustic signals) in the time domain and frequency domain and human auditory characteristics. ing. As this encoding method,
Broadly speaking, time domain coding, frequency domain coding,
Examples include analysis and synthesis coding.

[0003] Examples of high-efficiency coding of voice signals and the like include harmonic coding and MBE (Multiband Ex).
citation: Sine wave analysis coding such as multi-band excitation coding, SBC (Sub-band Coding: band division coding), LPC (Linear Predictive Coding: linear predictive coding), or DCT (discrete cosine transform), MDC
T (Modified DCT), FFT (Fast Fourier Transform), etc. are known.

[0004]

By the way, various encoding techniques for dividing an input signal into a plurality of bands and encoding the same are known, but the low side and the high side are encoded by a uniform method. Therefore, even if the encoding method is suitable for encoding the low-frequency side signal, the efficiency may be low in the encoding of the high-frequency side signal, or vice versa. When attempting to transmit with a low value, good coding may not be performed.

Further, although various rates of encoding / decoding apparatus are provided, it is inconvenient to use a different apparatus for each different rate, and one apparatus has a plurality of different bits. It is desirable to be able to encode or decode rate signals.

The present invention has been made in view of such circumstances, and when band-splitting and encoding, a high-quality reproduced sound can be obtained with a small number of bits, and a predetermined band, for example, It is an object of the present invention to provide a voice coding method and apparatus that can realize an independent codec for a telephone band.

[0007]

In order to solve the above-mentioned problems, a speech decoding method according to the present invention divides an input signal into bands, and divides each band signal into signals of respective bands. The feature is that different encodings are applied according to the characteristics.

Here, the band division may be performed by at least dividing an input voice signal of a wide band into a signal of a telephone band and a signal of a higher band than the telephone band.

Further, a short-term prediction is performed on the signal on the low frequency side of the band obtained by the band division to obtain a short-term prediction residual, and a long-term prediction is performed on the obtained short-term prediction residual. An example is orthogonal transformation of long-term prediction residuals.

[0010]

BEST MODE FOR CARRYING OUT THE INVENTION Preferred embodiments of the present invention will be described below. First, FIG. 1 shows a wideband speech signal coding apparatus to which an embodiment of a signal coding method according to the present invention is applied.

Here, the basic idea of the encoding apparatus of FIG. 1 is that an input signal is divided into a plurality of bands, and the signals for each divided band are mutually dependent on the signal characteristics of each band. Different encoding is applied. Specifically, the wideband input voice signal is divided into a plurality of bands, for example, a telephone band in which sufficient intelligibility as voice is obtained and a band higher than the telephone band. For signals in the low-frequency side telephone band, short-term prediction such as LPC (linear predictive coding) analysis is performed, and then long-term prediction such as pitch prediction is performed, and then orthogonal transformation is performed. Quantization is also performed, and parameters that represent short-term prediction coefficients such as LPC coefficients, and information related to long-term prediction such as pitch and pitch gain are also quantized. For signals in the higher band than the telephone band, short-term predictions are vector-quantized as they are on the time axis.

MDCT (Modified Discrete Cosine Transform) is used as the orthogonal transform, and the vector quantization is easily weighted by shortening the transform length at this time.
Furthermore, by setting this transform length to 2 ^N , that is, a power of 2, it is possible to speed up using FFT (Fast Fourier Transform). Further, the LPC coefficient (also for the post filter) for calculating the residual of the above short-term prediction and calculating the weight when vector-quantizing the orthogonal transform coefficient is
The optimum LPC coefficient is used in each analysis subframe by using a smooth interpolation of the one obtained in the current frame and the one obtained in the past frame. Further, in the case of performing the long-term prediction, prediction or interpolation is performed a plurality of times in one frame, and the pitch lag and pitch gain are quantized as they are, or the difference is taken, or a flag indicating an interpolation method is transmitted. There is. Furthermore, with respect to the prediction residual whose variance decreases as the number of predictions (frequency) increases, multi-stage vector quantization is performed to quantize the difference between the orthogonal transform coefficients, or the divided band is divided. Among these, only a parameter for one certain band is used, and a plurality of decoding operations with different rates are possible by the whole or a part of one encoded bit stream.

A description will be given below with reference to FIG. FIG.
Of the sampling frequency Fs
A wide band audio signal of about 0 to 8 kHz at 16 kHz is supplied. The wideband audio signal from the input terminal 101 is a low-pass filter 102 and a subtractor 106 for a low-frequency side signal, for example, a so-called telephone band signal of about 0 to 3.8 kHz and a high-frequency side signal, for example, a signal of 3.8 kHz to 8 kHz. Band divided into and. The low frequency side is thinned by the sampling frequency converter 103 within a range that satisfies the sampling theorem to be a signal of 8 kHz sampling, for example.

As the processing on the low frequency side, first, the LPC analysis / quantization unit 130 performs a Hamming window with an analysis length of, for example, about 256 samples per block.
The LPC coefficient of about 10th order, that is, the α parameter is calculated, and the LPC inverse filter 111 obtains the LPC residual. In this LPC analysis, 96 samples of one block of 256 samples, which is the unit of analysis, are overlapped with the next block, so that the frame interval, that is, the frame interval becomes 160 samples.
This frame interval is 20 msec at 8 kHz sampling
become. Further, in this LPC analysis / quantization unit 130,
The α parameter, which is an LPC coefficient, is converted into an LSP (line spectrum pair) parameter, and the quantized one is transmitted.

That is, in the LPC analysis / quantization unit 130, the LPC analysis circuit 132, to which the low-frequency side signal from the sampling frequency converter 103 is input, hums the length of about 256 samples of the input signal waveform as one block. Through a window, a linear prediction coefficient, a so-called α parameter, is obtained by the autocorrelation method. The framing interval, which is the unit of data output, is 20 for 160 samples, for example.
msec.

The α parameter from the LPC analysis circuit 132 is sent to the α → LSP conversion circuit 133 and converted into a line spectrum pair (LSP) parameter. This converts the α parameter obtained as the direct type filter coefficient into, for example, 10 pieces, that is, 5 pairs of LSP parameters. The conversion is performed using, for example, the Newton-Raphson method. The conversion to the LSP parameter is because it has better interpolation characteristics than the α parameter.

The LSP parameter from the α → LSP conversion circuit 133 is vector-quantized or matrix-quantized by the LSP quantizer 134. At this time, vector quantization may be performed after obtaining the difference between frames, or matrix quantization may be performed on a plurality of frames at once. Here, 20 msec is taken as one frame, and LSP parameters calculated every 20 msec are collectively matrix-quantized for two frames.

The quantized output from the LSP quantizer 134, that is, the index of the LSP vector quantization, is taken out via the terminal 131 and is also the quantized LSP.
The vector or dequantized output is the LSP interpolation circuit 13
Sent to 6.

The LSP interpolation circuit 136 is an LS that is vector-quantized by the LSP quantizer 134 every 20 msec.
This is for interpolating the set of the previous frame and the current frame of the vector of P to obtain the rate required for the subsequent processing. In this example, the rate is 8 times and the rate is 5 times. At the 8 times rate, the LSP vector is updated every 2.5 msec. This is because when the residual waveform is analyzed and synthesized, the envelope of the synthesized waveform becomes a very smooth and smooth waveform, so that the LPC coefficient is 20 m.
This is because an abnormal sound may be generated if it changes rapidly every sec. That is, if the LPC coefficient is gradually changed every 2.5 msec, the occurrence of such abnormal noise can be prevented.

In order to execute the inverse filtering of the input voice using the LSP vector for every 2.5 msec which has been interpolated in this way, the LSP → α conversion circuit 137
The LSP parameter is converted into, for example, an α parameter which is a coefficient of a direct type filter of about 10th order. This LSP → α
The output from the conversion circuit 137 is sent to the LPC inverse filter circuit 111 for obtaining the LPC residual, and the LP
The C inverse filter 111 performs an inverse filtering process with the α parameter updated every 2.5 msec to obtain a smooth output.

Further, the LSP coefficient every 4 msec interpolated by the LSP interpolation circuit 136 at the rate of 5 is sent to the LSP → α conversion circuit 138 to be converted into an α parameter and MD.
It is sent to the VQ (vector quantization) weight calculation circuit 139 for weight calculation used for the quantization of the CT coefficient.

The output from the LPC inverse filter 111 is the pitch inverse filter 11 for pitch prediction which is a long-term prediction.
2 and 122.

Next, the long-term prediction will be described. The long-term prediction is performed by subtracting the waveform shifted on the time axis by the pitch period or pitch lag obtained by the pitch analysis from the original waveform to obtain a pitch prediction residual. In this example, three-point pitch prediction is performed. Has gone by. The pitch lag refers to the number of samples corresponding to the pitch cycle of the sampled time axis data.

That is, in the pitch analysis circuit 115, the pitch analysis is performed once per frame, that is, the analysis length is 1 frame, and the pitch lag L in the pitch analysis results is obtained.
₁ is sent to the pitch inverse filter 112 and the output terminal 142, and the pitch gain is sent to the pitch gain VQ (vector quantization) circuit 116. Pitch gain VQ circuit 11
In 6, the pitch gain at three points corresponding to the above three-point prediction is vector quantized, the codebook index g ₁ is taken out from the output terminal 143, and the representative value vector or the dequantized output is subtracted from the pitch inverse filter 115 and subtracted. Bowl 11
7 and the adder 127, respectively. The pitch inverse filter 112 outputs a pitch prediction residual with three-point pitch prediction based on the pitch analysis result. This pitch prediction residual is, for example, the MDCT circuit 113 which is an orthogonal transform means.
And is subjected to MDCT processing and then subjected to auditory weighting vector quantization by a VQ (vector quantization) circuit 114. In this VQ circuit 114, the VQ weight calculation circuit 139
Is subjected to perceptually weighted vector quantization, and the output, index IdxVq _1, is taken out from the output terminal 141.

Further, in this example, a pitch inverse filter 122, a pitch analysis circuit 124, and a pitch gain VQ circuit 126, which are a separate pitch prediction system, are further provided in order to improve the prediction accuracy. That is, the analysis center is placed at an intermediate position between the pitch analysis centers, and the pitch analysis circuit 125 performs the analysis at 1/2 frame period. The pitch lag L ₂ from the pitch analysis circuit 125 is the pitch inverse filter 122 and the output terminal 14.
5, the pitch gain is sent to the pitch gain VQ (vector quantization) circuit 126, respectively. Pitch gain VQ
The circuit 126 vector-quantizes the pitch gain vector at the three points, sends the pitch gain index g ₂ which is a quantized output to the output terminal 144, and sends the representative vector or the inverse quantized output to the subtractor 117. Here, the pitch gain at the position of the analysis center of the original frame period is considered to be a value close to the pitch gain from the pitch gain VQ circuit 116. Therefore, regarding the pitch gain at this position, the pitch gain VQ circuit 116, The subtractor 117 takes the difference between the respective inverse quantized outputs, and the pitch gain VQ circuit 118 vector-quantizes the difference to obtain the pitch gain difference index g _1d.
I am sending to 6. The representative vector of the pitch gain difference or the inverse quantized output is sent to the adder 127, and the sum of the representative vector or the inverse quantized output from the pitch gain VQ circuit 126 is sent to the pitch inverse filter 122 as the pitch gain. . The pitch gain index g ₂ obtained from the output terminal 143 is the pitch gain index at the intermediate position. The pitch prediction residual from the pitch inverse filter 122 is the MDCT circuit 1
23, MDCT processing is performed, and this is sent to the subtractor 128, and the representative vector or dequantized output from the VQ (vector quantization) circuit 114 is subtracted, and the difference is VQ.
It is sent to the circuit 124, vector-quantized, and indexed.
IdxVq2 is sent to the output terminal 147. This VQ circuit
The output from the VQ weight calculation circuit 139 performs vector quantization with auditory weighting.

Next, the signal processing on the high frequency side will be described.

In the signal processing on the high frequency side, basically, the input signal is band-divided, at least one of the divided high-frequency side signals is frequency-converted to the low frequency side, and then converted to the low frequency side. The sampling rate of the signal is reduced, and the signal with the reduced sampling rate is predictively encoded.

The wideband signal supplied to the input terminal 101 of FIG. 1 is input to the subtractor 106, and from this wideband signal,
A low-frequency side signal extracted by the LPF (low-pass filter) 102, for example, a so-called telephone band signal of about 0 to 3.8 kHz is subtracted. As a result, from the subtractor 106, a high frequency side signal, for example, 3.8 kHz to 8 kHz
The Hz signal is output. However, due to the characteristics of the actual LPF 102, the output from the subtractor 106 is 3.8 k
A small amount of components below Hz remains, and signal processing on the high frequency side is performed on components above 3.5 kHz or above 3.4 kHz.

The high-frequency side signal has a frequency width of 4.5 kHz from 3.5 kHz to 8 kHz from the subtractor 106, but the frequency is shifted or converted to the low-frequency side by down-sampling or the like to perform signal processing. For example, 4k
It is necessary to narrow down to the Hz width. Here, in consideration of the synthesis with the low frequency side after that, since the region around 3.5 kHz to 4 kHz is sensitive to hearing, the component or power is small as the property of the audio signal without cutting it and the influence on the hearing is not affected. LPF or B for a small amount of 0.5 kHz from 7.5 kHz to 8 kHz
It is cut by a PF (band pass filter) 107.

Next, frequency conversion to the low frequency side is performed. In this example, orthogonal conversion means, for example, FFT (Fast Fourier Transform) circuit 161, is used to convert to data on the frequency axis, and on this frequency axis. Data of the frequency shift circuit 162
After being shifted by, the inverse FFT circuit 164, which is an inverse orthogonal transform means, performs inverse FFT processing.

From the inverse FFT circuit 164, signals of, for example, 3.5 kHz to 7.5 kHz on the high frequency side of the input signal are output from 0 to 4.
The signal converted to the low frequency side of kHz is extracted. Since this signal can be expressed at a sampling frequency of 8 kHz, it is down-sampled by the down-sampling circuit 164 to obtain a sampling frequency of 8 kHz from 3.5 kHz to 7.5 kHz.
The signal should be in the Hz band. This down sampling circuit 1
The output from 64 is the LPC inverse filter 171 and the LPC.
It is sent to the LPC analysis circuit 182 of the analysis / quantization unit 180.

Since the LPC analysis / quantization unit 180 has a configuration similar to that of the LPC analysis / quantization unit 130 on the low frequency side, it will be briefly described.

That is, in the LPC analysis / quantization unit 180, the LPC analysis circuit 182 to which the low-frequency-converted signal from the downsampling circuit 164 is input, has a length of about 256 samples of the input signal waveform as one block. A linear prediction coefficient, so-called α parameter, is obtained by, for example, an autocorrelation method by applying a Hamming window. The α parameter from the LPC analysis circuit 182 is sent to the α → LSP conversion circuit 183 and converted into a line spectrum pair (LSP) parameter. L from the α → LSP conversion circuit 183
The SP parameters are vector-quantized or matrix-quantized by the LSP quantizer 184. At this time,
Vector quantization may be performed after taking the difference between frames. Alternatively, a plurality of frames may be collectively subjected to matrix quantization. Here, 20 msec is set as one frame, and LSP parameters calculated every 20 msec are vector-quantized.

The quantized output from the LSP quantizer 184, that is, the LSP vector quantization index LSPidx _H of the high frequency side signal is extracted via the terminal 181, and the quantized LSP vector or dequantized The output is sent to the LSP interpolation circuit 186.

The LSP interpolation circuit 186 is an LS that has been vector quantized by the LSP quantizer 184 every 20 msec.
This is for interpolating the set of the previous frame and the current frame of the vector of P to obtain the rate required for the subsequent processing, and in this example, the rate is quadrupled.

L for each 5 msec in which such interpolation is performed
In order to execute the inverse filtering of the input voice using the SP vector, the LSP → α conversion circuit 187 causes the LS
The P parameter is converted into an α parameter which is a coefficient of the LPC synthesis filter. The output from the LSP → α conversion circuit 187 is sent to the LPC inverse filter circuit 171 for obtaining the LPC residual, and the LPC inverse filter 171 is supplied.
Then, the inverse filtering process is performed with the α parameter updated every 5 msec to obtain a smooth output.

The LPC prediction residual output from the LPC inverse filter 171 is the LPC residual VQ (vector quantization) circuit 1
It is sent to 72, vector-quantized, and its output is L
The PC residual index LPCidx is taken out from the output terminal 173.

In the signal coding apparatus having the above-described configuration, by partially setting the low-frequency side configuration as a coding apparatus of an independent codec, or by switching the whole and a part of the output bit stream, It enables signal transmission and decoding with different bit rates.

That is, when transmitting all data from each output terminal of the configuration of FIG. 1, the transmission bit rate is 16
It becomes kbps (k bits / second), and the transmission bit rate becomes 6 kbps by transmitting data from some terminals.

Alternatively, 16k from all terminals in FIG.
transmit all data in bps, ie send or record,
By decoding all 16 kbps data on the receiving or playback side, a high quality 16 kbps voice signal can be obtained. By decoding 6 kbps data, a simple decoder can produce a 6 kbps quality voice signal. .

Here, in the configuration of FIG. 1, the output data from the output terminals 131, 141 to 143 corresponds to the data of 6 kbps, and the output terminals 144 to 147, 17 are further provided.
16 kbps by adding output data from 3 and 181
All data of s are obtained.

Next, a signal decoding device corresponding to the signal coding device of FIG. 1 will be described with reference to FIG.

In FIG. 2, the input terminal 200 has
LSP corresponding to the output from the output terminal 131 of FIG.
The vector quantized output of, the so-called codebook index LSPidx, is supplied.

The index LSPidx of this LSP is LPC
It is sent to the inverse VQ (inverse vector quantization) circuit 241 of the LSP of the parameter reproducing unit 240 to be subjected to inverse vector quantization or inverse matrix quantization to LSP (line spectrum pair) data, and sent to the LSP interpolation circuit 242 to be sent to the LSP. After the interpolation processing is performed, LP is performed by the LSP → α conversion circuit 243.
The C parameter is converted into an α parameter which is a C (linear prediction code) coefficient, and the α parameter is converted into the LPC synthesis filters 215 and 2.
25 and pitch spectral post filters 216, 2
Sent to 26.

Further, the input terminals 201, 202, 2 of FIG.
03, the output terminals 141, 142, 14 of FIG.
Index Is of vector quantization of MDCT coefficients from 3
xVq ₁ , pitch lag L ₁ , and pitch gain g ₁ are respectively supplied.

The vector quantization index IsxVq ₁ of the MDCT coefficient from the input terminal 201 is supplied to the inverse VQ (inverse vector quantization) circuit 211, inverse vector quantized, and inverse MDCT processed by the inverse MDCT circuit 212. After that, the signals are superposed and added by the superposition and addition (overlap add) circuit 213 and sent to the pitch synthesis filter 214. The pitch synthesizing circuit 214 is supplied with the pitch lag L ₁ and the pitch gain g ₁ from the respective input terminals 202 and 203. In the pitch synthesizing circuit 214, the inverse process of the pitch predictive coding in the pitch inverse filter 112 in FIG. The LPC-combined output is sent to the pitch spectral post filter 216, subjected to post filter processing, and taken out from the output terminal 219 as an audio signal corresponding to a bit rate of 6 kbps.

The input terminals 204, 205, 206 and 207 of FIG. 4 are respectively connected to the output terminals 144, 145 of FIG.
Pitch gain g ₂ of vector quantization of MDCT coefficients from 146 and 147, pitch lag L ₂ , index IsxV
q ₂ and pitch gain g _1d are respectively supplied.

The vector quantization index IsxVq ₂ of the MDCT coefficient from the input terminal 207 is supplied to the inverse VQ circuit 220, inverse vector quantized, and sent to the adder 221, and the inverse vector quantum from the inverse VQ circuit 211. M
After being added to the DCT coefficient and subjected to inverse MDCT processing by the inverse MDCT circuit 222, it is superimposed and added by the superposition and addition (overlap add) circuit 223, and the pitch synthesis filter 2
14 is sent. The pitch synthesis filter 224 has
Pitch lag L from each input terminal 202, 204, 205
₁ , the pitch gain g ₂ , and the pitch lag L ₂ are supplied, and the sum of the pitch gain g ₁ from the input terminal 203 and the pitch gain g _1d from the input terminal 206 in the adder 217 is supplied. . After the pitch residual synthesis processing is performed by the pitch synthesis filter 224, the LPC
It is sent to the synthesis filter 225 and subjected to LPC synthesis processing. The LPC synthesized output is sent to the pitch spectral post filter 226, subjected to post filter processing, sent to the upsampling circuit 227, and the sampling frequency is upsampled from 8 kHz to 16 kHz, for example, and then the adder 228 is added. Sent to.

Further, the LSP index LSPidx _H on the high frequency side is supplied from the output terminal 181 of FIG. 1 to the input terminal 207, and the index LSPidx _{H of} this LSP is
After being sent to the LSP inverse VQ (inverse vector quantization) circuit 246 of the LPC parameter reproducing unit 245 and inverse vector quantized into LSP data, and sent to the LSP interpolation circuit 247 to be subjected to the LSP interpolation processing, → α conversion circuit 2
At 48, it is converted into an α parameter of the LPC coefficient, and this α parameter is sent to the high frequency side LPC synthesis filter 232.

The input terminal 209 has an output terminal 17 shown in FIG.
The index LPCidx, which is the vector quantization output of the LPC residual on the high frequency side from 3, is supplied to the high frequency inverse VQ circuit 23.
Inverse vector quantization is performed at 1, and the result is sent to the high frequency side LPC synthesis filter 232. The output subjected to the LPC synthesis processing by the high frequency side LPC synthesis filter 232 has a sampling frequency of, for example, 8 kHz to 16 kHz in the upsampling circuit 233.
After being up-sampled to Hz, the FFT circuit 234, which is an orthogonal transform means, performs a fast Fourier transform to convert the signal into a signal on the frequency axis, a frequency shift circuit 235 performs frequency shift processing to the high frequency side, and an inverse FFT circuit 236. By performing the inverse fast Fourier transform, it becomes a time domain signal on the high frequency side, and is sent to the adder 228 via the superposition addition circuit 237.

In the adder 228, the signal from the upsampling circuit 227 is added, and the added signal is taken out from the output terminal 229 as an audio signal corresponding to a part of the bit rate of 16 kbps. A signal having a bit rate of 16 kbps as a whole is taken out by combining the signal from the output terminal 219.

Scalability will now be described. In the configurations shown in FIGS. 1 and 2, 6 kbps and 16
Two kinds of transmission bit rates of kbps are realized by almost the same encoding / decoding method, and the scalability of completely enclosing a bit stream of 6 kbps in a bit stream of 16 kbps is realized.
It is difficult to obtain such a perfect inclusion relationship when performing encoding and decoding with extremely different bit rates such as.

Here, even when the same encoding / decoding method cannot be applied, it is preferable to have scalability while having a maximum sharing relationship.

For this reason, the coding apparatus having the structure shown in FIG. 3 performs coding at 2 kbps to have the maximum shared portion or shared data with the structure shown in FIG. 16 of these in the stream
The case of using all kbps, the case of using 6 kbps, and the case of transferring 2 kbps are properly used according to the intended use.

Strictly speaking, as will be described later, 2 kbps
s uses all 2 kbps information, but in the 6 kbps mode, it is 6 kbps when the frame to be coded is voiced (V), 5.65 kbps when unvoiced (UV), and 16 kbps in the 16 kbps mode. 15.2 kbps when frame is voiced (V), 1 when unvoiced (UV)
It is 4.85 kbps.

The configuration and operation of the 2 kbps coding device shown in FIG. 3 will be described.

The basic idea of the coding apparatus shown in FIG. 3 is to obtain a short-term prediction residual of an input speech signal, for example, an LPC (linear prediction coding) residual, and perform a sine wave analysis (sinusoidal analysis).
lysis) coding, eg harmonic coding (har
monic coding) and a second coding section 320 that performs coding by waveform coding that performs phase transmission on the input speech signal. : Voiced) portion is encoded by the first encoding unit 31.
0 is used, and the second encoding unit 320 is used for encoding the unvoiced sound (UV: Unvoiced) portion of the input signal.

The first encoding unit 310 has, for example, L
Harmonic coding and multi-band excitation (M
A configuration for performing sine wave analysis encoding such as BE) encoding is used. For the second encoding unit 320, for example, a configuration of code excitation linear prediction (CELP) encoding using vector quantization by closed loop search of an optimum vector using an analysis method by synthesis is used.

In the example of FIG. 3, the audio signal supplied to the input terminal 301 is sent to the LPC inverse filter 311 and the LPC analysis / quantization unit 313 of the first encoding unit 310. The LPC coefficient or the so-called α parameter obtained from the LPC analysis / quantization unit 313 is sent to the LPC inverse filter 311, and the LPC inverse filter 311 extracts a linear prediction residual (LPC residual) of the input speech signal. . The LPC analysis / quantization unit 313 extracts a quantized output of an LSP (line spectrum pair) as described later, and sends this to the output terminal 302. The LPC residual from the LPC inverse filter 311 is sent to the sine wave analysis coding unit 314. In the sine wave analysis coding unit 314, pitch detection and spectrum envelope amplitude calculation are performed, and the V (voiced sound) / UV (unvoiced sound) determination unit 3 is performed.
15 is used to determine V / UV. The spectral envelope amplitude data from the sine wave analysis coding unit 314 is sent to the vector quantization unit 316. The codebook index from the vector quantization unit 316 as the vector quantization output of the spectrum envelope is the switch 3
The output from the sine wave analysis coding unit 314 is sent to the output terminal 303 via the switch 17, and the output from the sine wave analysis coding unit 314 is sent to the output terminal 304 via the switch 318. In addition, the V / UV determination unit 315
V / UV judgment output from the output terminal 305 is sent to the output terminal 305 and is also sent as a control signal for the switches 317 and 318. When the voiced sound (V) is used, the index and pitch are selected and the output terminals are output. From 303 and 304 respectively.

The second coding section 320 of FIG. 3 has a CELP (code excitation linear prediction) coding configuration in this example, and outputs the output from the random codebook 321 by the weighted synthesis filter 322. The weighted voice obtained by the synthesis processing is sent to the subtractor 323, the voice signal supplied to the input terminal 301 is taken out as an error from the voice obtained through the auditory weighting filter 325, and this error is calculated by the distance calculation circuit. 32
4, the distance calculation is performed, and a vector that minimizes the error is searched by the noise codebook 321. A time-axis waveform using a closed loop search using an analysis by synthesis method is used. Vector quantization is performed. This CELP coding is used for coding the unvoiced sound portion as described above, and the noise codebook 32 is used.
The codebook index as the UV data from 1 is taken out from the output terminal 307 via the switch 327 that is turned on when the V / UV determination result from the V / UV determination unit 315 is unvoiced (UV).

The LPC analysis / quantization unit 313 of such an encoding device can be shared as a part of the LPC analysis / quantization unit 130 of FIG. 1, and the output from the terminal 302 is as it is.
Can be used as an output from the output terminal 131. Also,
A part of the pitch data obtained by the sine wave analysis coding unit 314 can be used as an output from the pitch analysis circuit 115 of FIG. 1, and this pitch analysis circuit 115 is shared with the pitch output part in the sine wave analysis coding unit 314. It is also possible to do so.

As described above, although the coding method of FIG. 3 is different from the coding method of FIG. 1, both have common information and have scalability as shown in FIG.

In FIG. 4, the internal structure of the 2 kbps bit stream S2 differs depending on whether the analysis and synthesis frame is V (voiced sound) or UV (unvoiced sound).
2 kbps bitstream S2v for V Is a 2 kbps bit stream S2u for UV from two parts S2ve and S2va. Consists of two parts S2ue and S2ua, respectively. Part S2ve has a pitch lag of 1
1 bit per 160 frame sample (hereinafter 1 bit /
(Shown as 160 samples), the amplitude Am is 15 bits / 160 samples, giving a total of 16 bits / 160 samples. This is 8kHz
It corresponds to data having a bit rate of 0.8 kbps in sampling. The partial S2ue has an LPC residual of 11 bits / 80 samples and a spare 1 bit / 160 samples, for a total of 23 bits / 160 samples, which corresponds to a bit rate of 1.15 kbps. Remaining part S2 of 2 kbps bit stream S2
va and S2ua are common parts or common parts with the above-mentioned 6 kbps and 16 kbps, and the part S2va is LSP data 32 bits / 320 samples and V / UV judgment data 1 bit / 16.
A total of 24 bits / one with 0 samples and pitch lugs 7 bits / 160 samples
It is 60 samples, which is equivalent to data at a bit rate of 1.2 kbps. Part S2ua consists of LSP data 32 bits / 320 samples and V / UV judgment data 1 bit / 160 samples, totaling 17 parts.
It becomes a bit / 160 sample, which corresponds to a bit rate of 0.85 kbps.

In addition, a 6 kbps bit stream S6
In the same manner as the bit stream S2, the internal structure differs only partially when the analysis frame is V and when it is UV. The 6 kbps bitstream S6v at V is
From 2 parts S6va and S6vb, 6 kbps for UV
Of the bit stream S6u of the two parts S6ua, S6
Each made up of ubs. As described above, the part S6va has the same data content as the part S2va, and the part S6va
vb is a pitch gain of 6 bits / 160 samples and a pitch residual of 18 bits / 32 samples, giving a total of 96 bits / 160 samples, 4.8k.
Corresponds to bps bit rate data. Also, part S
6ua has the same data content as the above part S2ua, and the part S6ub has the same data content as the above part S6vb.

In addition, a bit stream S1 of 16 kbps
6, like the bit streams S2 and S6, the internal structure differs only partially when the analysis frame is V and when it is UV. 16 kbps bit stream S for V
16v is four parts S16va, S16vb, S16vc,
From S16vd, the 16 kbps bitstream S16u for UV is divided into four parts S16ua, S16ub, and S1.
It consists of 6uc and S16ud. Part S16va
Is the data content common to the parts S2va and S6va, and S16vb is the data content common to the parts S6vb and S6ub. Part S16vc is pitch lag 2 bits / 160 samples, pitch gain 11 bits / 160 samples, and pitch residual
18 bits / 32 samples and 1 bit / 160 samples of S / M mode data give a total of 104 bits / 160 samples, which corresponds to a bit rate of 5.2 kbps. The S / M mode data is sent to the VQ circuit 124 for voice (Speech) and music (Musi).
It is for switching between two types of codebooks (codebooks), which are different for and for c). Part S16vd is high frequency LP
C data 5 bits / 160 samples and high band LPC residual 15 bits / 32 samples gives a total of 80 bits / 160 samples, which corresponds to a bit rate of 4 kbps. The portion S16ua has the same data content as the portions S2ua and S6ua.
ub is the portion S16vb, that is, the portions S6vb and S6.
It has the same data contents as ub. Furthermore, part S16uc
Is the data content common to the above-mentioned part S16vc, and the part S16ud is the data content common to the above-mentioned part S16vd.

The configuration of FIGS. 1 and 3 for obtaining the above bit stream is summarized as shown in FIG.

In FIG. 5, the input terminal 11 is shown in FIG.
This corresponds to the input terminal 101 of FIG. 3, and this is the LPF 1 of FIG.
02, sampling frequency converter 103, subtractor 10
6, sent to the band division circuit 12 corresponding to the BPF 107 and the like, and divided into the low frequency side and the high frequency side. Band division circuit 1
The low frequency side signal from 2 is sent to the 2k encoding unit 21 and the common partial encoding unit 22 corresponding to the configuration of FIG. The common partial encoding unit 22 is the LPC analysis / quantization unit 13 of FIG.
0, or almost equivalent to the LPC analysis / quantization unit 310 in FIG. 3, and the pitch extraction unit in the sine wave analysis coding unit in FIG. 3 and the pitch analysis circuit 115 in FIG. 1 are also included in the common partial coding unit 22. You can also

The low-frequency side signal from the band division circuit 12 is also sent to the 6k coding unit 23 and the 12k coding unit 24. The 6k encoding unit 23 includes the circuits 111 to 116 shown in FIG.
, And the 12k encoding unit corresponds to the circuit 117 of FIG.
118, 122 to 128 are substantially equivalent.

The high band side signal from the band dividing circuit 12 is sent to the high band 4k encoding unit 25. High frequency 4k encoding unit 25
Substantially correspond to the circuits 161-164, 171, 172 of FIG.

The relationship between the bit stream output from the output terminals 31 to 35 of FIG. 5 and the respective portions of FIG. 4 will be described. The data of the portion S2ve or S2ue in FIG. 4 is output from the 2k encoding unit 21 through the output terminal 31, and the portion S2va (= S6va = S16va in FIG. 4 is output from the common partial encoding unit 22 through the output terminal 32. ) Or S2ua (= S6ua =
The data of S16ua) is output. Further, the portion S6vb (= in FIG. 4 is transmitted from the 6k encoding unit 23 via the output terminal 33.
The data of S16vb) or S6ub (= S16ub) is output. Further, the data of the portion S16vc or S16uc of FIG. 4 is output from the 12k encoding unit 24 via the output terminal 34, and the high frequency 4k encoding unit 25 outputs the data of the portion S16vd or S16ud of FIG. 4 via the output terminal 35. Data is output.

Generalizing the above-described technique for realizing scalability, the first coded signal obtained by performing the first coding on the input signal and the first coded signal on the input signal are obtained. The second coded signal obtained by performing the second coding independent of the first coding and having a part common to only a part of the coding and a part not common to the first coding is multiplexed. At this time, the first coded signal is multiplexed with a signal of the second coded signal excluding a portion common to the first coded signal.

This makes it possible to maximize the share of what can be shared even if the encoding systems are essentially different, thereby providing scalability.

Next, a more specific operation of each unit shown in FIGS. 1 and 2 will be described.

First, as shown in FIG. 6A, the frame interval, so-called frame interval, is set to N samples,
For example, a case will be described where 160 samples are used and analysis is performed once per frame.

The pitch analysis center is t = kN (where k = 0,
1,2,3, when the ...), the LPC prediction residuals from the LPC inverted filter 111, a vector of t = kN-N / 2 ~ kN + N / 2 number of dimensions consisting of components present in the N X And t = kN-N / 2- which is shifted L samples forward of the time axis.
The N-dimensional vector consisting of components of L ~ kN + N / 2- L as X _L, ‖ X -g X _L ‖ ² searches the L = L _opt to be the smallest, in the L _opt in this section The optimum pitch lag L ₁ is set. Alternatively, in order to avoid a rapid change in pitch, the value after pitch tracking may be set as the optimum pitch lag L ₁ .

Next, with respect to this optimum pitch lag L ₁ ,

[0077]

(Equation 1)

The set of g _i that minimizes

[0079]

(Equation 2)

Solving for the pitch gain vector g ₁
Ask for. The codebook index of vector-quantized pitch gain vector g ₁ is g ₁ .

Next, in order to further improve the prediction accuracy, t =
Consider putting the analysis center on (k-1 / 2) N. At this time,
It is assumed that the pitch lag and the pitch gain at t = kN and (k-1) N are obtained in advance.

In the case of a voice signal, its fundamental frequency is considered to change gently, so that the pitch lag L (kN) at t = kN and the pitch lag L ((k- 1)
It is considered that there is no significant change between N) and N, and that change is also linear, so pitch lag L ((k-1 / 2) when t = (k-1 / 2) N It is possible to limit the possible values of N). In this example, the following is done.

L ((k-1 / 2) N) = L (kN) = (L (kN) + L ((k-1) N)) / 2 = L ((k-1) N) Which value is adopted is determined by calculating the power of the pitch residual corresponding to each lag.

That is, t = centered on t = (k-1 / 2) N
Let X be a vector of dimension number N / 2 of (k-1 / 2) NN / 4 to (k-1 / 2) N + N / 4, and let L (kN) and (L (kN) + L ((k -1) N)) / 2, L ((k-1) N)
The vector of the number of dimensions N / 2 delayed respectively by X ₀ ⁽⁰⁾ ,
X ₁ ⁽⁰⁾ , X ₂ ^(0), and these respective vectors X ₀ ⁽⁰⁾ , X ₁
⁽⁰⁾ and X ₂ ⁽⁰⁾ each vector near X ₀ ^(-1) ,
X ₀ ⁽¹⁾ , X ₁ ^(-1) , X ₁ ⁽¹⁾ and X ₂ ^(-1) , X ₂ ⁽¹⁾ . Also, each of these vectors X ₀ ⁽ⁱ⁾ , X ₁ ⁽ⁱ⁾ , X ₂ ⁽ⁱ⁾
For the nuclear pitch gains g ₀ ⁽ⁱ⁾ , g ₁ ⁽ⁱ⁾ and g ₂ ⁽ⁱ⁾ corresponding to (where i = −1, ₀ , ₁ ⁾ ,

[0085]

(Equation 3)

The lag with respect to the smallest one of the three, D _j , is the optimum lag L ₂ at t = (k-1 / 2) N, and the pitch gain g _j ^{(i) at} that time (where i =- 1, 0, 1) are vector-quantized and then the pitch gain is obtained. There are three possible values for L ₂ , which are the current and past L ₁ values.
Therefore, a flag representing an interpolation scheme, rather than a straight value, may be transmitted as an interpolation index. Further, when it is determined that either L (kN) or L ((k-1) N) is 0, that is, there is no pitch and the pitch prediction gain cannot be obtained, L ((k-1 / 2) N ) Above (L (kN) + L
((k-1) N)) / 2 is excluded.

[0087] Thus, when the number of dimensions of the vector X used for the pitch lag calculated half N / 2, t = it kN is L _k when the analysis center may be used as it is, X
Even though the pitch gain is obtained when the number of dimensions is analyzed by N, the gain must be calculated again and the data must be transmitted. Here, to reduce the number of bits,

[0088]

(Equation 4)

Of the elements (g ₀ , g ₁ , g ₂ ) of the vector g ,
Since g ₁ is the largest and g ₀ and g ₂ are close to 0 or vice versa and the vector g has a strong correlation among the three points, the vector g _1d has a variance larger than that of the original vector g. It is expected to be smaller and can be quantized with a smaller number of bits.

Accordingly, there are five pitch parameters L ₁ , g ₁ , L ₂ , g ₂ , g _1d to be transmitted in one frame.

Next, FIG. 5B shows the phase of the LPC coefficient interpolated at a rate of 8 times the frame frequency. This LPC coefficient is the LPC inverse filter 111 of FIG.
2 is used for the prediction residual calculation, and is also used for the LPC synthesis filters 215 and 225 and the pitch spectral post filters 216 and 226 of FIG.

Next, vector quantization of the pitch residual obtained from the pitch lag and the pitch gain will be described.

In order to easily and accurately perform the perceptual weighting of the vector quantization, the pitch residual is windowed by 50% overlap, the MDCT transformation is performed, and the weighting vector quantization is performed in this region. The conversion length at this time is arbitrary, but in consideration of the following points, a considerably small dimension is used in this example.

(1) The large-dimensional vector quantization requires an enormous amount of calculation, and there is no choice but to perform splitting or rearrangement in the MDCT domain. (2) When splitting, it is very difficult to perform precise bit allocation between the split bands. (3) When the number of dimensions is not a power of 2, the high-speed arithmetic method of MDCT using FFT cannot be used.

This time, the frame length is 20 msec (= 160
160/5 = 3 because it is for sample / 8 kHz)
Since 2 = ²⁵ , the MDCT transform size is set to 64 in consideration of 50% overlap and the above points (1) to (3) are solved.

The framing state is as shown in FIG.

That is, in FIG. 6C, 2
Pitch residual r in the frame of 0 msec = 160 samples
_p (n) (where n = 0,1, ..., 191, where n = 160, ..., 191 is
(Meaning 0, ..., 31 of the next frame) is divided into 5 subframes, and the pitch residual r _pi (n) of the i-th (i = 0,1, ..., 4) subframe of the 5 subframes (However, n = 0,1,
, 31) is set as r _pi (n) = r _p (32i + n). For the pitch residual r _pi (n) of this subframe, M
Window function w (n) that can cancel aliasing of DCT
MDCT conversion is applied to w (n) · r _pi (n) obtained by multiplying by. As the window function w (n), for example, w (n) = √ (1- (cos2π (n + 0.5)) / 64) may be used.

Since the conversion length of the MDCT conversion process is 64 (= 2 ⁶ ), it can be calculated as follows using FFT.

(1) x (n) = w (n) _.rpi (n) .exp ((-2π
j / 64) (n / 2)). (2) x (n) is 64-point FFT processed and this is y (k)
And (3) Take the real part of y (k) ・ exp ((-2πj / 64) (k + 1/2) (1/2 + 64/4)), and use this as the MDCT coefficient c _i (k) (however, , K = 0,1,
…, 31).

Next, the MDCT coefficient c of each subframe
Vector quantization is performed on _i (k), and the weighting at this time will be described.

Letting the pitch residual r _pi (n) be the vector r _i , the distance D after synthesis is

[0102]

(Equation 5)

Here, M is considered to diagonalize H ^t H (where H ^t is a transposed matrix of H) because of its nature.

[0104]

(Equation 6)

Here, h _i is taken as the frequency response of the synthesis filter. Therefore,

[0106]

(Equation 7)

As described above, when h _k is used as it is for weighting the quantization of c _i (k), the noise after synthesis becomes flat, that is, so-called 100% noise shaping. Control so that it becomes a noise of similar shape.

[0108]

(Equation 8)

Note that h _i ² and w _i ² are synthesis filters H (z)
And perceptual weighting filter W (z)

[0110]

(Equation 9)

The FFT power spectrum of the impulse response of is obtained.

Here, α _ij is the LPC coefficient corresponding to the i-th subframe, and is obtained from the interpolated LSP coefficient. That is, the LSP obtained by the analysis of the previous frame
₀ (j) and LSP ₁ (j) of the current frame are internally divided, and in this example, the LSP of the i-th subframe is

[0113]

(Equation 10)

LSP ⁽ⁱ⁾ (j) is obtained as Then L
Obtain α _ij by SP → α conversion.

With respect to H and W thus obtained, W '= WH is newly set and used as a distance measure in vector quantization.

Vector quantization is performed by shape and gain vector quantization. Optimal encoding and decoding conditions at the time of learning will be described.

Let s be the shape codebook at a certain point in learning, g be the gain codebook, x be the MDCT coefficient in the training input, that is, in each subframe, and W ′ be the weight in that subframe. The distortion power D ^{2 at this time} is defined by the following equation.

D ² = ‖W ′ ( x− g s ) ‖ ² The optimum encoding condition is to select (g, s ) that minimizes this D ² .

[0119]

[Equation 11]

Therefore, first of all, as the first step,
About shape code book,

[0121]

(Equation 12)

Search for s _opt that maximizes, and for the gain codebook, for this s _opt ,

[0123]

(Equation 13)

It suffices to search g _opt closest to.

Next, the optimum decoding condition is obtained.

As the second step, regarding the shape codebook, at some point during learning, the shape codebook is
A set of encoded x in _{s x k (k = 0,} ..., N-
In contrast to 1), the total distortion E _{s at} this time is

[0127]

[Equation 14]

Therefore, s that minimizes this is

[0129]

(Equation 15)

From

[0131]

(Equation 16)

Is calculated.

For the gain codebook, a set x _{k of} x encoded in gain codebook g (weight W
′ _K , shape s _k ), the sum of distortion E _g is

[0134]

[Equation 17]

Is as follows.

A shape and gain codebook can be obtained by GLA (generalized Lloyd algorithm) while repeatedly obtaining the first and second steps.

[0137] In this embodiment, in order to emphasize the noise when the signal level is low, W'not itself, W'was weighted levels (inverse of) / ‖ x ‖ performing learning using There is.

In this way, vector quantization is performed on the MDCT of the pitch residual using the codebook created in this way, and its index is LPC (actually LSP),
By transmitting together with the pitch and the pitch gain and performing the inverse vector quantization, the pitch, and the LPC synthesis on the decoding side, the reproduced sound can be obtained, but in this example, since the operation with a higher rate is possible, The aforementioned pitch lag,
As the frequency of pitch gain calculation is improved, the pitch residual MD
This is dealt with by making the CT vector quantization multistage.

An example is shown in FIG. The number of stages here is two, which is sequential multi-stage vector quantization, but the input of the second stage is the decoding result of the first stage is L ₂ ,
_Used as subtracted from the highly accurate pitch residual made from g ₂ and g _1d . That is, the representative vector or the dequantized output after vector-quantizing the output from the MDCT circuit 113 of the first stage by the VQ circuit 114 is the inverse MDC.
The result of the inverse MDCT processing by the T circuit 113a is subtracted by the subtracter 1
28 ', and subtracted from the residual of the second stage (output from the pitch inverse filter 122 of FIG. 1). This subtractor 12
The output from 8'is MDCTed by the MDCT circuit 123 '.
It is processed and quantized by the VQ circuit 124. This can have an equivalent configuration as shown in FIG. 7B in which the inverse MDCT of the first stage is not performed, and this configuration in FIG. 1 is used.

When performing decoding using both the MDCT coefficient indexes I _dx V _q1 and I _dx V _q2 on the decoder side in FIG. 2, the sum of the results of inverse vector quantization of I _dx V _q1 and I _dx V _q2. Inverse MDCT and overlap addition are performed, and then pitch synthesis and LPC synthesis are performed to obtain a reproduced sound. Naturally, the pitch lag and pitch gain update frequency at the time of pitch synthesis are doubled in the case of only one stage, and in the present application, a pitch synthesis filter that switches every 80 samples is driven.

Next, the post filters 216 and 226 on the decoder side in FIG. 2 will be described.

The post filters 216 and 226 realize the post filter characteristic p (Z) by cascading pitch enhancement, high frequency enhancement and spectrum enhancement filters.

[0143]

(Equation 18)

In this equation, g _i and L are the pitch gain and pitch lag obtained by pitch prediction, and ν is a parameter indicating the degree of pitch enhancement (for example, ν =
0.5). Further, ν _b is high-frequency emphasis (for example, ν _b = 0.
4), ν _n and ν _d are the degree of emphasis of the spectrum (for example, ν _n
= 0.5, ν _d = 0.8).

Next, the output s (n) of the LPC synthesis filter,
Gain correction is performed on the output s _p (n) of the post filter. The coefficient k _{adj at} this time is

[0146]

[Equation 19]

However, k _adj is not fixed within the frame, but is changed for each sample after passing through the LPF. For example, 0.1 is used as p.

K _adj (n) = (1-p) k _adj (n-1) + pk _adj Next, in order to smooth the frame connection, two pitch enhancement filters are prepared as follows, The final output is the result of crossfading.

[0149]

(Equation 20)

[0150]

(Equation 21)

The final output s _out (n) for the outputs s _p0 (n) and s _p (n) of the post filter thus constructed is s _out (n) = (1-f (n)). Let s _p0 (n) + f (n) · s _p (n). Here, f (n) is a window as shown in FIG. 8, for example. 8A shows a low rate and FIG. 8B shows a high rate, and the 80 sample width window in FIG. 8B is used twice when combining 160 samples and 20 msec.

Next, the VQ (vector quantization) circuit 124 on the encoder side in FIG. 1 will be described.

This VQ circuit 124 has two types of codebooks (codebooks) that are different from each other for speech and music, and these two types of codebooks are used as input signals. The switching is selected accordingly.

That is, when the structure of the quantizer is determined in the quantization of the voice and musical tone signals, the codebook of this quantizer becomes the optimum one in the nature of the voice and musical tone used during learning. Therefore, when both are learned together, if the properties of both greatly differ, the codebook after learning will have the average properties of both. Therefore, when a quantizer is configured with one codebook, it is expected that its performance or average S / N will not be so high.

Therefore, in the present example, the code amount created by using the respective learning data for a plurality of signals having different properties as described above is switched to improve the performance of the quantizer.

FIG. 9 shows such two kinds of codebooks C.
1 shows a schematic configuration of a vector quantizer having B _A and _{C B B.}

In FIG. 9, the input signal supplied to the input terminal 501 is sent to the vector quantizers 511 and 512. These vector quantizers 511, 512
Have code lengths (codebooks) CB _A and CB _B , respectively. The representative vector or dequantized output from these vector quantizers 511 and 512 is sent to subtractors 513 and 514, respectively, and the difference from the original input signal is taken, and these respective error components are compared. Sent to 515.
In the comparator 515, the respective error components are compared with each other, and the index which is the quantized output from the vector quantizer 511, 512 having the smaller error is selectively selected by the selector switch 516 and sent to the output terminal 502.

In this case, each vector quantizer 511, 5
The changeover cycle of the changeover switch 516 is made longer than the 12 quantization unit times or cycles. For example, when the quantization unit is a subframe obtained by dividing a frame into eight, the changeover switch 516 is switched in the frame unit.

As an example, assuming that there are codebooks CB _A and CB _{B of} the same size N and the same dimension M, which are learned only by voice and only by tone, respectively, an L-dimensional data consisting of L data of a certain frame is given. Data X is subframe length M (= L
/ N), when the vector quantization is performed, the distortion after the quantization is E _A (k) when the codebook CB _A is used, and the codebook CB
_Let E _B (k) be the case when _B is used. These distortions E _A (k),
E _B (k), as the index i, j was chosen _{respectively, E A (k) = ‖W} k (X - C Ai) ‖ _{E B (k) = ‖W k} (X - C Bj) ‖ Is. In this formula, W _k represents a weighting matrix of a subframe k, C _Ai, C _Bj denote representative vectors corresponding to the codebook CB _A, each index i of the CB _B, j.

Consideration will be given to adopting the optimum codebook for the two distortions thus obtained, based on the total sum of distortions within one frame. The following two methods can be considered as the selection method at this time.

As a first method, in all subframes, quantization is performed using only the codebooks CB _A and CB _B , and the total sum Σ _k E _A (k) of the distortions, Σ _k E _B (k ), And gives the sum of the smaller distortions, codebooks CB _A , C
One of B _B is used over one frame.

A configuration example for realizing the first method is shown in FIG.
Shown in In FIG. 10, parts corresponding to those in FIG. 9 are given the same reference numerals, and a, b, and
The subscripts such as ... Correspond to the subframe k. The codebook CB _A, takes the subtracters 513a distortion of each sub-frame is obtained, 513b, · · ·, the frame total of the output from 513n in adder 517, each sub-frame for the codebook CB _B An adder 518 obtains the total sum of distortions for each frame, and a comparator 515 compares them to obtain a control signal or a selection flag for codebook switching from a terminal 503.

Next, the second method compares distortions E _A (k) and E _B (k) for each sub-frame, and judges the comparison result over all sub-frames in the frame. By doing so, the codebook is switched and selected.

FIG. 11 shows an example of implementation of this second method.
In FIG. 11, the output from the comparator 516 that performs comparison for each sub-frame is sent to the judgment logic 519, and judgment processing is performed by, for example, a majority decision so that a 1-bit codebook switching selection flag is obtained from the terminal 503. ing.

Note that this selection flag indicates that the S / M
It is transmitted as (voice / tone) mode data.

In this way, a plurality of signals having different characteristics can be efficiently quantized by one quantizer.

Next, the frequency conversion processing by the FFT circuit 161, the frequency shift circuit 162, and the inverse FFT circuit 163 of FIG. 1 will be described.

This frequency conversion processing includes a band extraction step of extracting at least one band of the input signal, an orthogonal conversion step of converting the extracted signal of at least one band into a signal on the frequency axis, and an orthogonal conversion. Shift step for shifting the shifted signal on the frequency axis (to another position, to another band), and inverse orthogonal transformation for transforming the signal shifted on the frequency axis to a signal on the time axis And the process.

FIG. 12 is a diagram showing in more detail the configuration for the above frequency conversion, and the parts corresponding to those in FIG. 1 are designated by the same reference numerals. In FIG. 12, the input terminal 1
01 is supplied with a wideband audio signal having a component of 0 to 8 kHz at 16 kHz sampling, for example. Of the wideband audio signals from the input terminal 101, for example, 0-3.
LPF (low-pass filter) with 8 kHz as the low frequency side signal
The signal is separated by 102, and the component obtained by subtracting the low frequency side signal from the original wide band signal by the subtractor 151 is separated as the high frequency side signal. The low-frequency side signal and the high-frequency side signal are processed independently.

The high-frequency side signal obtained here has a frequency width of 4.5 kHz from 3.5 kHz to 8 kHz, which is slightly left even through the LPF 102, but since it is down-sampled for signal processing, We need to narrow it down to 4 kHz. In this example, 0.5 kHz from 7.5 kHz to 8 kHz
BPF (bandpass filter) 107 or LP
Cut by F.

Next, for example, FFT (Fast Fourier Transform) is performed as frequency conversion to the low frequency side. Prior to this, the frame division circuit 108 makes the number of samples a power of 2, for example, (A) of FIG. As shown in FIG. 5, each sample is divided into 512 samples. However, in order to facilitate later signal processing, it is advanced every 80 samples.

Next, the Hamming windowing circuit 109
A Hamming window of 320 samples in length is applied. This sample number 320 is advanced by 80 samples at the time of the frame division so that four waveforms can be added in an overlapping manner as shown in FIG. Therefore, it is set to four times 80.

Next, the data of 512 samples in length is subjected to FFT processing by the FFT circuit 161, and converted into data on the frequency axis.

Next, the frequency shift circuit 162 shifts or moves the data to another position or another band on the frequency axis. The principle of lowering the sampling frequency by the shift on the frequency axis is to move the high frequency side signal in the shaded area of (A) to the low frequency side as shown in (B), as shown in FIG. Down-sampling is performed as shown in (C). 14 (A) to (B)
The components that are folded back around fs / 2 when moving on the frequency axis are in opposite movement directions. As a result, the sampling frequency can be reduced to fs / n if the subband is within fs / 2n.

The frequency shift circuit 162 shown in FIG.
As shown in, the process of shifting or moving the shaded data corresponding to the high frequency side of the data on the frequency axis to the position or band on the frequency axis corresponding to the low frequency side may be performed. Specifically, with respect to 512 data on the frequency axis obtained by performing FFT processing on the data on the time axis of 512 samples, 127 pieces of data from the 113th to 239th positions are placed at positions 1 to 127. Alternatively, it is moved to the band, and 127 pieces of data from the 273rd to the 399th are moved to the 395th to 511th positions or bands. At this time, it is important not to move the 112th data on the frequency axis to the 0th position. This is because in the frequency domain signal, the 0th is the DC component,
Since there is no phase component, the data at this position must be a real number, and the frequency component, which is generally a complex number, cannot be included here. Also, 256 representing fs / 2
The second data (generally N / 2nd data) is also invalid and is not used, that is, 0 to 4 in this case.
Strictly speaking, the range of kHz is a range expressed as 0 <f <4 kHz.

Next, the inverse FFT circuit 163 performs inverse FFT.
By processing, the data on the frequency axis is returned to the signal on the time axis. In this case, a signal on the time axis is obtained every 512 samples. The time-axis signals for every 512 samples are overlapped by 80 samples by the superposition addition circuit 166 as shown in FIG. 7B, and the overlapping portions are added.

Since the signal obtained from the superposition addition circuit 166 is limited to 0 to 4 kHz by 16 kHz sampling, it is downsampled by the downsampling circuit 164. 8 frequency shifted by this
A signal of 0 to 4 kHz can be obtained by kHz sampling, and this signal is taken out through the output terminal 169,
It is sent to the LPC analysis / quantization unit 130 and the LPC inverse filter 171 in FIG.

Next, the restoration process on the decoding side is shown in FIG.
It can be realized by the configuration shown in.

The structure of FIG. 16 corresponds to the structure of the upsampling circuit 233 and the subsequent parts of FIG. 2 described above, and corresponding parts are designated by the same reference numerals. However, although the upsampling process is performed before the FFT process in FIG. 2, the upsampling process is performed after the FFT process in the example of FIG. 16.

In FIG. 16, a terminal 241 is supplied with a high-frequency side signal, such as the output signal from the high-frequency side LPC synthesis filter 232 in FIG. 2, which is shifted to 0 to 4 kHz by 8 kHz sampling. It

This signal is divided by the frame division circuit 242 into a signal having a frame length of 256 samples and an advance amount of 80 samples. This is for the same reason as the frame division on the encoding side, but since the sampling frequency is 1/2, the number of samples is also 1/2. Also, the Hamming window application circuit 243 applies a Hamming window having a length of 160 samples to the signal from the frame division circuit 242, similarly to the encoding side (however, the number of samples is 1/2).

Next, the FFT circuit 234 sets the length 256.
The FFT processing is performed on the sample, and the signal on the time axis is converted to the signal on the frequency axis. In the next up-sampling circuit 244, the frame length is effectively reduced to 21 by performing the zero padding processing as shown in FIG.
From 6 samples to 512 samples. This is shown in FIG.
Of (C) to (B).

Next, the frequency shift circuit 235 shifts or moves the data to another position or another band on the frequency axis to shift the frequency by +3.5 kHz. This corresponds to the conversion from (B) to (A) in FIG.

The signal on the frequency axis thus obtained is subjected to the inverse FFT processing by the inverse FFT circuit 236 to be returned to the signal on the time axis. This inverse FFT circuit 236
The signal from is 3.5kHz from 16kHz sampling
It is 7.5 kHz.

In the next superposition addition circuit 237, the length 512
For each frame of samples, 80 samples are overlapped and added together to restore a continuous time axis signal. The high frequency side signal thus obtained is added to the low frequency side signal by the adder 228 and is taken out from the output terminal 229.

In such frequency conversion,
The specific numerical value is not limited to the above example, and the number of bands for shifting is not limited to one.

For example, as shown in FIG. 17, when the narrow band signal is set to 300 Hz to 3.4 kHz and the wide band signal is set to 0 to 7 kHz by 16 kHz sampling, the low band 0 to 300 Hz not included in the narrow band and the high band are included. Of the 3.4kHz to 7kHz range, if you move the high range to 300Hz to 3.9kHz and collect it so that it contacts the low range, it becomes a signal of 0 to 3.9kHz, which is the same as above. 1 / sampling frequency fs
It can be 2 to 8 kHz.

If this is generalized, when a wideband signal and a narrowband signal that fits inside the wideband signal are multiplexed,
The high band side component of the rest of the wide band signal obtained by subtracting the narrow band signal is shifted to the low band side to lower the sampling rate.

As described above, a subband of an arbitrary frequency can be created from an arbitrary frequency, processing can be performed at a sampling frequency twice the sum of the frequency widths, and the application can be flexibly handled.

Further, when the quantization error is large at a low bit rate, the aliasing noise, which should have been generated in the vicinity of the division frequency when the QMF is generally used, can be avoided by the frequency conversion method.

The present invention is not limited to the above-described embodiment, and for example, the configuration of the speech encoding side (encoding side) of FIG. 1 and the speech decoding side (decoding side) of FIG. Regarding the configuration, each unit is described as hardware, but it is also possible to realize it by a software program using a so-called DSP (digital signal processor) or the like. Also, instead of the vector quantization, the data of a plurality of frames may be collectively subjected to matrix quantization. Furthermore, the speech encoding method and the decoding method to which the present invention is applied are not limited to the above encoding / decoding method, but can be applied to various speech encoding / decoding methods, and can be used for transmission and Not limited to recording and playback,
Of course, it can be applied to various applications such as pitch conversion, speed conversion, regular speech synthesis, or noise suppression.

[0192]

As is apparent from the above description, according to the present invention, the input signal is band-divided, and the signals for each divided band are encoded differently according to the signal characteristics of each band. By performing the above, it is possible to operate the decoders having different rates, and it is possible to perform the efficient and optimal coding for each band, thereby improving the coding efficiency.

A short-term prediction is performed on the signal on the low frequency side of the band obtained by the band division to obtain a short-term prediction residual, a long-term prediction is performed on the obtained short-term prediction residual, and the obtained long-term prediction is performed. By orthogonally transforming the residual, high coding efficiency is obtained and the quality of reproduced sound is excellent.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a basic configuration of an audio signal encoding device to which an embodiment of an audio encoding method according to the present invention is applied.

FIG. 2 is a block diagram showing a basic configuration of a speech signal decoding device.

FIG. 3 is a block diagram showing the configuration of another audio signal encoding device.

FIG. 4 is a diagram for explaining scalability of a bit stream of encoded data to be transmitted.

FIG. 5 is a block diagram schematically showing an entire encoding side system to which the present invention can be applied.

FIG. 6 is a diagram for explaining a cycle and a phase relationship of main operations of encoding and decoding.

FIG. 7: MDCT (Modified Discrete Cosine Transform)
It is a figure which shows the structural example of vector quantization of a coefficient.

FIG. 8 is a diagram showing an example of a window function applied to a post filter output.

FIG. 9 is a diagram showing an example of a vector quantization device having two types of codebooks.

FIG. 10 is a diagram showing a specific example of a vector quantization device having two types of codebooks.

FIG. 11 is a diagram showing another specific example of the vector quantization device having two types of codebooks.

FIG. 12 is a block diagram showing a configuration on the encoder side of frequency conversion.

FIG. 13 is a block diagram showing a configuration of a frequency conversion decoder side.

FIG. 14 is a diagram for explaining frame division and superposition addition processing.

FIG. 15 is a diagram showing a shift process of data on the frequency axis.

FIG. 16 is a block diagram showing a configuration of a frequency conversion decoder side.

FIG. 17 is a diagram showing another example of frequency shift on the frequency axis.

[Explanation of symbols]

111, 171 LPC inverse filter 112, 122 Pitch inverse filter 113, 123 MDCT (Modified Discrete Cosine Transform) circuit 114, 124 VQ (Vector Quantization) circuit 115, 125 Pitch analysis circuit 116, 118, 126 Pitch gain VQ circuit 130, 180 LPC analysis / quantization unit

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification number Reference number within the agency FI Technical indication location H03M 7/30 9382-5K H03M 7/30 A (72) Inventor Kazuyuki Iijima Kitashinagawa, Shinagawa-ku, Tokyo 6th-7th 35th Sony Corporation

Claims (10)

[Claims] 1. A band division step of band-dividing an input signal, and a step of performing different encoding on the divided signals for each band according to the signal characteristics of each band. Signal coding method. 2. The signal encoding method according to claim 1, wherein said band dividing step divides at least a wide band input voice signal into a telephone band signal and a signal on a higher band side than the telephone band. . 3. The signal coding according to claim 1, wherein the signal on the low band side of the band-divided band is coded by combining short-term predictive coding and orthogonal transform coding. Method. 4. A short-term prediction step of performing a short-term prediction on a signal on the low-frequency side of the band-divided band to obtain a short-term prediction residual, and performing a long-term prediction on the obtained short-term prediction residual. The signal coding method according to claim 1, further comprising a long-term prediction step of obtaining a long-term prediction residual by performing an orthogonal transformation step of orthogonally transforming the obtained long-term prediction residual. 5. The signal coding method according to claim 4, further comprising the step of perceptually weighting and quantizing the orthogonal transform coefficient obtained in the orthogonal transform step on a frequency axis. 6. The signal code according to claim 4, wherein the orthogonal transform step uses MDCT (Modified Discrete Cosine Transform) to shorten the transform length and to make it a power of two. Method. 7. The signal coding method according to claim 1, wherein short-term predictive coding is performed on the signal on the high frequency side of the band that has been divided. 8. A band dividing means for dividing an input signal into bands, and an encoding means for performing different encoding on the divided signals for each band according to the signal characteristics of each band. And a signal encoding device. 9. The signal coding apparatus according to claim 8, wherein said band dividing means divides a wide band input voice signal into at least a telephone band signal and a signal on a higher band side than the telephone band. . 10. The encoding means includes short-term prediction means for performing short-term prediction on a signal on the low frequency side of the band-divided band to obtain a short-term prediction residual, and the calculated short-term prediction residual. 9. The signal coding apparatus according to claim 8, further comprising: long-term prediction means for performing long-term prediction to obtain a long-term prediction residual, and orthogonal transformation means for orthogonally transforming the obtained long-term prediction residual.