JPH09127997A - Voice coding method and device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To simply switch the bit rate by providing multiple coding sections for the vector quantization of the time base wave-form, reference-inputting the quantization error of the (N-1)th stage for the coding of the Nth stage, and selecting the quantization output of each stage to switch the bit rate. SOLUTION: The second coding section 120 having a CELP coding structure is constituted of multiple vector quantization process sections, e.g. two coding sections 1201 , 1202 . When the vector quantization by the closed loop search of the first stage is finished, the quantization error of the (N-1)th stage is used as the reference input for the quantization of the Nth stage (2<=N). The calculation quantity is reduced. The case that both index outputs of two coding sections 1201 , 1202 are used and the case that only the output of the coding section 1201 of the first stage is used are switched, and the bit number can be simply switched.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice encoding method and apparatus for dividing an input voice signal into blocks and performing an encoding process in units of the divided blocks.

[0002]

BACKGROUND OF THE INVENTION There are various coding methods for performing signal compression 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. Are known. This encoding method is roughly classified into encoding in the time domain, encoding in the frequency domain, and analysis-synthesis encoding.

[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. Further, as an example of high-efficiency encoding of a speech signal or the like, there is also code-excited linear prediction (CELP) encoding using vector quantization by closed loop search of an optimum vector using an analysis method by synthesis.

[0004]

By the way, as an example of the high-efficiency coding of the above speech signal, consider a system in which, for example, code excitation linear predictive coding is used and the coded speech signal is transmitted to a transmission line. In this case, in the code excitation linear predictive coding, straight vector quantization is performed, and since a conjugate codebook or the like is used, for example, even if it is desired to switch the transmission bit rate according to the capacity of the transmission line, Due to its structure, it is very difficult. Also,
When the bit rate on the encoding side (encoder) and the bit rate on the decoding side (decoder) are different, that is, for example, the bit rate in the encoder is Mkbps and the bit rate in the decoder is Nkbps. In such a case, the decoder corresponding to Nkbps cannot support the encoded data string of Mkbps supplied from the encoder.

Therefore, the present invention has been made in view of such a situation, and in the case where the transmission bit rate can be easily switched and the bit rate is different between the encoding side and the decoding side. It is an object of the present invention to provide a speech coding method and apparatus capable of easily generating a coded data string that can be easily dealt with on the decoding side.

[0006]

A speech encoding method and apparatus according to the present invention encodes an input speech signal in units of blocks divided on a time axis, and uses an analysis method by synthesis. There are multiple stages of encoding for performing vector quantization of the time-axis waveform by the closed loop search of the optimum vector. Among them, when encoding the Nth stage, the quantization error of the (N-1) th stage is used as the reference input. , The above-mentioned problem is solved by selecting the quantized output by the encoding of each stage and switching the bit rate.

That is, according to the speech encoding method and apparatus of the present invention, the output bit rate can be switched by selecting all or a part of the encoded outputs of the multistage configuration. Further, since the encoded outputs of the multi-stage configuration are output together, even if the bit rates on the encoding side and the decoding side are different, on the decoding side, for example, all of these encoded outputs or If a part of them is selected, it becomes possible to easily obtain a coded output with a desired bit rate.

[0008]

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

Here, the basic idea of the speech signal coding apparatus of FIG. 1 is that the short-term prediction residual of the input speech signal, for example, LP.
Sine wave analysis (sinu
soidal analysis) First encoding unit 11 that performs encoding, for example, harmonic coding
0 and a second coding unit 120 that performs coding by waveform coding that performs phase transmission on the input speech signal, and is the first for coding the voiced sound (V: Voiced) portion of the input signal. Of the input signal unvoiced sound (UV: Unvoic
The second encoding unit 120 is used for encoding the portion (ed).

The first encoding unit 110 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. The second encoding unit 120 employs, for example, a configuration of code excitation linear prediction (CELP) encoding using vector quantization by closed loop search of an optimal vector using an analysis method based on synthesis.

In the example of FIG. 1, the audio signal supplied to the input terminal 101 is sent to the LPC inverse filter 111 and the LPC analysis / quantization unit 113 of the first encoding unit 110. The LPC coefficient or the so-called α parameter obtained from the LPC analysis / quantization unit 113 is sent to the LPC inverse filter 111, and the LPC inverse filter 111 extracts a linear prediction residual (LPC residual) of the input audio signal. . Also, a quantized output of an LSP (line spectrum pair) is extracted from the LPC analysis / quantization unit 113 and sent to the output terminal 102 as described later. The LPC residual from LPC inverse filter 111 is sent to sine wave analysis encoding section 114. In the sine wave analysis encoding unit 114, pitch detection and spectrum envelope amplitude calculation are performed, and a V (voiced sound) / UV (unvoiced sound) determination unit 1 is performed.
15 is used to determine V / UV. The spectrum envelope amplitude data from the sine wave analysis encoding unit 114 is sent to the vector quantization unit 116. The codebook index from the vector quantization unit 116 as the vector quantization output of the spectrum envelope is
The output from the sine wave analysis encoding unit 114 is sent to the output terminal 104 via the switch 118. Also, the V / UV determination unit 115
Is output to the output terminal 105 and is also sent as a control signal for the switches 117 and 118. In the case of the above-mentioned voiced sound (V), the index and the pitch are selected and each output terminal is output. 103 and 104 respectively.

The second coding unit 120 of FIG. 1 has a CELP (code excitation linear prediction) coding structure in this example, and outputs the output from the random codebook 121 by a weighting synthesis filter 122. The weighted speech obtained by the synthesis processing is sent to the subtractor 123, the speech signal supplied to the input terminal 101 is taken out as an error from the speech obtained through the auditory weighting filter 125, and this error is calculated by the distance calculation circuit. 12
4 to calculate the distance, and search for a vector that minimizes the error in the noise codebook 121 by using a closed-loop search using an analysis by synthesis method. Vector quantization is performed. This CELP coding is used for coding the unvoiced sound portion as described above,
The codebook index as UV data from No. 1 is extracted from the output terminal 107 via a switch 127 that is turned on when the V / UV determination result from the V / UV determination unit 115 is unvoiced (UV).

Next, FIG. 2 shows a speech signal decoding apparatus to which an embodiment of the speech decoding method according to the present invention is applied, which corresponds to the speech signal decoding apparatus of FIG. It is a block diagram which shows the basic composition of an apparatus.

In FIG. 2, a codebook index as a quantized output of the LSP (line spectrum pair) from the output terminal 102 of FIG. 1 is input to the input terminal 202. The input terminals 203, 204, and 205 include the output terminals 103, 104, and 105 of FIG.
, That is, an index, a pitch, and a V / UV determination output as an envelope quantization output. The input terminal 207 receives an index as UV (unvoiced sound) data from the output terminal 107 shown in FIG.

The index as the envelope quantization output from the input terminal 203 is the inverse vector quantizer 212.
, And is subjected to inverse vector quantization, and the spectrum envelope of the LPC residual is obtained and sent to the voiced sound synthesis unit 211. The voiced sound synthesizer 211 synthesizes an LPC (linear predictive coding) residual of the voiced sound part by sine wave synthesis.
, And the pitch and V / UV determination outputs from 205 are also provided. The LPC residual of the voiced sound from the voiced sound synthesis unit 211 is sent to the LPC synthesis filter 214. Further, the index of the UV data from the input terminal 207 is sent to the unvoiced sound synthesizer 220, and the LPC residual of the unvoiced sound portion is extracted by referring to the noise codebook. This LP
The C residual is also sent to LPC synthesis filter 214. LPC
In the synthesis filter 214, the LPC residual of the voiced part and the LPC residual of the unvoiced part are independently LPC residuals.
A combining process is performed. Alternatively, LPC synthesis processing may be performed on the sum of the LPC residual of the voiced sound part and the LPC residual of the unvoiced sound part. Here, the index of the LSP from the input terminal 202 is sent to the LPC parameter reproducing unit 213, the α parameter of the LPC is extracted, and this is sent to the LPC synthesis filter 214. An audio signal obtained by LPC synthesis by the LPC synthesis filter 214 is extracted from the output terminal 201.

Next, a more specific structure of the speech signal coding apparatus shown in FIG. 1 will be described with reference to FIG. In FIG. 3, parts corresponding to the respective parts in FIG. 1 are given the same reference numerals.

In the speech signal coding apparatus shown in FIG. 3, the speech signal supplied to the input terminal 101 is filtered by a high-pass filter (HPF) 109 to remove a signal in an unnecessary band. After that, the LPC analysis circuit 132 of the LPC (linear predictive coding) analysis / quantization unit 113.
To the LPC inverse filter circuit 111.

The LPC analysis circuit 132 of the LPC analysis / quantization unit 113 obtains a linear prediction coefficient, so-called α parameter, by the autocorrelation method by applying a Hamming window with one block having a length of about 256 samples of the input signal waveform. .
The framing interval, which is the unit of data output, is 160
It is about a sample. When the sampling frequency fs is, for example, 8 kHz, one frame interval is 20 for 160 samples.
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 quantized into a matrix or vector by the LSP quantizer 134. At this time, vector quantization may be performed after obtaining an inter-frame difference, or matrix quantization may be performed on a plurality of frames at once. Here, 20 msec is defined as one frame, and LSP parameters calculated every 20 msec are combined for two frames, and are subjected to matrix quantization and vector quantization.

The quantized output from the LSP quantizer 134, that is, the index of the LSP quantizer is the terminal 102.
And the quantized LSP vector is sent to the LSP interpolation circuit 136.

The LSP interpolation circuit 136 has the above-mentioned 20 msec.
Alternatively, the LSP vector quantized every 40 msec is interpolated to make the rate eight times higher. That is, 2.5 mse
The LSP vector is updated every c. This is because when the residual waveform is analyzed and synthesized by the harmonic encoding / decoding method, the envelope of the synthesized waveform becomes a very smooth and smooth waveform.
This is because an abnormal sound may be generated if it changes abruptly every msec. 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, where the LPC inverse filter 111
Inverse filtering is performed using the α parameter updated every 2.5 msec to obtain a smooth output. An output from the LPC inverse filter 111 is output to an orthogonal transform circuit 145 of a sine wave analysis encoding unit 114, specifically, for example, a harmonic encoding circuit,
(Discrete Fourier transform) circuit.

The α parameter from the LPC analysis circuit 132 of the LPC analysis / quantization unit 113 is sent to the perceptual weighting filter calculation circuit 139 to obtain data for perceptual weighting. Vector quantizer 116 and second encoding unit 12
0 and a synthesis filter 122 with a perceptual weight.

A sine wave analysis coding unit 114 such as a harmonic coding circuit analyzes the output from the LPC inverse filter 111 by a harmonic coding method. That is, pitch detection, calculation of the amplitude Am of each harmonic, determination of voiced sound (V) / unvoiced sound (UV) are performed, and the number of the envelopes or amplitudes Am of the harmonics that change with the pitch is dimensionally converted to a constant number. .

In the concrete example of the sine wave analysis coding unit 114 shown in FIG. 3, general harmonic coding is assumed, but particularly in the case of MBE (Multiband Excitation) coding, The modeling is performed on the assumption that there is a voiced sound (Voiced) portion and an unvoiced sound (Unvoiced) portion in each frequency axis region of the same time (in the same block or frame), that is, in each band. In other harmonic coding, an alternative determination is made as to whether voice in one block or frame is voiced or unvoiced. In the following description, the term “V / UV for each frame” means that when all bands are UV when applied to MBE coding, the UV of the frame is used.

The open-loop pitch search section 141 of the sine wave analysis coding section 114 of FIG.
The input voice signal from 01 is again the zero cross counter 1
Signals from the HPF (high-pass filter) 109 are supplied to 42 respectively. The LPC residual or the linear prediction residual from the LPC inverse filter 111 is supplied to the orthogonal transform circuit 145 of the sine wave analysis encoding unit 114. In the open loop pitch search section 141,
An LPC residual of the input signal is used to perform a relatively rough pitch search by an open loop, and the extracted coarse pitch data is sent to a high-precision pitch search 146, and a high-precision closed loop as described later is used. A pitch search (fine search of the pitch) is performed. From the open loop pitch search section 141, a normalized autocorrelation maximum value r (p) obtained by normalizing the maximum value of the autocorrelation of the LPC residual with power together with the coarse pitch data is extracted.
V / UV (voiced sound / unvoiced sound) determination unit 115.

The orthogonal transform circuit 145 performs an orthogonal transform process such as DFT (discrete Fourier transform) to transform the LPC residual on the time axis into spectrum amplitude data on the frequency axis. The output from the orthogonal transform circuit 145 is a high precision pitch search unit 146 and a spectrum evaluation unit 148 for evaluating the spectrum amplitude or envelope.
Sent to

High precision (fine) pitch search unit 146
Is supplied with relatively rough coarse pitch data extracted by the open loop pitch search unit 141 and data on the frequency axis, for example, DFT performed by the orthogonal transform unit 145. The high-precision pitch search unit 146 oscillates ± several samples at intervals of 0.2 to 0.5 around the coarse pitch data value to drive the value of the fine pitch data with a decimal point (floating) to an optimum value. At this time, as a method of fine search, a so-called analysis by synthesis method is used, and the pitch is selected so that the synthesized power spectrum is closest to the power spectrum of the original sound. The pitch data from the high-precision pitch search unit 146 by such a closed loop is output via the switch 118 to the output terminal 10.
4

The spectrum evaluation unit 148 evaluates the size of each harmonics and the spectrum envelope which is a set thereof based on the spectrum amplitude and the pitch as the orthogonal transformation output of the LPC residual, and the high precision pitch search unit 146, V / It is sent to the UV (voiced sound / unvoiced sound) determination unit 115 and the perceptual weighted vector quantizer 116.

V / UV (voiced sound / unvoiced sound) determination section 115
Are the output from the orthogonal transformation circuit 145, the optimum pitch from the high-precision pitch search unit 146, and the spectrum evaluation unit 1
48 and the normalized autocorrelation maximum value r (p) from the open loop pitch search unit 141.
And the zero-cross count value from the zero-cross counter 412, the V / UV determination of the frame is performed. Further, the boundary position of the V / UV determination result for each band in the case of MBE may be used as one condition for the V / UV determination of the frame. The determination output from the V / UV determination unit 115 is taken out via the output terminal 105.

By the way, an output unit of the spectrum evaluation unit 148 or an input unit of the vector quantizer 116 is provided with a data number conversion unit (a kind of sampling rate conversion unit). The number-of-data converters are used to make the amplitude data | A _m | of the envelope a constant number in consideration of the fact that the number of divided bands on the frequency axis varies according to the pitch and the number of data varies. It is. That is, for example, if the effective band is up to 3400 kHz, this effective band is divided into 8 bands to 63 bands according to the pitch, and the amplitude data | A _m | of each of these bands is obtained. The number m _MX +1 also changes from 8 to 63. Therefore, the data number conversion unit 119 converts the variable number m _MX +1 of amplitude data into a fixed number M, for example, 4
It is converted into four data.

The fixed number M (for example, 44) of amplitude data or envelope data from the data number conversion unit provided at the output unit of the spectrum evaluation unit 148 or the input unit of the vector quantizer 116 is a vector quantum. By the digitizer 116, a predetermined number, for example, 44 pieces of data are put together into a vector, and weighted vector quantization is performed. This weight is given by the output from the auditory weighting filter calculation circuit 139. The index of the envelope from the vector quantizer 116 is
It is taken out from the output terminal 103 via the switch 117. Prior to the weighted vector quantization, an inter-frame difference using an appropriate leak coefficient may be calculated for a vector composed of a predetermined number of data.

Next, the second encoder 120 will be described. The second encoding unit 120 has a so-called CELP (Code Excited Linear Prediction) encoding configuration, and is particularly used for encoding an unvoiced sound portion of an input audio signal. In this unvoiced CELP coding configuration,
A noise output corresponding to an LPC residual of unvoiced sound, which is a representative value output from a noise codebook, that is, a so-called stochastic codebook 121, is passed through a gain circuit 126 to a synthesis filter 1 with auditory weights.
22. The weighted synthesis filter 122 performs an LPC synthesis process on the input noise, and sends the obtained weighted unvoiced sound signal to the subtractor 123. A signal obtained by subjecting the audio signal supplied from the input terminal 101 via the HPF (high-pass filter) 109 to auditory weighting by the auditory weighting filter 125 is input to the subtractor 123, and the difference from the signal from the synthesis filter 122 is input to the subtractor 123. Alternatively, the error is extracted. This error is sent to the distance calculation circuit 124 to calculate the distance, and a representative value vector that minimizes the error is searched in the noise codebook 121. Vector quantization of the time axis waveform is performed using the closed loop search using such an analysis by synthesis method.

As the data for the UV (unvoiced sound) portion from the second encoding unit 120 using this CELP encoding structure, the shape index of the codebook from the noise codebook 121 and the code from the gain circuit 126 are used. The gain index and the book are retrieved. Noise codebook 121
Is sent to the output terminal 107s via the switch 127s, and the gain index which is UV data of the gain circuit 126 is sent to the output terminal 107g via the switch 127g.

Here, these switches 127s, 12s
7g and the switches 117 and 118 are connected to the V / UV
On / off control is performed based on the V / UV determination result from the determination unit 115, and the switches 117 and 118 are turned on when the V / UV determination result of the audio signal of the frame to be currently transmitted is voiced (V). 127s, 127g
Is turned on when the audio signal of the frame to be transmitted at present is unvoiced (UV).

Next, FIG. 4 shows a more specific structure of the speech signal decoding apparatus as the embodiment according to the present invention shown in FIG. In FIG. 4, parts corresponding to the respective parts in FIG. 2 are denoted by the same reference numerals.

In FIG. 4, the input terminal 202 has:
L corresponding to the output from the output terminal 102 in FIGS.
An SP vector quantization output, a so-called codebook index, is supplied.

This LSP index is the LSP inverse vector quantizer 231 of the LPC parameter reproducing unit 213.
Is subjected to inverse vector quantization to LSP (line spectrum pair) data, sent to LSP interpolation circuits 232 and 233 and subjected to LSP interpolation processing, and then subjected to LPC (linear) by LSP → α conversion circuits 234 and 235. The α parameter is transmitted to the LPC synthesis filter 214. Here, the LSP interpolation circuit 232 and the LSP → α conversion circuit 234 are for voiced sound (V).
The SP interpolation circuit 233 and the LSP → α conversion circuit 235 are for unvoiced sound (UV). Also, the LPC synthesis filter 214
Separates the LPC synthesis filter 236 for the voiced portion and the LPC synthesis filter 237 for the unvoiced portion. That is, LPC coefficient interpolation is performed independently for voiced and unvoiced parts, and LSPs having completely different properties are interpolated between the transition from voiced to unvoiced and the transition from unvoiced to voiced. To prevent the adverse effects of doing so.

The input terminal 203 of FIG. 4 is supplied with the coded index data of the spectral envelope (Am) weighted vector quantized corresponding to the output from the encoder side terminal 103 of FIGS. The input terminal 204 is supplied with the pitch data from the terminal 104 in FIGS. 1 and 3, and the input terminal 205 is supplied with the V / UV determination data from the terminal 105 in FIGS. ing.

The vector-quantized index data of the spectral envelope Am from the input terminal 203 is
The data is sent to the inverse vector quantizer 212, subjected to inverse vector quantization, subjected to an inverse transform corresponding to the above-described data number conversion, becomes spectral envelope data, and becomes a sine wave synthesizing circuit of the voiced sound synthesizer 211. 215.

If the interframe difference is taken before the vector quantization of the spectrum at the time of encoding, the number of data is converted after decoding the interframe difference after the inverse vector quantization here, and the spectrum envelope is converted. Get the data of.

The sine wave synthesis circuit 215 has an input terminal 2
04 and the V / U from the input terminal 205
V determination data is supplied. Sine wave synthesis circuit 21
5, the LPC inverse filter 11 shown in FIGS.
LPC residual data corresponding to the output from 1 is extracted and sent to the adder 218.

The envelope data from the inverse vector quantizer 212, the pitch from the input terminals 204 and 205, and the V / UV determination data are used as a noise synthesis circuit for adding noise in the voiced sound (V) portion. Have been sent to 216. The output from the noise synthesis circuit 216 is sent to an adder 218 via a weighted superposition addition circuit 217. This is an excitation (Excitati) which is input to the LPC synthesis filter of voiced sound by sine wave synthesis.
When on (excitation, excitation) is made, there is a case where there is a feeling of nasal congestion with a low pitch sound such as a male voice, and the sound quality changes suddenly between V (voiced sound) and UV (unvoiced sound) and feels unnatural. Considering a certain point, the LPC synthesis filter input of the voiced sound portion, that is, the excitation, was considered in consideration of parameters based on the speech coded data, for example, pitch, spectrum envelope amplitude, maximum amplitude in a frame, residual signal level, and the like. Noise is added to the voiced portion of the LPC residual signal.

The addition output from the adder 218 is sent to the voiced sound synthesis filter 236 of the LPC synthesis filter 214 and subjected to LPC synthesis processing to become time waveform data, and further to the voiced sound post filter 238v.
, And sent to the adder 239.

Next, the input terminals 207s and 207 of FIG.
The shape index and the gain index as UV data from the output terminals 107 s and 107 g in FIG. 3 are supplied to g, and are sent to the unvoiced sound synthesis unit 220. The shape index from the terminal 207s is stored in the noise codebook 221 of the unvoiced sound synthesizer 220 in the terminal 2
The gain index from 07g is sent to the gain circuit 222, respectively. The representative value output read from the noise codebook 221 is a noise signal component corresponding to the LPC residual of the unvoiced sound. The noise signal component has an amplitude of a predetermined gain in the gain circuit 222 and is sent to the windowing circuit 223. A windowing process is performed to smooth the connection with the voiced sound portion.

The output from the windowing circuit 223 is the output from the unvoiced sound synthesizing section 220, which is the LPC synthesis filter 21.
4 is sent to the synthesis filter 237 for UV (unvoiced sound). The synthesis filter 237 performs LPC synthesis processing to obtain unvoiced sound time waveform data. The unvoiced sound time waveform data is filtered by the unvoiced sound post filter 238u, and then sent to the adder 239.

In the adder 239, the time waveform signal of the voiced sound portion from the voiced sound post filter 238v and the time waveform data of the unvoiced sound portion from the unvoiced sound post filter 238u are added together and taken out from the output terminal 201.

In the above audio signal encoding device, the bit rate of the output data is changed and output. In particular,
The bit rate of output data can be switched between a low bit rate and a high bit rate. For example, when the low bit rate is 2 kbps and the high bit rate is 6 kbps, the data of each bit rate shown in Table 1 below is output.

[0050]

[Table 1]

The pitch data from the output terminal 104 is always output at 8 bits / 20 msec during voiced sound, and the V / UV determination output from the output terminal 105 is always 1 bit / 20 msec. The LSP quantization index output from the output terminal 102 is 32 bits / 4
Switching is performed between 0 msec and 48 bits / 40 msec. The index of the voiced sound (V) output from the output terminal 103 is 15 bits / 20 msec and 87 bits.
Switching between s / 20 msec and output terminal 10
The index at the time of unvoiced sound (UV) output from 7s and 107g is switched between 11 bits / 10 msec and 23 bits / 5 msec. Thus, the output data at the time of voiced sound (V) is 40 bits / 20 ms at 2 kbps.
ec, which is 120 bits / 20 msec at 6 kbps.
The output data at the time of unvoiced sound (UV) is 39 bits / 20 msec at 2 kbps, and 117 bits / 20 msec at 6 kbps.
20 msec.

The LSP quantization index, the voiced sound (V) index, and the unvoiced sound (UV) index will be described together with the configuration of each unit described later.

Next, the matrix quantization and the vector quantization in the LSP quantizer 134 will be described in detail with reference to FIGS. 5 and 6.

As described above, the α parameter from the LPC analysis circuit 132 is sent to the α → LSP conversion circuit 133 and converted into the LSP parameter. For example, when the P order LPC analysis is performed by the LPC analysis circuit 132, P α parameters are calculated. The P α parameters are converted into LSP parameters and stored in the buffer 610.

This buffer 610 outputs LSP parameters for two frames. LS for 2 frames
The P parameter is subjected to matrix quantization by the matrix quantization unit 620. Matrix quantizer 620 consists of a first matrix quantizer 620 ₁ and a second matrix quantizer 620 _2. The LSP parameters for two frames are subjected to matrix quantization in the first matrix quantization section 620 ₁ , and the resulting quantization error is further subjected to matrix quantization in the second matrix quantization section 620 ₂ . By these matrix quantization, the correlation in the time axis direction is removed.

The quantization error for two frames from the matrix quantizer 620 ₂ is input to the vector quantizer 640. The vector quantization unit 640 includes a first vector quantization unit 640 ₁ and a second vector quantization unit 640 ₂ . Further, the first vector quantizer 640 ₁
The second vector quantizer 640 ₂ comprises two vector quantizers 650 and 660, and the second vector quantizer 640 ₂ comprises two vector quantizers 670 and 680. First vector quantizer 64
The vector quantization units 650 and 660 of 0 ₁ vector-quantize the quantization error from the matrix quantization unit 620 for each frame. The quantization error vector thus obtained is further vector-quantized by the vector quantizers 670 and 680 of the _second vector quantizer 640 ₂ . By these vector quantizations, the correlation in the frequency axis direction is processed.

As described above, the matrix quantizing unit 620 for performing the matrix quantizing process has the first matrix quantizing unit 620 for performing the first matrix quantizing process.
₁ and at least a _second matrix quantizer 620 ₂ for performing a second matrix quantization step of matrix quantizing a quantization error due to the first matrix quantization, and performing the above vector quantization. The vector quantizer 640 for performing the first vector quantization step 640 ₁ for performing the first vector quantization step, and the second vector quantizer 640 for vector quantizing the quantization error vector at the time of the first vector quantization.
Vector quantization unit 6 that performs the vector quantization process of
Having at least a 40 _2.

Next, the matrix quantization and the vector quantization will be specifically described.

The LSP parameters for two frames, that is, the 10 × 2 matrix held in the buffer 610 are sent to the matrix quantizer 620 ₁ . In the first matrix quantizer 610 ₁ , the LSP parameters for two frames are sent to the weighted distance calculator 6 via the adder 621.
23, and a minimum weighted distance is calculated.

The distortion measure d _MQ1 at the time of codebook search by the first matrix quantizer 620 ₁ is expressed by the equation (1) using the LSP parameter X ₁ and the quantized value X ₁ .

[0061]

(Equation 1)

Here, t is a frame number and i is a P-dimensional number.

Further, at this time, the weight W when the limitation of the weight is not taken into consideration in the frequency axis direction and the time axis direction is (2)
It is shown by the formula.

[0064]

(Equation 2)

The weight W in the equation (2) is also used in matrix quantization and vector quantization in the latter stage.

The calculated weighted distance is sent to the matrix quantizer (MQ ₁ ) 622 and matrix quantization is performed. The 8-bit index output by the matrix quantization is sent to the signal switch 690. The quantization value obtained by the matrix quantization is added to an adder 621.
Then, it is subtracted from the LSP parameters for the next two frames from the buffer 610. The weighted distance calculator 623 uses the output from the adder 621 to calculate the minimum weighted distance. In this way, every two frames
The weighted distance calculator 623 sequentially calculates the weighted distance, and the matrix quantizer 622 performs matrix quantization. The output from the adder 621 is the second
Is sent to the adder 631 of the matrix quantizer 620 ₂ .

The second matrix quantizer 620 ₂ also performs matrix quantization in the same manner as the first matrix quantizer 620 ₁ . The output from the adder 621 is sent to the weighted distance calculator 633 via the adder 631, and the minimum weighted distance is calculated.

The distortion measure d _MQ2 at the time of codebook search by the second matrix quantizer 620 _{2 is calculated} by the quantization error X ₂ and the quantization value X ₂ from the _first matrix quantizer 620 _1. It is shown by the equation 3).

[0069]

(Equation 3)

This weighted distance is sent to the matrix quantizer (MQ ₂ ) 632 for matrix quantization. The 8-bit index output by the matrix quantization is sent to the signal switch 690. Further, the quantized value obtained by the matrix quantization is subtracted from the quantization error for the next two frames by the adder 631. The weighted distance calculator 633 sequentially calculates the minimum weighted distance using the output from the adder 631. Also,
The output from the adder 631 is the first vector quantization unit 6
One frame is sent to the adders 651 and 661 of 40 ₁ .

The first vector quantizer 640 ₁ performs vector quantization for each frame. The output from the adder 631 is added to the adder 651,
The weighted distance is sent to weighted distance calculators 653 and 663 via 661 to calculate the minimum weighted distance.

The difference between the quantization error X ₂ and the quantization value X ₂ is
It is a 10 × 2 matrix and is represented by the vector quantizers 652 and 662 of the _first vector quantizer 640 ₁ when X ₂ −X ₂ ′ = [ X _3-1 , X _3-2 ]. The distortion measures d _VQ1 and d _VQ2 at the time of codebook search are shown by equations (4) and (5).

[0073]

(Equation 4)

This weighted distance is calculated by the vector quantizer (V
Q ₁ ) 652 and vector quantizer (VQ ₂ ) 662, respectively, and vector quantization is performed. Each 8-bit index output by this vector quantization is sent to a signal switch 690. Further, the quantized value obtained by the vector quantization is subtracted by the adders 651 and 661 from the quantized error vector of the next two frames.
In the weighted distance calculators 653 and 663, the adder 65
Using the outputs from 1 and 661, the minimum weighted distance is sequentially calculated. The outputs from the adders 651 and 661 are the outputs of the adder 6 of the _second vector quantizer 640 _2.
71 and 681 respectively.

Here, this second vector quantizer 640 when X _4-1 = X _3-1 −X ′ _3-1 X _4-2 = X _3-2 −X ′ _3-2 is expressed. the distortion measure d _VQ3, d _VQ4 during codebook search by the _second vector quantizer 672, 682, (6), shown in equation (7).

[0076]

(Equation 5)

This weighted distance is calculated by the vector quantizer (V
Q ₃ ) 672 and vector quantizer (VQ ₄ ) 682, respectively, for vector quantization. Each 8-bit index output by this vector quantization is sent to a signal switch 690. Further, the quantized value obtained by the vector quantization is subtracted from the quantized error vectors for the next two frames to be inputted by the adders 671 and 681.
In the weighted distance calculators 673 and 683, the adder 67
Using the outputs from 1 and 681, the minimum weighted distance is sequentially calculated.

Further, at the time of learning the codebook, the generalized Lloyd algorithm (GL
Learning is performed according to A).

The distortion scales at the time of codebook search and at the time of learning may have different values.

The matrix quantizers 622, 632,
The 8-bit indexes from the vector quantizers 652, 662, 672, and 682 are switched by the signal switch 690 and output from the output terminal 691.

Specifically, when the bit rate is low, the first matrix quantizing section 620 ₁ for performing the first matrix quantizing step and the second matrix quantizing section for performing the second matrix quantizing step are described. 620 ₂ , and a first vector quantization unit 64 that performs the first vector quantization step.
0 output at ₁ taken out, at the time of high bit-rate, taken together outputs of the second vector quantizer 640 ₂ carrying out the second vector quantization process on the output when the low bit rate.

As a result, at 2 kbps, 32 bits /
An index of 40 msec is output and 4 at 6 kbps.
The 8bits / 40msec index is output.

Further, in the matrix quantizing unit 620 and the vector quantizing unit 640, in the frequency axis direction or the time axis direction, or in the frequency axis and time axis direction, which is in accordance with the characteristic of the parameter expressing the LPC coefficient. Weight with restrictions.

First, description will be given of weighting having a limit in the frequency axis direction, which is matched with the characteristic of the LSP parameter. For example, when the degree P = 10, the LSP parameter X (i) is set to three regions of low band, middle band, and high band: L ₁ = {X (i) | 1 ≦ i ≦ 2} L ₂ = {X (i) | 3 ≦ i ≦ 6} L ₃ = {X (i) | 7 ≦ i ≦ 10}. Then, each group L ₁ , L ₂ , L ₃
Is 1/4, 1/2, and 1/4, the weights of each group L ₁ , L ₂ , and L ₃ having restrictions only in the frequency axis direction are (8), (9), and (10). It becomes an expression.

[0085]

(Equation 6)

As a result, the weighting of each LSP parameter is performed only within each group, and the weighting is limited by the weighting for each group.

Here, when viewed from the time axis direction, the sum of weighting of each frame is always 1, so that the limitation in the time axis direction is one frame unit. The weight having a restriction only in the time axis direction is given by equation (11).

[0088]

(Equation 7)

According to the equation (11), weighting is performed between two frames having frame numbers t = 0 and 1 which are not limited in the frequency axis direction. The weighting having a limitation only in the time axis direction is performed between two frames on which matrix quantization is performed.

At the time of learning, all voice frames used as learning data, that is, the number T of frames of all data, is weighted by the equation (12).

[0091]

(Equation 8)

Weighting with restrictions in the frequency axis direction and the time axis direction will be described. For example, the order P = 1
When L is 0, the LSP parameter X (i, t) is set to three regions of low band, middle band, and high band. L ₁ = {X (i, t) | 1 ≦ i ≦ 2,0 ≦ t ≦ 1} L ₂ = {X (i, t) | 3 ≦ i ≦ 6, 0 ≦ t ≦ 1} L ₃ = {X (i, t) | 7 ≦ i ≦ 10, 0 ≦ t ≦ 1} and a group Turn into. Assuming that the weights of the groups L ₁ , L ₂ , L ₃ are １ ／, 、, １ ／, the weights of the groups L ₁ , L ₂ , L ₃ are limited in the frequency axis direction and the time axis direction. Becomes the expressions (13), (14), and (15).

[0093]

(Equation 9)

According to the equations (13), (14) and (15), weighting is performed for every three bands in the frequency axis direction and in the time axis direction by weighting limitation between two frames for matrix quantization. . This is effective for both codebook search and learning.

At the time of learning, the number of frames of all data is weighted. LSP parameter X
Let (i, t) be the three regions of low band, middle band, and high band. L ₁ = {X (i, t) | 1 ≦ i ≦ 2, 0 ≦ t ≦ T} L ₂ = {X ( i, t) | 3 ≦ i ≦ 6, 0 ≦ t ≦ T} L ₃ = {X (i, t) | 7 ≦ i ≦ 10, 0 ≦ t ≦ T}, and each group L ₁ , L _If the weighting of ₂ and L ₃ is 1/4, 1/2, and 1/4, each group L ₁ ,
The weights L ₂ and L ₃ having restrictions in the frequency axis direction and the time axis direction are given by equations (16), (17) and (18).

[0096]

(Equation 10)

By the equations (16), (17) and (18), weighting is performed for each of the three bands in the frequency axis direction,
In the time axis direction, weighting can be performed between all frames.

Further, the matrix quantizer 620 and the vector quantizer 640 perform weighting according to the magnitude of change in the LSP parameter. V → UV, U, which is a small number of frames in the entire audio frame
In the transition (transient) portion of V → V, the LSP parameter greatly changes due to a difference in frequency characteristics between the consonant and the vowel. Therefore, by multiplying the weight W '(i, t) described above by the weight shown in the equation (19), weighting with emphasis on the transition part can be performed.

[0099]

[Equation 11]

It is also possible to use equation (20) instead of equation (19).

[0101]

(Equation 12)

Thus, in the LSP quantizer 134,
By performing two-stage matrix quantization and two-stage vector quantization, the number of bits of an output index can be made variable.

Next, the basic configuration of the vector quantizer 116 is shown in FIG. 7, and the more specific configuration of the vector quantizer 116 of FIG. 7 is shown in FIG. 8, and the spectrum envelope (Am) of the vector quantizer 116 is shown. A specific example of weighted vector quantization will be described.

First, in the speech signal coding apparatus of FIG. 3, data number conversion to make the number of amplitude envelope data provided on the output side of the spectrum evaluation section 148 or the input side of the vector quantizer 116 a fixed number. A specific example of will be described.

Various methods are conceivable for this data number conversion, but in the present embodiment, for example, for the amplitude data for one block of the effective band on the frequency axis, from the last data in the block to the block. After adding dummy data for interpolating the values up to the first data in the table to expand the number of data to N _F , the band-limited O _S times (for example, 8 times) oversampling is performed to obtain O. obtain an amplitude data of _S times the number, the O _S times the number ((m _MX
(+1) × O _S pieces of amplitude data are linearly interpolated to be expanded to a larger number of N _M pieces (for example, 2048 pieces), and the N _M pieces of data are thinned out to obtain the fixed number of M pieces (for example, 44 pieces) of data. Has been converted to.

The vector quantizer 116 for performing the weighted vector quantization shown in FIG. 7 includes a first vector quantizer 500 for performing the first vector quantizing process and a first vector quantizer 500 in the first vector quantizer 500. And a second vector quantizer 510 that performs a second vector quantization step of quantizing a quantization error vector at the time of vector quantization of 1. The first vector quantization unit 500 is a so-called first-stage vector quantization unit, and the second vector quantization unit 510 is a so-called second-stage vector quantization unit.

The output vector of the spectrum evaluation unit 148 is connected to the input terminal 501 of the first vector quantization unit 500.
X , that is, a fixed number M of envelope data is input. This output vector X is weighted vector quantized by the vector quantizer 502. As a result, the shape index output from the vector quantizer 502 is output from the output terminal 503, the quantized value X ₀ 'is output from the output terminal 504, and the adders 505, 5
Sent to 13. In the adder 505, the quantized value X ₀ 'is subtracted from the output vector X to obtain a multidimensional quantization error vector Y.

This quantization error vector Y is sent to the vector quantization unit 511 in the second vector quantization unit 510. This vector quantization section 511 is composed of a plurality of vector quantizers, and in FIG. 7, is composed of two vector quantizers 511 ₁ and 511 ₂ . Quantization error vector Y
Is dimensionally divided into two vector quantizers 511 ₁ ,
At 511 ₂ , each is weighted vector quantized.
The shape indexes output from these vector quantizers 511 ₁ and 511 _{2 are} output to output terminals 512 ₁ and 512 _2.
Respectively, and the quantized values Y ₁ 'and Y ₂ ' are connected in the dimension direction and are sent to the adder 513. In the adder 513, the quantized values Y ₁ 'and Y ₂ ' and the quantized value X
_0'and are added to generate a quantized value X ₁ '. This quantized value X ₁ 'is output from the output terminal 514.

As a result, when the bit rate is low, the output of the first vector quantization step by the first vector quantizer 500 is taken out, and when the bit rate is high, the output is obtained.
The output in the first vector quantization step and the output in the second vector quantization step by the second quantization unit 510 are extracted.

Specifically, as shown in FIG. 8, the vector quantizer 502 of the first vector quantizer 500 in the vector quantizer 116 has an L-dimensional, for example 44-dimensional, two-dimensional vector quantizer 502.
It has a stage configuration.

That is, the sum of output vectors from a vector quantization codebook having a 44-dimensional codebook size of 32 and a gain g _i is multiplied to obtain a quantized value X ₀ 'of a 44-dimensional spectral envelope vector X. To use as. As shown in FIG. 8, the two shape codebooks are CB0 and CB1, and the output vectors are s _0i and s _1j , where 0 ≦ i, j ≦ 31. The output of the gain codebook CBg is represented by _gl , where 0
≦ l ≦ 31. _gl is a scalar value. This final output X ₀ 'is g _i ( s _0i + s _1j ).

Let X be a constant dimension of the spectral envelope Am obtained by the above MBE analysis of the LPC residual. At this time, how to efficiently quantize X is important.

[0113] Here, the quantization error energy E, E = ‖W {H X -Hg l (s 0i + s 1j)} ‖ ^{2 ··· (21) = ‖WH {} X -g l (s 0i + S _1j )} ∥ ² . In the equation (21), H is a characteristic of the LPC synthesis filter on the frequency axis, and W is a matrix for weighting that represents the characteristic of the auditory weighting on the frequency axis.

The α parameter according to the LPC analysis result of the current frame is α _i (1 ≦ i ≦ P),

[0115]

(Equation 13)

The value of each corresponding point in the L dimension, for example, 44 dimensions, is sampled from the frequency characteristic of.

The calculation procedure is, for example, 1,
α ₁ , α ₂ ,..., α _p are padded with 0, that is, 1,
α ₁ , α ₂ ,..., α _p , 0, 0,.
For example, data of 256 points is used. After that, 256 points FF
Perform T, it is calculated for points corresponding to 0~π the ^{_{(r e 2 + I m 2}} ) 1/2, taking its reciprocal. A matrix having diagonal elements obtained by thinning it out to L points, for example, 44 points,

[0118]

[Equation 14]

It is assumed that

The perceptual weighting matrix W is

[0121]

(Equation 15)

It is assumed that In this equation (23), α _i is the LPC analysis result of the input. Further, λa and λb are constants,
As an example, λa = 0.4 and λb = 0.9 can be mentioned.

The matrix or the matrix W is (2)
It can be calculated from the frequency characteristic of equation 3). As an example, 1,
α ₁ λb, α ₂ λb ² ,..., α _p λb ^p , 0, 0,.
0 As performs an FFT as 256 points data for 0 or π following section ^{_{(r e 2 [i] +}} I m 2 [i]) 1/2, 0
≦ i ≦ 128 is calculated. Next, 1, α ₁ λa, α ₂ λ
a ² , ..., α _p λ a ^p , 0, 0, ..., 0, the frequency characteristic of the denominator is 256 points FFT, and the interval of 0 to π is 128.
Calculate in points. This is (r _e ' ² [i] + I _m ' ² [i]) ^1/2 ,
0 ≦ i ≦ 128.

[0124]

(Equation 16)

As the above, the frequency characteristic of the equation (23) is obtained.

This is obtained by the following method for the corresponding points of the L-dimensional, for example, 44-dimensional vector. More precisely, linear interpolation should be used, but the following example substitutes the value of the closest point.

That is, ω [i] = ω ₀ [nint (128i / L)] 1 ≦ i ≦ L where nint (X) is a function that returns an integer closest to X.

Further, with respect to the above H, in the same manner, h
(1), h (2), ..., h (L) are calculated. That is,

[0129]

[Equation 17]

Is obtained.

Here, as another example, in order to reduce the number of FFTs, H (z) W (z) may be obtained first, and then the frequency characteristic may be obtained. That is,

[0132]

(Equation 18)

The result of expanding the denominator of the equation (25) is

[0134]

[Equation 19]

It is assumed that Here, 1, β ₁ , β ₂ , ...,
As β _2p , 0, 0,... After that, a 256-point FFT is performed, and the frequency characteristic of the amplitude is

[0136]

(Equation 20)

It is assumed that Than this,

[0138]

(Equation 21)

This is calculated for the corresponding points of the L-dimensional vector. When the number of FFT points is small, it should be obtained by linear interpolation, but the nearest value is used here. That is,

[0140]

(Equation 22)

It is The matrix with this as diagonal elements is W '
Then

[0142]

(Equation 23)

It becomes: Equation (26) is the same matrix as equation (24).

Using the matrix, that is, the frequency characteristics of the weighted synthesis filter, the above equation (21) can be rewritten.

[0145]

(Equation 24)

[0146]

Here, a method of learning the shape codebook and the gain codebook will be described.

First, the expected value of distortion is minimized for all frames k that select the code vector s _0c for CB0. Assuming there are M such frames,

[0149]

(Equation 25)

It is only necessary to minimize In the equation (28), W ′ _k is a weight for the _kth frame, X _k is an input of the _kth frame, g _k is a gain of the kth frame, and s _1k is a codebook for the kth frame. C
Output from B1.

To minimize the equation (28),

[0152]

(Equation 26)

[0153]

[Equation 27]

Next, optimization regarding gain will be considered.

The expected distortion value J _{g for} the kth frame selecting the gain codeword g _c is

[0156]

[Equation 28]

The above equations (31) and (32) are the optimum centroid condition (Centroid Condition) of the shapes s _0i , s _1i and the gain g _i , 0 ≦ i ≦ 31, that is, the optimum conditions. It provides a decoder output. In addition, s
_{The value of 1i} can be _{calculated in the} same manner as s _0i .

Next, the optimum encoding condition (Nearest Neig
hbour Condition).

[0159] The obtaining the distortion measure (27), i.e., E = ‖W '(X -g l (s 0i + s 1j)) ‖ ² minimizes s _0i, the s _1j, the input X, the weight It is determined every time the matrix W ′ is given, that is, every frame.

Originally, all g _l (0 ≦ l ≦
31), s _0i (0 ≦ i ≦ 31), s _{1 j} (0 ≦ j ≦ 31)
It is necessary to find E for 32 × 32 × 32 = 32768 combinations of the above, and find the set of g _l , s _0i , and s _1j that gives the minimum E, but this is an enormous amount of calculation. In this form, a sequential search of shape and gain is performed. Note that a brute force search is performed for the combination of s _0i and s _1j . This is 32
× 32 = 1024 patterns. In the following description, for simplicity, the s _0i + s _1j referred to as s _m.

[0161] The above equation (27) is, E = ‖W '(X -g l s
_m) ‖ ² to become. For further _{simplicity, X w = W 'X,} s w
= When W 's _m,

[0162]

(Equation 29)

It becomes Therefore, assuming that the accuracy of _gl is sufficiently high,

[0164]

[Equation 30]

The search can be performed by dividing it into two steps. Rewriting using the original notation,

[0166]

(Equation 31)

It becomes: This equation (35) is the optimum encoding condition (Nearest Neighbor Condition).

Here, using the conditions (Centroid Condition) of the above equations (31) and (32) and the condition of the equation (35), an LBG (Linde-Buzo-Gray) algorithm, a so-called generalized Lloyd algorithm is used. Algori
thm: GLA) and the codebook (CB0, CB1, C
Bg) can be trained simultaneously.

By the way, the weight W'used for perceptual weighting at the time of vector quantization in the vector quantizer 116.
Is defined by the above equation (26), and W'in consideration of temporal masking is also obtained by obtaining the current W'in consideration of the past W '.

Wh (1), wh (2), ..., w in the above equation (26)
respect h (L), the time n, that each wh _n (1) those which are calculated in the n-th _{frame, wh n (2), ···} , and wh _n (L).

At time n, the weight considering the past value is A
_{If n} (i) and 1 ≦ i ≦ L are defined, A _n (i) = λA _n-1 (i) + (1-λ) wh _n (i) (wh _n (i) ≦ A _n-1 ( i)) = wh _n (i) (wh _n (i)> A _n-1 (i)). Here, λ may be, for example, λ = 0.2.
For A _n (i), 1 ≦ i ≦ L obtained in this way, a matrix having diagonal elements may be used as the weight.

Shape indexes s _0i and s _1j thus obtained by weighted vector quantization are output terminal 5
20, 522, and the gain index _gl is output from the output terminal 521. The quantized value X ₀ ′ is output from the output terminal 504 and is also sent to the adder 505.

In this adder 505, the quantized value X ₀ 'is subtracted from the output vector X , and the quantized error vector Y
Is generated. Specifically, this quantization error vector Y is sent to a vector quantizer 511 composed of _eight vector quantizers 511 _{1 to} 511 ₈ , and is dimensionally divided, and each vector quantizer 511 _{1 to} 511 is quantized. Weighted vector quantization is applied at ₈ .

In the second vector quantizer 510, the first vector quantizer 510
Since a considerably large number of bits are used as compared with the vector quantization unit 500, the memory capacity of the codebook and the amount of calculation (Complexity) for the codebook search become very large, and the first vector quantization unit 500 Same as 500 4
It is impossible to perform vector quantization with four dimensions. Therefore, the vector quantizer 511 in the second vector quantizer 510 is composed of a plurality of vector quantizers, and the input quantized value is dimensionally divided into a plurality of low-dimensional vectors and weighted. Perform vector quantization with.

Table 2 shows the relationship between the quantized values Y ₀ to Y ₇ used in the vector quantizers 511 _{1 to} 511 ₈ , the number of dimensions, and the number of bits.

[0176]

[Table 2]

[0177] index Idvq ₀ ~Idvq ₇ output from the vector quantizer 511 _{1 to 511 8} are outputted from the output terminals 523 _{1 to 523 8.} The sum of these indexes is 72 bits.

Further, the vector quantizers 511 _{1 to} 511 ₈
If a value obtained by connecting the quantized values Y ₀ 'to Y ₇ ' in the dimension direction is Y ', the quantized value
Y ′ and the quantized value X ₀ ′ are added to each other to obtain the quantized value X ₁ ′.
Is obtained. Therefore, the quantized value X ₁ ′ is represented by X ₁ ′ = X ₀ ′ + Y ′ = X − Y + Y ′. That is, the final quantization error vector is Y'- Y .

On the audio signal decoding device side, this second
When decoding the quantized value X ₁ 'from the vector quantizer 510, the quantized value X ₀ ' from the first vector quantizer 500 is unnecessary, but the first vector quantizer 500 And the index from the second vector quantizer 510 is required.

Next, the learning method and codebook search in the vector quantizer 511 will be described.

First, in the learning method, the quantization error vector Y and the weight W'are used to divide into eight low-dimensional vectors Y ₀ to Y ₇ and a matrix as shown in Table 2. At this time, the weight W ′ is, for example, a matrix having diagonal elements obtained by thinning out 44 points,

[0182]

(Equation 32)

Then, it is divided into the following eight matrices.

[0184]

[Equation 33]

The low-dimensional divisions of Y and W'in this way are defined as Y _i and W _i '(1≤i≤8), respectively.

Here, the distortion measure E is defined as E = ‖W _i '( Y _i − s ) ‖ ² ... (37). This code vector s is the quantization result of Y _i , and the code vector s of the codebook that minimizes the distortion measure E is searched.

Note that W _i ′ may be weighted at the time of learning and not weighted at the time of searching, that is, may be a unit matrix, and different values may be used at the time of learning and at the time of codebook search.

In the codebook learning, a generalized Lloyd algorithm (GLA) is used and further weighting is performed. First, an optimal centroid condition for learning will be described. When there are M input vectors Y in which the code vector s is selected as the optimum quantization result, and the training data is Y _k , the expected value J of the distortion is that the center of the distortion at the time of weighting is the minimum for all frames k. The equation (38) becomes

[0189]

(Equation 34)

S shown in the above equation (39) is an optimum representative vector, which is an optimum centroid condition.

The optimum encoding condition is ‖W _i '( Y
_It is sufficient to search for s that minimizes the value of _i − s ) ‖ ² .
Here, W _i 'at the time of searching is not always the same as W _i ' at the time of learning.
Need not be, with no weight

[0192]

(Equation 35)

The matrix of may be used.

As described above, by configuring the vector quantizer 116 in the speech signal coding apparatus with the two-stage vector quantizer, the number of bits of the output index can be made variable.

Next, the second coding unit 120 using the CELP coding structure of the present invention is more specifically a multi-stage vector quantization processing unit (example of FIG. 9) as shown in FIG. In the above, the two-stage encoding units 120 ₁ and 120 ₂ ) are included. The configuration of FIG. 9 shows a configuration corresponding to a transmission bit rate of 6 kbps when the transmission bit rate can be switched between 2 kbps and 6 kbps, for example, and the shape and gain index output are 23 bits / The switching is made between 5 msec and 15 bits / 5 msec. The flow of processing in the configuration of FIG. 9 is as shown in FIG.

In FIG. 9, for example, the first encoding unit 200 of FIG. 9 substantially corresponds to the first encoding unit 113 of FIG. 3, and the LPC analysis circuit 302 of FIG. Corresponding to the LPC analysis circuit 132 shown, the LSP parameter quantization circuit 303 of FIG. 9 corresponds to the configuration from the α → LSP conversion circuit 133 to the LSP → α conversion circuit 137 of FIG. Reference numeral 304 corresponds to the auditory weighting filter calculation circuit 139 and the auditory weighting filter 125 of FIG. Therefore, this FIG.
At the terminal 305, the first encoding unit 1 of FIG.
The same output from the LSP → α conversion circuit 137 of FIG. 13 is supplied, and the same output from the auditory weighting filter calculation circuit 139 of FIG. 3 is supplied to the terminal 307.
At the terminal 306, the perceptual weighting filter 125 of FIG.
The same output from is supplied. However, FIG.
Unlike the perceptual weighting filter 125 of FIG. 3, the perceptual weighting filter 304 of FIG. 3 performs the perceptual weighting from the input voice data and the α parameter before quantization without using the output of the LSP → α conversion circuit 137. Signal (ie, the same signal as the output from the perceptual weighting filter 125 of FIG. 3 above).

In addition, in the two-stage second coding units 120 ₁ and 120 ₂ shown in FIG. 9, the subtractors 313 and 323 correspond to the subtractor 123 of FIG. 3, and the distance calculation circuit 3
Reference numerals 14 and 324 denote the distance calculation circuit 124 of FIG. 3, gain circuits 311 and 321 denote the gain circuit 126 of FIG. 3, and stochastic codebooks 310 and 320 and gain codebooks 315 and 325 denote the noise codebook 12 of FIG.
It corresponds to 1 respectively.

In the configuration of FIG. 9 as described above, first, as shown in step S1 of FIG.
In 2, the input voice data x supplied from the terminal 301 is divided into appropriate frames as described above and LPC analysis is performed to obtain the α parameter. The LSP parameter quantization circuit 303 converts the α parameter from the LPC analysis circuit 302 into an LSP parameter, quantizes the LSP parameter, interpolates the quantized LSP parameter, and converts it into an α parameter. Next, the LSP parameter quantizing circuit 303 generates an LPC synthesis filter function 1 / H (z) from the α parameter obtained by converting the quantized LSP parameter, that is, the quantized α parameter. 305 via a letter to the second encoding unit 120 ₁ of the perceptually weighted synthesis filter 312 of the first stage.

On the other hand, in the perceptual weighting filter 304,
From the α parameter from the LPC analysis circuit 302 (that is, the α parameter before quantization), the same data for perceptual weighting as obtained by the perceptual weighting filter calculation circuit 139 in FIG. 3 is obtained. 307, the second encoding unit 120 _{1 in the} first stage
To the auditory weighted synthesis filter 312. Also,
In the perceptual weighting filter 304, as shown in step S2 in FIG. 10, the perceptually weighted signal (see FIG. 3 above) is calculated from the input voice data and the α parameter before quantization.
The same signal as the output from the auditory weighting filter 125)
Generate That is, first, the perceptual weighting filter function W (z) is generated from the α parameter before quantization, the input audio data x is further multiplied by the filter function W (z) to generate x _W, and the perceptual weighting is performed. As a generated signal via the terminal 306 to the second encoding unit 120 of the first stage.
_{1 to} the subtractor 313.

In the second encoding section 120 ₁ of the first stage, 9
A representative value output (a noise output corresponding to the LPC residual of unvoiced sound) from the stochastic code book 310 of the bit shape index output is sent to the gain circuit 311, and the stochastic code is output from the gain circuit 311. The representative value output from the book 310 is multiplied by the gain (scalar value) from the gain codebook 315 of the 6-bit gain index output, and the representative value output multiplied by the gain in the gain circuit 311 is:
1 / A (z) = (1 / H (z)) · W (z) is sent to the synthesis filter 312 with the auditory weight. From the weighted synthesis filter 312, the zero input response output of 1 / A (z) is subtracted by the subtractor 313 as in step S3 of FIG.
Sent to In the subtractor 313, the zero-input response output from the auditory weighting synthesis filter 312, is subtracted with the above perceptually weighted signal x _W from the perceptually weighted filter 304 is performed, the difference or error reference Extracted as a vector r . As shown in step S4 of FIG. 10, the second encoding unit 120 ₁ in the first stage
At the time of the search, the reference vector r is sent to the distance calculation circuit 314, where the distance calculation is performed, and the shape vector s and the gain g that minimize the quantization error energy E are searched. Here, 1 / A (z) is in a zero state. That is, the shape vector s in the codebook is synthesized by 1 / A (z) of the zero state to s
_{When syn} is set, a shape vector s and a gain g that minimize Expression (40) are searched.

[0201]

[Equation 36]

Here, s and g which minimize the quantization error energy E may be fully searched, but the following method can be adopted in order to reduce the amount of calculation.

As a first method, a shape vector s that minimizes E _s defined in the following equation (41) is searched for.

[0204]

(37)

As the second method, since the ideal gain is as shown in equation (42) from s obtained by the first method, g which minimizes equation (43) is searched for.

[0206]

(38)

E _g = (g _ref −g) ² (43) Here, since E is a quadratic function of _g , _g that minimizes E _g minimizes E.

S obtained by the above first and second methods
And g, the quantization error vector e (n) is given by the following equation (4
It can be calculated as in 4).

E (n) = r (n) −g s _syn (n) (44) Quantize this as the reference input of the second-stage second encoding unit 120 _{2 in} the same manner as in the first-stage. To do.

That is, the second encoding unit 1 in the first stage
From 20 ₁ of the auditory weighting synthesis filter 312 is sent to the second perceptually weighted synthesis filter 322 of the encoding unit 120 ₂ of the signal supplied to the terminal 305 and the terminal 307 as the second stage. Further, the quantization error vector e (n) obtained by the second encoding unit 120 ₁ of the first stage is supplied to the second encoding unit 120 ₂ subtractor 323 of the second stage. .

Next, in step S5 of FIG. 10, the second coding section 120 _{2 in the second} stage also performs the same processing as in the first stage. That is, a representative value output from the stochastic codebook 320 of the 5-bit shape index output is sent to the gain circuit 321, and the gain of the 3-bit gain index output is added to the representative value output from the codebook 320 by the gain circuit 321. The output from the gain circuit 321 is multiplied by the gain from the codebook 325 and sent to a synthesis filter 322 with auditory weights. The output from the weighted synthesis filter 322 is sent to the subtractor 323, and the subtractor 323 outputs the difference between the output from the auditory weighted synthesis filter 322 and the first-stage quantization error vector e (n). Is calculated, and this difference is sent to the distance calculation circuit 324 where distance calculation is performed, and the shape vector s and the gain g that minimize the quantization error energy E are searched.

The second-stage coding unit 1 in the first stage as described above
20 ₁ of the gain index output from the shape index output and the gain codebook 315 of the strike cast codebook 310, second encoding unit 12 of the second stage
0 index output from the index output and the gain codebook 325 of the _second strike cast codebook 320 are sent to the index output switching circuit 330. Here, the second encoding unit 120
When performing the 23-bit output from, the indexes from the astcast codebooks 310 and 320 and the gain codebooks 315 and 325 of the first-stage and second-stage second encoding units 120 ₁ and 120 ₂ are combined. When the 15-bit output is performed, the above 1
The respective indexes from the cast code codebook 310 and the gain codebook 315 of the second encoding unit 120 _{1 in} the second stage are output.

After that, the filter state is updated as in step S6.

By the way, in the present embodiment, the number of index bits of the second coding section 120 _{2 in} the second stage is 5 bits for the shape vector and 3 for the gain.
A bit and very little. In such a case, if the appropriate shape and gain do not exist in the codebook, the quantization error may be increased rather than reduced.

To prevent this problem, it is sufficient to prepare 0 for the gain, but the gain has only 3 bits, and making one of them 0 makes the performance of the quantizer largely deteriorate. I will end up. Therefore, a vector having all zero elements is prepared for a shape vector to which a relatively large number of bits are allocated. Then, the above-described search is performed excluding the zero vector, and when the quantization error finally increases, the zero vector is selected. The gain at this time is arbitrary. Thus, second-stage second encoding unit 120 ₂ can be prevented from increasing the quantization error.

In the example of FIG. 9, the case of a two-stage configuration is taken as an example, but the number of stages is not limited to two, and a multi-stage configuration is also possible. In this case, when the vector quantization by the closed loop search of the first stage is completed, the Nth stage (2 ≦ N) performs the quantization with the quantization error of the N−1th stage as a reference input, and further the quantization error. Is the N + 1th stage reference input.

As described above, from FIG. 9 and FIG. 10, by using a multi-stage vector quantizer in the second coding unit, straight vector quantization with the same number of bits and a conjugate codebook as in the conventional art are performed. The amount of calculation is smaller than that using. In particular, in CELP coding,
Since the vector quantization of the time-axis waveform is performed using a closed-loop search using an analysis by synthesis method, it is important that the number of searches is small. Also, two-stage second encoding units 120 ₁ and 120 1
20 in the case of using both index outputs of the _2, (not using the second output index encoding unit 120 ₂ of the second stage) of the first-stage second encoding unit 120 using only _one index output when The number of bits can be easily switched by switching between. Further, as described above, the second encoding units 120 _{1 in the} first and second stages
If by performing the like and outputs the combined both index outputs of 120 _2, by to choose one example the decoder side after, so that can be easily associated with the decoder side. That is, for example, when a parameter encoded at 6 kbps is decoded by a 2 kbps decoder, the decoder can easily cope with it. Further, for example, the second encoding unit 120 at the second stage
By including a zero vector in the shape codebook of ₂ , even if the number of allocated bits is small,
It is possible to prevent the quantization error from increasing with less performance degradation than adding 0 to the gain.

Next, the code vector (shape vector) of the above stochastic code book can be generated as follows, for example.

For example, the stochastic codebook code vector can be generated by so-called Gaussian noise clipping. More specifically, a codebook can be constructed by generating Gaussian noise, clipping it with an appropriate threshold value, and normalizing it.

However, there are various forms of voice, and Gaussian noise is suitable for consonant voices that are close to noise such as "sa, shi, su, se, so". It is not possible to deal with consonant sounds with a sharp rise (a sharp consonant sound) such as "pi, pu, pe, po".

Therefore, in the present invention, an appropriate number of all code vectors is Gaussian noise, and the rest is obtained by learning so as to be able to deal with both the consonant with a sharp rise and the consonant close to noise. For example, when the threshold value is increased, a vector having several large peaks is obtained. On the other hand, when the threshold value is decreased, the vector approaches Gaussian noise itself. Therefore, by increasing the variation of the clipping threshold value in this way, for example, “ぱ, ぴ,
ぷ, ぺ, ぽ ”and consonants with a sharp rise, such as consonants close to noise such as さ, ，, ，, ，, ，, な ど, etc., thereby improving clarity. . Note that FIG. 11 shows the Gaussian noise shown by the solid line in the figure and the noise after clipping shown by the dotted line in the figure. Further, (A) of FIG. 11 shows a case where the clipping threshold value is 1.0 (that is, a large threshold value), and (B) of FIG. 11 shows that the clipping threshold value is 0.4. The case (that is, the case where the threshold value is small) is shown. From FIGS. 11A and 11B, when the threshold value is increased, a vector having several large peaks is obtained, while when the threshold value is decreased, Gaussian noise itself is generated. It turns out that it will be close.

In order to realize such a thing, first, an initial codebook is constructed by clipping Gaussian noise, and an appropriate number of code vectors which are not learned are determined in advance. The non-learned code vectors are selected in ascending order of variance. This is to make it correspond to a consonant close to noise, for example, "sa, shi, su, se, so". On the other hand, the code vector obtained by performing the learning uses the LBG algorithm as the learning algorithm. Here is the optimal encoding condition (Nearest Neighbor Conditio
Encoding in n) is performed using both the fixed code vector and the code vector to be learned. In the Centroid Condition,
Update only the code vector to be learned. As a result, the code vector to be learned is “ぱ,
ぴ, ぷ, ぺ, ぽ ”and so on.

The gain can be learned optimally with respect to these code vectors by performing learning as usual.

FIG. 12 shows the flow of processing for constructing the codebook by clipping the Gaussian noise described above.
Shown in

In FIG. 12, in step S10, as the initialization, the learning number n = 0 and the error D ₀ = ∞
The maximum learning number n _max is determined, and the threshold value ε that determines the learning end condition is determined.

In the next step S11, an initial codebook is generated by clipping Gaussian noise, and in step S12, some code vectors are fixed as non-learned code vectors.

Next, in step S13, the above codebook is used for encoding, in step S14 an error is calculated, and in step S15, (D _n-1 -D _n ) / D _n <
It is determined whether or not ε or n = _nmax . If the determination is Yes, the process ends. If the determination is No, the process proceeds to step S16.

In step S16, the code vector not used for encoding is processed, and the next step S16 is executed.
At 17, the code book is updated. Next, in step S18, the number of times of learning n is incremented by one, and thereafter, the process returns to step S13.

The signal coding apparatus and the signal decoding apparatus as described above can be used as an audio codec used in, for example, a mobile communication terminal or a mobile phone as shown in FIGS. 13 and 14.

That is, FIG. 13 shows a transmitting side configuration of a mobile terminal using the voice encoding unit 160 having the configuration as shown in FIGS. 1 and 3 above. The audio signal collected by the microphone 161 of FIG.
2 and is converted to a digital signal by an A / D (analog / digital) converter 163.
Sent to 60. The audio encoding section 160 has a configuration as shown in FIGS. 1 and 3 described above, and a digital signal from the A / D converter 163 is input to the input terminal 101. In the audio encoding unit 160, FIG.
The encoding process described with reference to FIG. 3 is performed, and FIG.
An output signal from each output terminal of FIG.
The output signal of 0 is sent to the transmission path coding unit 164. In the transmission path coding section 164, a so-called channel coding process is performed, and the output signal is output to the modulation circuit 165.
Is sent to the D / A (Digital / Analog)
Antenna 1 via converter 166 and RF amplifier 167
68.

Further, FIG. 14 shows the configuration of the receiving side of a portable terminal using the speech decoding unit 260 having the configuration as shown in FIGS. The audio signal received by the antenna 261 of FIG. 14 is amplified by the RF amplifier 262, and the A / D (analog / digital) converter 26
3, the signal is sent to the demodulation circuit 264, and the demodulated signal is sent to the transmission path decoding unit 265. The output signal from the H.264 is sent to the audio decoding unit 260 having the configuration as shown in FIGS. The audio decoding unit 260 performs the decoding process as described with reference to FIGS.
The output signal from the output terminal 201 in FIGS. 2 and 4 is sent to the D / A (digital / analog) converter 266 as a signal from the audio decoding unit 260. This D / A converter 2
The analog audio signal from 66 is sent to the speaker 268.

Note that the present invention is not limited to the above-described embodiment. For example, regarding the configuration on the voice analysis side (encoding side) and the configuration on the voice synthesis side (decoding side), each unit is hardware-based. However, 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 coding method and the corresponding decoding method to which the present invention is applied are not limited to the speech analysis / synthesis method using the multi-band excitation described above. It can be used or applied to various voice analysis / synthesis methods such as synthesizing an unvoiced part based on a noise signal, and the application is not limited to transmission and recording / reproduction, and pitch conversion, speed conversion, regular voice synthesis, or Of course, it can be applied to various applications such as noise suppression.

[0233]

As is apparent from the above description, according to the speech coding method of the present invention, the code for performing the vector quantization of the time base waveform by the closed loop search of the optimum vector by using the analysis method by the synthesis. There are a plurality of stages of encoding, and when encoding the Nth stage, the quantization error of the N-1th stage is used as a reference input, and the quantization output by the encoding of each stage is selected to set the bit rate. By switching,
The transmission bit rate can be easily switched, and even if the encoding side and the decoding side have different bit rates, the decoding side can easily generate an encoded data string that can be handled. it can.

[Brief description of the drawings]

FIG. 1 is a block circuit diagram showing a basic configuration of a speech signal coding apparatus to which an embodiment of a speech coding method according to the present invention is applied.

FIG. 2 is a block circuit diagram showing a basic configuration of a speech signal decoding apparatus to which an embodiment of a speech decoding method according to the present invention is applied.

FIG. 3 is a block circuit diagram showing a more specific configuration of the audio signal encoding device according to the embodiment of the present invention.

FIG. 4 is a block circuit diagram showing a more specific configuration of the audio signal decoding device according to the embodiment of the present invention.

FIG. 5 is a block diagram illustrating a basic configuration of an LSP quantization unit.

FIG. 6 is a block diagram illustrating a more specific configuration of an LSP quantization unit.

FIG. 7 is a block diagram illustrating a basic configuration of a vector quantization unit.

FIG. 8 is a block diagram illustrating a more specific configuration of a vector quantization unit.

FIG. 9 is a block circuit diagram showing a specific configuration of a CELP coding section (second coding section) of the speech signal coding apparatus of the present invention.

10 is a flowchart showing the flow of processing in the configuration of FIG.

FIG. 11 is a diagram showing a state of Gaussian noise and noise after clipping with different threshold values.

FIG. 12 is a flowchart showing a flow of processing when a shape codebook is generated by learning.

[Fig. 13] Fig. 13 is a block circuit diagram showing a configuration of a transmission side of a mobile terminal to which the audio signal encoding device of the present invention is applied.

FIG. 14 is a block circuit diagram showing a configuration of a receiving side of a mobile terminal to which the audio signal decoding device of the present invention is applied.

[Explanation of symbols]

110 first coding unit 111 LPC inverse filter 113 LPC analysis / quantization unit 114 sine wave analysis coding unit 115 V / UV determination unit 120, 120 ₁ , 120 ₂ second coding unit 121 noise codebook 122, 312,322 Weighted synthesis filter 123,313,323 Subtractor 124,314,324 Distance calculation circuit 125 Auditory weighting filter 302 LPC analysis circuit 303 LPC parameter quantization circuit 304 Auditory weighting filter 310,320 Stochastic codebook 315,325 Gain codebook 330 Index output switching circuit

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Shiro Omori 6-735 Kita-Shinagawa, Shinagawa-ku, Tokyo Sony Corporation

Claims (6)

[Claims] 1. A speech coding method in which an input speech signal is divided into blocks on the time axis and is coded in each block, wherein a time axis waveform by a closed loop search of an optimum vector using an analysis method by synthesis. Has a plurality of stages of vector quantization, and the Nth stage of the encoding process uses the N−1th stage quantization error as a reference input, and all of the quantization outputs of the respective stages of the encoding process. Alternatively, a voice encoding method is characterized in that a bit rate is selected and a bit rate is switched. 2. The voice encoding method according to claim 1, wherein a signal of an unvoiced sound portion of the input voice signal is encoded by the encoding step. 3. In each of the encoding steps, a quantized value is obtained by multiplying a shape vector extracted from the shape codebook by a gain from the gain codebook, and the shape codebook has a shape vector of size 0. The speech coding method according to claim 1, wherein 4. A speech coder that divides an input speech signal into blocks on the time axis and performs coding on a block-by-block basis. In a speech coder, a time-axis waveform by a closed-loop search of an optimum vector using an analysis method by synthesis. Has a plurality of stages of vector quantization, and the Nth stage encoding unit uses the N-1th stage quantization error as a reference input, and all the quantization outputs of the encoding units of each stage. Alternatively, a voice encoding device is characterized in that a bit rate is selected and a bit rate is switched. 5. The speech coding apparatus according to claim 4, wherein the unvoiced part of the input speech signal is coded by the coding means. 6. In each of the encoding means, a quantized value is obtained by multiplying a shape vector extracted from a shape codebook by a gain from a gain codebook, and the shape codebook has a shape vector of size 0. The speech coding apparatus according to claim 4, wherein the speech coding apparatus is a speech coding apparatus.