JPS625378B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to differential pulse code modulation used for band compression of audio signals.
Modulatisr-DPCM) method improvement.

Various techniques have been developed to compress and transmit or store audio signals, focusing on the redundancy of audio signals. The DPCM method is a band compression method that uses the property that the amplitude of an audio signal at a certain time can be predicted fairly well using past audio signals because audio signals have a fairly large correlation. the easiest
An example of the DMCM method is to predict the signal amplitude at a certain time using a signal one sample before, and then quantize the difference between the actual signal and the predicted value and transmit or store it.

In order to perform band compression more effectively, it is sufficient to use a plurality of samples instead of just one past sample. Furthermore, if the weighting coefficients used in prediction are optimally matched to the audio signal to be encoded, the degree of band compression will be further increased.

However, the degree of band compression obtained by encoding an audio signal by DPCM is about 2 bits or more. Therefore, even if the above-mentioned technique is used, for example, when an audio signal is encoded by DPCM with 4-bit quantization, the SN ratio obtained is about 30 dB. Like this
The noise generated by quantization in the DPCM method (quantization noise) can be considered almost white noise. Therefore, when you listen to the reproduced sound, it sounds as if a noise called "sards" is superimposed on the real audio signal. The reason for this is that the high-frequency spectrum of noise is stronger than that of speech. A noise shaping method is used as a method of compressing the high-frequency components of such noise. This method is described in the publication “IEEETRANSACTIONS VOL ASSP−27No.” published in February 1979.
1 Pages 63-73” “Adaptive Noise”
Spectral Shaping and Entropy Coding in
The noise shaping method is explained in detail in the paper entitled "Predic-tive Coding of Speech", so here we will briefly explain the noise shaping method using diagrams.

FIGS. 1 and 2 are block diagrams showing a conventional DPCM encoding device using a noise shaping method. FIG. 1 shows a DPCM encoding device, in which an audio signal X _j to be encoded is input from terminal 1, a predicted value including quantization noise is subtracted by a first subtracter 10, and then quantized by a quantizer 20. Ru. The quantized signal e^ _j is output from the terminal 2 to the DPCM decoding device and is also input to the predictor 30. The first adder 32 supplies the first filter 31 with a signal X^ _j obtained by adding the quantized signal e^ _j and the predicted value X^ _j . If the quantizer 20 is not provided, the signal X〓 _j and the input signal X _j match. The predicted value X^ _j is calculated using the following formula.

X＾ _j =a ₁ X〓 _j-1 +a ₂ X〓 _j-2 +……+a _N X〓 _jN …(1
) Here, a ₁ , a _{2 .} . . a _N is a weighting coefficient, and N is the number of weighting coefficients. Such a filter is called a transversal filter. The weighting coefficient may be set in advance to a fixed value, or may be adaptively modified from time to time depending on the nature of the audio signal. Since this adaptation technique can also be referred to in the above-mentioned literature, its explanation will be omitted. Quantizer 2
The input e _j of 0 and the output e^ _j are input to the second subtractor 50 . The output −n _j of the second subtractor 50 is equal to the noise component generated by the quantizer 20. Signal-n _j
is added to the second adder 60 via the second filter 70.

Next, the DPCM decoding device
It is composed of an adder 32' and a third filter 31'. These are exactly the same as the first adder 32 and first filter 31 of the DPCM encoding device. Unless there is an error in the transmission path, the third adder 3
The output X〓' _j of 2' matches the X〓 _j of the DPCM encoding device. In the above-described DPCM system, the relationship between the reproduced signal X〓' _j (corresponds to X〓 _j when there is no transmission error) and the quantization noise n _j is expressed by the following equation.

X〓' _j =X _j +(1+B)n _j (2) Here, B is the transfer characteristic of the second filter 70. Let us now assume that the transfer characteristic of the first filter 31 is A. Predictor 7 predicts the transfer characteristic B of the second filter 70
In order to make the transfer characteristic completely equal to 0, it is sufficient to set B=A/(1-A) …………(3). Substituting equation (3) into equation (2), we get X〓' _j =X _j + (1/(1-A))n _j
………(4) becomes. If the first filter 31 is the optimal filter for speech, then 1/(1-A) is an approximation of the speech power spectrum, so from equation (4), the quantization noise n _j is It can be seen that the spectrum is the same as that of voice. As described above, the noise shaping method eliminates the "sears" and various other harsh sounds caused by quantization noise.

However, if the second filter 70 is made of a polar model as shown in equation (3), the low frequency range is emphasized too much, so that low frequency noise such as "rumbling" becomes audible. Therefore, it is pointed out in the above-mentioned literature that the second filter 70 should be realized as a zero-type model (transversal filter). In other words, the second filter 70 must perform the calculation b ₁ n _j-1 +b ₂ n _j-2 +...+b _M n _j - _M (5) on the input signal n _j . The method of converting the weighting coefficients a ₁ , a ₂ . . . a _N of the first filter 31 into _the coefficients b ₁ , b ₂ . It's complicated.
If the weighting coefficients of the first filter 31 are fixed in advance, the coefficients of the second filter 70 may also be determined at the same time. However, the first filter 3
When a weighting factor of 1 is changed depending on the voice, the amount of calculation for conversion becomes extremely large.

The purpose of the present invention is to minimize the influence of quantization noise on reproduced sound, and to reduce the scale of the device.
The purpose is to provide a DPCM encoding device.

Next, the present invention will be explained in detail using the drawings.

FIG. 3 is a block diagram showing an embodiment of a DPCM encoding device according to the present invention. In this embodiment, the predictor 30 in FIG. 1 is composed of a fourth filter 130. The fourth filter 130 is a transversal filter (zero model), and its internal calculations are expressed as in the following equation.

X^ _j =a ₁ e^ _j-1 +a ₂ e^ _j-2 +...+a _N e^ _j - _N
...... ₍ 6) Here, the coefficients a ₁ , a ₂ . The greatest advantage of using the method of prediction using a transversal filter as in this embodiment is that stability on the receiving side is guaranteed. In conventional embodiments, the decoder consists of a recursive filter as shown in FIG. Therefore, when the first filter 31 and the third filter 31' operate adaptively, if an error occurs in the transmission path, the first filter 31 and the third filter 31'
There is a possibility that the decoder may become unstable due to a difference in characteristics between the filter 31' and the filter 31'. On the other hand, if the predictor is configured with a transversal filter as in this example,
As will be described later, the decoder is not a recursive filter but a transversal filter, so even if there is an error in the transmission path, it will not become unstable. Correcting the coefficient of equation (6) is a _i = a _i (1-δ) + g・e^ _j・e _ji ……(7) a _i = a _i (1-δ) + g・₁ (e ＾ _j )・₂ (e＾ _ji ) ......(8) Here, g and δ are positive numbers sufficiently smaller than 1, and ₁ and ₂ are arbitrary monotonically increasing nonlinear functions. A method of changing the coefficients a ₁ , a _{2 .}
Since the specification of the present invention can be referred to, a detailed explanation of the operating principle will be omitted here, and a brief explanation will be given with reference to FIG. Similar to the conventional embodiment shown in FIG. 1, the quantized signal e^ _j is processed in the fourth filter 130 as shown in equation (6). Fourth filter 1
The output X^ _j of 30 is added to the output of the second filter 70 and input to the first subtracter 10. It is assumed that the coefficients of the second filter 70 for noise shaping are the same as the coefficients of the fourth filter 130.

FIG. 4 is a block diagram showing an embodiment of a DPCM decoding device according to the present invention. The signal output from the DPCM encoding device is input from the terminal 3 and input to the fifth filter 130 and the fourth adder 80.
The fifth filter 130' has exactly the same characteristics as the fourth filter 13. Therefore, the output X^' _j of the fifth filter 130' in the decoder coincides with the output X^' _j of the fourth filter 130. Assuming that the transfer functions of the fourth filter 130, the fifth filter 130, and the second filter 70 are A, it can be seen from FIG. 3 that the following equation holds true.

−n _j =e _j −e＾ _j ………(9) e _j =X _j −A′e＾ _j +A′n _j ………(10) From equations (9) and (10), e ^ _j +A′e^ _j =X _j +n _j +A′n _j ………(11) is obtained. Therefore, the regenerated signal X _j of the decoder is X〓' _j =X^' _j +e^ _j =X _j +(1+A')n _j ......
It appears as (12). If the fourth filter 130 changes optimally according to the audio signal (1
+A) closely approximates the power spectrum of the audio signal. Therefore, the quantization noise contained in the reproduced signal X〓' _j resembles the spectrum of the audio signal. As a result, the high-frequency noise called "saerts" becomes extremely small. Since the second filter 70 is a zero-type filter, there is no surplus gain in the low frequency range, resulting in very high quality sound with little low frequency noise. In this embodiment, since the coefficients of the second filter 70 are the same as the coefficients of the fourth filter 130, calculations for conversion are completely unnecessary, and the amount of calculations for the DPCM apparatus as a whole does not increase much.

As described above, according to the present invention, it is possible to obtain a DPCM encoding device that has less high-frequency noise due to quantization, no unpleasant low-frequency noise, and is small in size.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of a conventional DPCM encoding device, FIG. 2 is a block diagram showing an embodiment of a conventional DPCM decoding device, and FIG. 3 is a block diagram showing an embodiment of a conventional DPCM encoding device. Block Diagram Showing One Embodiment FIG. 4 is a block diagram showing one embodiment of the DPCM decoding apparatus according to the present invention. In the figure, 1 is an input terminal, 2 is an output terminal, and 3 is an input terminal.
is an input terminal, 4 is an output terminal, 10 is a first subtractor, 20 is a quantizer, 50 is a second subtractor, 60
is the second adder, 70 is the second filter, 80 is the fourth adder, 130 is the fourth filter, 13
0' is the fifth filter.

Claims (1)

[Claims] 1 Quantize and output the difference between the input signal and predicted value,
The quantized signal is input to a first transversal filter whose characteristics adaptively change, and the difference between the signals before and after the quantization is input to a first transversal filter having the same characteristics as the first transversal filter. 2. A DPCM encoding device characterized in that, in addition to the second transversal type filter, the sum of the outputs of both the first transversal filter and the second transversal filter is used as the predicted value.