JPS5994797A - Adaptive conversion coding system for voice - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an adaptive transform coding method that transforms an audio signal into a frequency domain and adaptively changes its quantization.

<Prior art> This type of audio encoding method is known, for example, from Japanese Patent Application Laid-Open No. 55-5'7.
No. 900 "Audio Signal Processing Circuit". In this method, as shown in FIG. 1, the input audio from the input terminal 11 is sampled at, for example, 8 kHz, and each sample value is input as a digital signal to the orthogonal transform section 12. A fixed number of input audio samples S1...S2n shown in are subjected to discrete Fourier transform to produce a frequency domain signal (spectrum) f1f2...
fn (FIG. 2B) and sent to the adaptive quantization section 13. On the other hand, the input voice of the terminal 11 is input to the spectral envelope extraction section 14, and the spectral envelope of the input voice is estimated by linear prediction 61 analysis, and the spectral envelope and pitch period are supplied to the adaptive information assignment section 15.

The adaptive information allocation unit 15 receives a frequency domain signal f>f2...
...In accordance with the instantaneous level of the spectral envelope for each of fn, each bit in the quantization unit 13 is set in consideration of the pitch period, so that if this level is thick, more bits are allocated, and if this level is small, fewer bits are allocated. signal f1
f2... Mineralizes adaptively with quantization for fn. The quantized □H\ information and the information indicating bit allocation are combined in the combining circuit 16 and sent out as encoded output.

With this method, the audio signal of 8KHz sampling is
It can be efficiently encoded with an amount of information of about 6 Kbps, and high quality speech can be obtained. However, the bit allocation information has 2K
Since the amount of information about bpS is required, the total amount is 9',
6 r<bps (1 of commonly used transmission layers)
1) When encoding with the following information amount, the signal f1. f
2... It is necessary to quantize fn to 1 bit/sample or less. At this time, almost no information can be assigned to sections of low intensity among frequency components, resulting in significant deterioration of voice quality.

<Summary of the Invention> This invention converts an input audio signal into the frequency domain, divides the converted spectrum into blocks, and adaptively performs vector quantization in each block according to spectral envelope information. For example, the encoding speed is 9.6Kbl)
An object of the present invention is to provide an adaptive conversion encoding system for audio that reduces deterioration in audio quality even when the audio quality is lower than S. Further, vector quantization can be efficiently performed by flattening the spectrum before dividing the spectrum into blocks.

<First Embodiment> FIG. 3 shows an embodiment of the speech encoding system according to the present invention. The input signal from terminal 11 is quadrature f? In unit A 12, each frame is converted into a frequency domain signal, that is, a spectrum, by orthogonal transformation such as discrete Fourier transform (DFT) or discrete cosine transform (DCT), and this spectrum is converted into a frequency domain signal, that is, a spectrum. In this step, the spectral envelope is globally flattened using separately determined and quantized spectral envelope information.

That is, the spectral envelope of the input voice at the terminal 11 is estimated by linear predictive analysis in the spectral envelope extraction section 14,
This spectral envelope information and audio power are quantized as auxiliary information in the quantization section 18, and this quantized output is
5 is decoded by the post-coding unit 19, and the cis vector f1'f2...fn is divided by the decoded iIi auxiliary information by the spectrum smoothing unit 17.

This flattened spectrum is divided into p consecutive blocks by the block dividing unit 21 as shown in FIG.
(f 11 f 12-...ftp), F2==(f
2tf22...f2),...Fs-(fS
lfS2...fSp). Spectral gender component ij (1==l...S.

j=1...p) consists of a real part R(fij) and an imaginary part (fij), and the vector R(F 1)=(R (f 11
)R(f 12 )...R(f'tp)'), R(
F2) - (R(f2t)R(fz2)...R(f
2p))”・R(Fs)=(R(fsx)R(fs2)
“”・R(fsp)) and 8 which similarly uses the valley imaginary part as an element.
Create a vector I(Fi)=(I(f i j).

The vector quantization unit 22 detects which standard vector in a dictionary prepared in advance these vectors most closely corresponds to, and performs vector quantization.

Store multiple predicted standard vectors as a dictionary, detect which standard vector the input speech vector is close to, and output a code such as a number indicating the matching or similar standard vector. do. Therefore, the intensity of each spectral component can be encoded with a smaller number of bits than when quantized. Furthermore, the bit allocation for this vector quantization is adaptively changed.

That is, the number of bits is assigned to each block according to the decoded output of the auxiliary information that is the output of the local decoding section 19. Generally, a large number of bits are allocated to blocks containing strong spectra, and when a large number of bits are allocated to the weak spectrum quantization unit 22, a dictionary with a large number of standard vectors to be compared is referred to, When a small number of bits is allocated, a dictionary with a small number of standard vectors is referred to. Since the number p of elements of a standard vector is constant, a dictionary with a large number of standard vectors is stored.Standard vectors that display even minute patterns are stored in a dictionary with a small number of standard vectors. It can be said that the marker vector shown only shows a rough pattern.

This adaptive information allocation (bit allocation) is performed with the aim of maximizing the S/N ratio of the input signal and output signal for each frame. Since the SN ratio remains unchanged even after orthogonal transformation, it is only necessary to minimize the distortion between the output of the spectral smoothing section 17 of the encoder 24 and the spectral reproduction output of the decoder 25 on the receiving side, and the distortion scale is the Euclidean distance. 1
The distortion per frame is as follows.

The total number of samples (spectrums) is still p's, and the average information amount (average number of quantized bits) B for each sample is B=ri bj/(p-8) −1. Since B is held constant, the number of quantization bits bj that minimizes distortion is given by the following equation.

Quantization is performed by converting bj to an integer value and selecting the one with the minimum distortion from a dictionary consisting of 25j items.

Note that the quantization in the quantization unit 18 can also be vector quantization. The quantized output of this spectrum envelope, that is, the auxiliary information, and the waveform information that is the output of the vector quantizer 22 are combined and sent to the decoder 25 as an encoded output.

In the decoder 25, the input waveform information is processed by a smoothing spectrum reproducing unit 26, which uses the same dictionary as that used by the vector quantizer 22 in the encoder 24 to convert the palm quasi-vector into a quantized code for each block. to reproduce the smoothed spectrum.

On the other hand, the spectrum envelope of the input auxiliary information is reproduced by the spectrum envelope reproduction section 27, and the spectrum envelope is multiplied by the power by the reproduced smoothed spectrum in the spectrum reproduction section 28 to reproduce the spectrum.

The reproduced spectrum is inversely transformed into the time domain by the inverse transformer 29 to obtain a reproduced audio signal at the output terminal 31.

<Second Embodiment> In the above description, spectrum smoothing was performed after orthogonality adjustment, but orthogonal transformation may be performed after passing the input audio through an inverse filter. For example, as shown in FIG. 5 with the same reference numerals attached to parts corresponding to those in FIG.
2.

- The spectral envelope of the actual input audio signal is analyzed by the linear prediction analyzer 33, the analyzed prediction coefficients are vector quantized by the quantizer 18, the quantized output is decoded by the local decoder 19, and the decoding The filter constant of the inverse filter 32 is controlled by the output, that is, the linear prediction coefficient. The output of this inverse filter 32 is a residual signal, which is orthogonally transformed, encoded in the same manner as described above, and sent out. In the decoder 25, the vector quantized code is decoded by the spectrum reproducing unit 26 to reproduce the spectrum of the residual signal, which is inversely transformed into the time domain and sent to the linear prediction synthesis filter unit 34.

The filter constant of this synthesis filter section 34 is controlled by the prediction coefficient reproduced by the spectrum envelope reproduction section 27,
The filter section 34 reproduces the audio signal.

FIG. 6A shows the waveform at of the input audio signal, the real part waveform a2, and the imaginary part waveform a3 of the orthogonal transform output, and FIG. 6B
shows a residual signal waveform b1 after passing the input audio signal waveform a1 through the inverse filter unit 32, a real part waveform b2 of the orthogonal transform output of the residual signal, and an imaginary part waveform b8.

The spectrum envelope 11b4 of the audio input waveform al and the allocated bits b5 for each block are shown, respectively. Tasoshi p
==13 and B=1.0.

In the above, by adapting the dimension P of the vector that is the unit of quantization to the pitch frequency of the input audio and making the length of one frame an integral multiple of the pitch period, the efficiency of ;j quantization can be further improved. In this case, since the pitch frequency changes over time, the pitch frequency is also included in the auxiliary information.

Furthermore, it is also possible to process vectors as a unit of complex numbers, without making the real and imaginary parts independent. Furthermore, each of the above-mentioned units can be processed independently or in a common electronic calculation manner.

<Effects> As explained above, quantization efficiency can be increased by dividing a signal flattened in the frequency domain into blocks and adaptively allocating information, especially when scalar quantization is performed at 9.6 Kbps or less. It is possible to reproduce audio with a higher SN ratio than the conventional adaptive transform coding method. Due to frequency domain flattening, the number of standard vectors for vector quantization can be reduced. Further, the amount of information allocated to one block only needs to be an integer, and the amount of information per sample is allocated in detail in units of 1/P bits. In this way, it is possible to avoid auditory deterioration caused by the existence of frequency components to which no amount of information is assigned, which is a drawback of the conventional method, and which is adaptively changed.

Next, an experimental example will be described. The sampling frequency is 8KHz,
The analysis order of the linear prediction analysis section 33 is 8th order, and the analysis length (conversion section) is 26 to 31 m5. FIG. 7 shows the S/N ratio for the information amount B (bits/sample) when the analysis overlap is 2m5 (connected by trapezoidal windows) and the vector dimensions are 6 to 12 (pitch adaptation). In Fig. 6, the curve g141 is uniform quantization and each sample is encoded, the curve 42 is uniform quantization and 6-dimensional fixed vector encoding is performed, the curve 43 is
Curve 44 is the case when encoding is performed by uniform quantization with the dimension of the vector changed according to the pitch frequency, curve 44 is the case when each sample is encoded using adaptive quantization (conventional method), and curve 45 is
curve 46 represents the case where the method of the present invention performs encoding by six-dimensional fixed vector quantization, and the curve 46 represents the case where the method of the present invention performs encoding by adaptively changing the dimension of the vector according to the pitch frequency. All these, uniform quantization (song @41~
43) In theory, adaptive quantization (curves 44 to 46) is better, and even in adaptive quantization, the invented method (curves 45 and 46) is better than the conventional method (curve 44). corner! will be played. In the 0.5 to 1.1 bit/sample range, this S/N ratio improvement is 2.5 dB for female voices even outside of the training samples.
, x in a male voice. Obtained oaBs degrees. By vector quantizing the spectrum envelope, it can be estimated that about 800 bps of auxiliary information including pitch, power, etc. is required, so the amount of information B per sample of the residual signal is 0. 5
So 4. 8 Kbps, 1. 9.6Kbps in 1
It is possible to encode

[Brief explanation of drawings]

Hs 1 is a block diagram showing a conventional adaptive transform encoding method, FIG. 2 is a diagram for explaining its operation, FIG. 3 is a block diagram showing an example of an adaptive transform encoding method according to the present invention, and FIG. 4 is a diagram showing an example of block division, FIG. 5 is a block diagram showing another example of this invention, FIG. 6 is a diagram showing an example of its operation, and FIG. 7 is a diagram showing SN ratio vs. information amount of various encoding methods. B
It is a diagram showing the relationship with □. 11: Audio input, 12: Orthogonal transform unit, 14 Varnish vector envelope extractor, 17: Spectral smoother, 18: Vector quantizer, 19: Local decoder, 21 Niblock division unit, 22: Vector quantizer , 23: Adaptive information allocation unit, 2
4: Encoder. Patent Applicant: Representative of Nippon Telegraph and Telephone Public Corporation Kusano Kusano Sojiru 1 Figure 7172 Figure Bowl 3 Figure

Claims (2)

[Claims] (1) The sample value series of the audio signal is divided into 1
In an encoding method that calculates a cis vector by orthogonal transformation for each frame and adaptively quantizes it,
Spectral envelope extraction means for determining the envelope of the spectrum, quantizing it together with power, and encoding it as auxiliary information;
local decoding means for decoding the auxiliary information; smoothing means for converting the spectrum into a spectral signal sequence flattened on the frequency axis using the decoded auxiliary information; block dividing means for dividing into blocks on the frequency axis; adaptive information allocation means for adaptively allocating information to each divided block using the above-mentioned auxiliary information; and vector quantization means for vector quantizing a spectral signal sequence. (2) In a coding method that analyzes and encodes the sample value series of the audio signal in units of one frame, the audio signal is analyzed with linear prediction, its spectral envelope is determined, and the power and spectral envelope are quantized and encoded as auxiliary information. a linear prediction analysis means; a local decoding means for decoding the auxiliary information;
an inverse filter means whose filter constant is controlled by the linear prediction coefficient of the decoded auxiliary information and which receives the sample value series of the audio signal and outputs a residual signal; orthogonal transformation means for obtaining a spectrum by orthogonal transformation; block division means for dividing this spectrum into blocks on the frequency axis; and adaptive information allocation for each divided block using the above-mentioned auxiliary information. 1. An adaptive transform coding method for speech, comprising an adaptive information allocation means and a vector quantization means for converting the divided spectral residual signal sequence into vectors/synthesizers according to the allocation.