JPH0445614A - One bit/sample quanitzation method - Google Patents

Abstract

PURPOSE:To realize one bit/sample quantization method by quantizing (m) sample values of a residual signal being the subtraction of a predictive signal from an input signal in m-bit with minimum distortion in a lump and sequentially compressing and expanding the roughness of the quantization of succeeding samples. CONSTITUTION:A signal Xn from an input terminal 21 is subtracted from a prediction signal X'n at a subtractor 22 and the residual signal En is normalized by a divider 23 and quantized in a lump in m-bit for (m) samples at a quantizer 24. Its vector number is inversely quantized by an inverse quantizer 25, multiplied with a normalizing level S at a multiplier 26 and its multiplication output E'n is added to a predictive signal X'n from an adaptive predictive device 27 at an adder 28 and a local decoding signal Y'n is obtained and given to the adaptive predictive device 27. A normalized level S is adaptively changed by a normalizing level adaptive processing section 29 in response to the state of the quantizer 24.

Description

[Detailed Description of the Invention] "Industrial Application Field" This invention divides a signal such as music or voice into bands, performs adaptive bit allocation to each band, and performs adaptive differential pulse code modulation for each band. Case 1
This invention relates to a quantization method used in ADPCM.

"Prior Art J The configuration of the conventional band division adaptive bit allocation ADPCM system is shown in FIG. 4. The input signal from the terminal 11 is, for example, 48i
This input signal is a voice or music signal sampled at iHz, and this input signal is divided into a plurality of subbands by a band division filter 12. For example, when divided into four, the sub-bands are O~6kl(z, 6~12kl(z, 12~18
kHz, 18 to 24kTo, and this band division is QMF (Quadrature former rrorFil
This is realized by the vertical connection of ter). The signals of each divided subband are encoded by the ADPCM quantizer 13, respectively. The adaptive bit side assigning unit 14 calculates the number of quantization focuses proportionally distributed to the logarithm of the electric cassette vector density of the quantized residual signal obtained from the encoded output so that the quantization distortion of the entire band is minimized. , are adaptively allocated to each ADPCM placer 13. The codes of each sub-band obtained from each ADPCM quantizer 13 are configured into a frame by a decoder 15 and output.

FIG. 5 shows an example of quantization characteristics when the ADPCM quantizer 13 operates at 2 bits/sample.

The input level in this figure corresponds to the actual input level divided by the normalized level S (also called quantization step width, a value corresponding to the short-time effective value of the signal), and has an average value of 0 and an effective value of 1. It is designed to minimize the quantization error for Laplace distributed signals. If the quantization value is appropriately selected depending on the level distribution characteristics of the applied sound source, more desirable characteristics can be obtained. In this example, the input is quantized to four levels, and each sample is quantized to two bits.

In addition, the normalization level adaptation rule for 2-bit/sample quantization is shown in Figure 6.As is clear from this figure, when the absolute value of the input level is larger than the normalization level, the normalization level is normalized at the next sample point. The normalized level is expanded, and when the absolute value of the input level is smaller than the normalized level, the normalized level is compressed at the next sample time. Through this operation, it is expected that the normalized signal will always have a thickness appropriate for the discrimination level of the quantizer.

"Problems to be Solved by the Invention" In the bit allocation to each subband in FIG. An allocation of ~1 bit may be desirable. However, the conventional ADPCM quantizer 13 requires a minimum number of quantization bits of 2 bits/sample. This is because when the ADPCM quantizer 13 decides to compress or expand the normalization level S, which is the standard for quantization, it
It depends on the information on which quantization level the quantization value before the sampling point is at, and for this, at least two quantization levels are required for each of the positive and negative Shingetsu,
This is because the required number of encoding levels cannot be obtained with 1 bit/sample that can only represent positive and negative signals.

If we observe the quantization history of multiple samples and, for example, if positive values continue multiple times, we can compress and expand the normalization level by expanding the normalization level.It is possible to compress and expand even 1 bit/sample. , which is ADM (
This is well known as an adaptive differential modulation (differential modulation) encoding method. However, in this case, it is necessary to observe the quantization history of several samples, and this method is applied to the adaptive bit allocation ADPCM method.
In this case, when transitioning from a frame operating with quantization of 2 bits/sample or more to a frame operating with 1 bit/sample quantization, at the beginning of the frame operating with 1 bit/sample quantization, the current quantization The conversion history is 2
The history of quantization is more than bit/sample, and there is no history of quantization of 1 bit/sample.Therefore, a predetermined quantization history of 1 bit/sample is used, but the compression/expansion at the normalization level is It is easy to become inappropriate, and
A 1-bit/sample quantizer was not adopted because the quantization error added to the input to the predictor becomes excessive with 1-bit/sample quantization, making prediction operation with high gain impossible. There wasn't.

For this reason, conventional band division adaptive bit allocation ADPC
The M method consumes 2 bits/sample even in the treble (high frequency) subband where there is almost no signal, which has the disadvantage that the noise level increases in the low frequency band, which requires a larger number of quantization bits. there were.

The purpose of this invention is to solve these problems by:
The objective is to provide a quantization method that operates on bits/sample.

"Means for Solving the Problem" According to the present invention, m sample values (m is an integer of 2 or more) of a residual signal obtained by subtracting a predicted signal from an input signal are collectively processed using m bits based on a minimum distortion measure. Then, the roughness of the quantization of subsequent sample values is sequentially compressed and expanded according to the history of the quantization code up to the present.

Embodiment FIG. 1 shows an embodiment of the present invention, in which, for example, a signal of one sub-band divided into bands shown in FIG. 4 is quantized. The signal L from the input terminal 21 (n represents the sampling time point) is the predicted signal sentence, and the subtracter 22
The residual signal E, is level-normalized by dividing by the normalization level S in the divider 23, and the normalized residual signal e,- is subtracted by the quantizer 24 into m samples. (m is 2
Each integer above) is collectively quantized with m bits, that is, vector quantized, and a code (vector number) indicating the vector is output, and the code is inversely quantized by the inverse quantizer 25. , the inverse quantization output L(i) is multiplied by the normalization level S in the multiplier 26, and the multiplication output is
is the prediction signal sentence from the adaptive predictor 27 as the inverse quantization prediction residual signal.

are added by a stop adder 28 to obtain a local decoded signal V,
This local decoded signal? , is input to the adaptive predictor 27.

Normalization level adaptation unit 29 according to the state of quantizer 24
The normalization level S is adaptively changed.

Next, this quantization operation will be specifically explained for m=2.

(b) Set the initial value of the normalization level S to 1. If the normalization level is defined in the previous frame, that value is inherited.

(b) Predicted signal from input signal X! , is subtracted, and the resulting prediction residual signal E is divided by the normalization level S.

(c) At the sampling time n, from the inverse quantization value table, for example, the table in FIG. 2, the inverse quantization value @++(i), i=1.2
．． ..., find 2@. Generally, there are 2- inverse quantization values, and in the example of FIG. 2, @1l (ll=0.395).

@, (21=0.395, @, (3)=-0
,395,@, (4)=0.395. Each of these inverse quantization values @, (i) is multiplied by the normalization level S to obtain an inverse quantization output, and the sum of this and the prediction signal at sampling time n, that is, the locally decoded signal 7. This is input to the predictor 27.

(d) Prediction residual signal E at sampling time n+1
−+(i) is calculated.

(e) Repeat (c) and (d) for the number of dequantized values 2'' in the dequantized value table (Figure 2).In other words, when m=2, the predicted signal is 0 for the four @ and (i). .1 is obtained, and the predicted signal M1.8 is obtained for the four @ka, , (i).

(f) Select the squared error and vector at sampling time points n and n+1, and use the code vector number as the quantized output.

That is, (ea −@s (]) )” ten (e++, + −@
a+-1(1))"(e,-kai,(2))"+(
e@*1 @−+ (2) )”(e,, −@, (
3) door + (e+t, + -@11.+ (3) )”(
ea −@s (4) )” + (ea−+−@□I
(41)'', and determine the minimum i, and use this as the quantized output.

In addition, Fig. 2 shows the adaptive bit allocation coding method.
When the bit allocation changes from a frame with sample quantization to a frame with 1 bit/sample quantization,
Since both of these frames have a continuous nature, the inverse quantization value of 1 bit/sample quantization and the inverse quantization value of 2 bit/sample quantization! As for the L image, it was experimentally determined from FIG. 5 so that the quantization error would not increase during frame transition, and the distortion during 1-bit/sample quantization would be small. Therefore @, 0.3950 of (i)
．． 395 corresponds to the dequantization output level in FIG.

(g) The normalization level for encoding at sampling time (n+2.n+3) is the code vector number (quantization output) i for each of the previous sampling time (n-2, nl) and (n n+1). When both are 2 or 4 in Fig. 2, it is considered to be larger than the short-term effective value of the input signal, so the normalization level is multiplied by a value smaller than 1, for example 0.8, and the previous sampling point (
When the code vector numbers i for each of n-2n-1) and (n, n+1) are both 1 or 3, the normalization level is considered to be smaller than the short-term effective value of the input signal, so it is greater than 1. Multiply the value, for example 1.25. The normalization level is saved for other code vector numbers.

(H) Repeat steps C) to (G) for continued sample values (n=1.3.5...frame length).

As the adaptive predictor 27, a Barcall lattice type one shown in FIG. 3 is suitable. For 1 bit/sample quantization,
Since quantization errors affect the calculation of prediction coefficients and high-order predictions are meaningless, it is better to keep the prediction order to about 2.The correlator 31 in Fig. 3 has a coefficient km (n+1).
N, (n>=(1-r) ・N, (
n-1)-2r-er, (n) ・eba(n) Da(
n)=(1-r) HD+e (n-1)+r ・(
era”(n)+eba”(n)) to m++(n+1)
= N-(n)/Da(n)r・0.03.

Similarly, when m>2, the optimal code vector can be selected by performing 2''' error calculations for vectors of degree m from sampling time n to sampling time n+m-1.In this case, the dequantized value The table can be created in advance by selecting two items that have the highest probability of appearance, reflecting the statistical properties of the prediction residual signal.

"Effects of the Invention" As explained above, according to the present invention, m-bit/m-sample quantization, that is, 1-bit/1-sample quantization is performed. When this 1-bit sample quantization method is applied to the band division adaptive focus allocation ADPCM method, stable and low-distortion operation can be achieved with a minimum bit allocation of 1 bit/sample while obtaining a high prediction gain using the Percoll lattice predictor. Therefore, it is possible to maintain the characteristics of the treble subband above a certain level, and to perform bit allocation very close to the optimum bit allocation according to the characteristics of the input signal, thereby improving the encoding characteristics. In addition, even though quantization is performed as a vector, there is no need to separately quantize and transmit information related to amplitude, so there is no need to significantly change the configuration of the conventional ADPCM system, low delay time, and separate information The advantage of the ADPCM method of not requiring transmission is preserved.

[Brief explanation of drawings]

FIG. 1 is a diagram showing a configuration example of the ADPCM encoding method according to the present invention, FIG. 2 is a diagram showing an example of an inverse quantization value table, FIG. 3 is a diagram showing a Percoll lattice filter as an adaptive predictor, and FIG. Figure 4 shows conventional band division adaptive bit allocation ADPCM
Figure 5 is a diagram showing the encoding level and decoding level of conventionally used 2-bit quantization, Figure 6 is the normalization level compression during 2-bit quantization,
FIG. 3 is a diagram illustrating an adaptation rule for decompression.

Claims (1)

[Claims] (1) Adaptive bit allocation of the signal A quantization method used for ADPCM (adaptive differential pulse code modulation), in which m sample values (m is an integer of 2 or more) of the residual signal obtained by subtracting the predicted signal from the signal are used. are quantized in m bits using the minimum distortion scale, and the coarseness of the quantization of subsequent sample values is sequentially compressed and expanded according to the history of the quantization code up to the present. Bit/sample quantization method.