JPS64857B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to an adaptive differential PCM encoding and decoding method used for band compression of audio signals, that is, so-called low bit rate encoding. Regarding conventional adaptive differential PCM encoding/decoding methods, see “Proceedings of
IEEE, pages 488-525, and adaptive differential PCM encoding/decoding with characteristics that are resistant to transmission line bit errors, see "Proceedings of ICASSP'82," published by IEEE May 1982, pages 960-- Details can be found on page 963.The conventional technology will be described below based on the above-mentioned second document.Figure 1 shows the conventional adaptive differential PCM encoding and decoding method. Adaptive difference consisting of a quantizer 3, an inverse quantizer 4, an adder 5, a predictor 6 and a code output terminal 7
PCM encoder, code input terminal 8, dequantizer 9, adder 10, predictor 11 and output terminal 12
shows an adaptive differential PCM decoder consisting of The quantizer 3 is a circuit that obtains an output signal with a length of N bits smaller than M when the input signal is expressed with a length of M bits, and judges the input signal using 2 ^N -1 thresholds. The determination result is output in N bits. In other words, if the quantization width at a certain sample time j is △ _j , the input signal x _j at this time is n _j・△ _j x _j <(n _j +1)・△ _j , n _j ∈{0, ±1, ± 2, ...±(2 ^m-1 -1), -2 ^N-1 }, (1) If N is the allocated number of quantization bits, the output signal is n _j and at the next sampling time (j+1) The quantization width Δj ₊₁ is companded using the following equation according to the quantizer input signal level. △ _j+1 = △〓 _j・M(n _j ) (2) Here, M(n _j ) is a multiplier uniquely determined by n _j , and the audio signal sampled at 8kHz is converted into 4 bits ( Table 1 shows an example of the multipliers used when encoding with m=4).

【table】

[Table] If β is set as a positive constant smaller than 1 in equation (2), as long as the predictor is a time-invariant filter, △
〓 It is known that the calculation of _j has the effect of leaking the past quantization width, making it resistant to bit errors in the transmission path. See page 1365. Inverse quantizers 4 and 9
When the N-bit output signal of the quantizer 3 and the transmitted N-bit quantizer output signal are inputted, x _j outputs an M-bit reproduced input signal corresponding to the threshold value. =n _j △ _j +0.5△ _j (3) The transmission signal is dequantized. The transfer functions of predictors 6 and 11 are the same, and if this is P(Z), then P(Z)= _k 〓 ⁱ⁼¹ a ^j _i Z ^-i (4). Here, {a ^j _i |i=1,...,k} is time j
If the predictor input signal at time j is x _j and the inverse quantizer output signal is e _j , each coefficient is changed so that e _j ² is minimized.
In other words, each coefficient changes from moment to moment as a ^j+1 _i = (1-δ)a ^j _i +g・e^ _j・x^ _ji (5). Here, δ and g are positive constants smaller than 1. A conventional adaptive differential PCM encoding and decoding method will be described below with reference to FIG. When the input signal sample value x _j at time j is input to the adaptive differential PCM encoder from terminal 1, the subtracter 2 calculates the difference between the input signal x _j and the output signal x^ _j of the predictor 6, and the error signal e _j
The signal is input to the quantizer 3 as a signal. As described above, the quantizer 3 converts e _j into an N-bit code n _j , which is output from the terminal 7 and simultaneously input to the inverse quantizer 4. The inverse quantizer 4 reproduces an M-bit error signal e^ _j from _nj . The reproduced error signal e^ _j and the output x〓 _j of the predictor 6 are added by the adder 5 to reproduce the quantized input signal e^ _j . Thereafter, the quantization widths of the quantizer 3 and the inverse quantizer 4 and the coefficients of the predictor 6 are modified in order to encode the next input signal as described above. As mentioned above, the coefficients of the predictor are modified to minimize the power of the error signal e^ _j , that is, e^ ² _j , so the e _j signal has a smaller dynamic range than the x _j signal, Considering that the same bits are used for encoding, the error generated by the quantizer 3 becomes smaller as the bit size becomes smaller, and more accurate encoding can be achieved. On the other hand, in a conventional decoder, the received quantized code n _j is inputted from a terminal 8 and an inverse quantizer 9 generates a reproduced error signal e^ _j . This e^ _j and predictor 11
The outputs x^ _j of are added by the adder 10 to synthesize x^ _j , output to the output terminal 12, and added to the predictor 11 for predicting the next sample time. On the decoder side, the quantization width of the inverse quantizer is changed moment by moment based on the quantization code n _j or the error signal e^ _j , and x^ _j
The coefficients of the predictor 11 are changed so as to minimize the difference between and x〓 _j , that is, the power of e^ _j . In the encoder and decoder, if the internal states of the inverse quantizers 4 and 9 and the predictors 6 and 11 match, e^ _j , x^ _j , x〓 of the encoder/decoder The values of _j match. Therefore, even if the encoder and decoder are provided at a distance, the input signal x _j applied to terminal 1 and x^ _j output from terminal 12 will take almost the same value. By the way, a transmission line is formed between the terminal 7 of the encoder and the terminal 8 of the decoder, and there is a possibility that bit errors may occur in the transmission line due to thermal noise or the like. In this case, the decoder often falls into an unstable state and cannot recover. This can be explained as follows. Output terminal 1 from the output e^ _j of the inverse quantizer 9 of the decoder
When the transfer function D(Z) up to 2 is obtained using equation (4) as the transfer function of the predictor 11, becomes. As mentioned above, a ^j _i is a value calculated from e^ _j , and when a transmission path bit error occurs, the correction value of the prediction coefficient of the predictor of the decoder is different from the prediction coefficient of the predictor of the encoder. The values will be different. Equation (6) has K poles determined by the prediction coefficients, and as a result of the above transmission line bit error, the position of the pole may move outside the unit circle on the Z plane. In such a situation, the decoder enters an oscillation state and cannot return to normal operation again. (Refer to the second document above) In order to eliminate this unstable state, the second document expands equation (6) as shown below to create an adaptive model with a transfer function that excludes the poles that move adaptively. A differential PCM encoder and decoder is realized. Here, the coefficient {a^ _i } is a fixed constant and {b ^j _i } is an adaptive coefficient. The term (1+ _M 〓 ⁱ⁼¹ b ^j _i Z ^-i ) transforms equation (6) into (1
+ _k 〓 ⁱ⁼¹ a _i Z ^-i ) is divided by the (M+1) term. If the fixed coefficient {a^ _i } is selected to a value that suits the average nature of speech, the above-mentioned truncation error will be small, and there will be almost no deterioration in encoding quality. Here, a fixed coefficient {a^ _i } that matches the average nature of the voice
The method for determining is detailed on page 498 of the above-mentioned first document. FIG. 2 shows a conventional adaptive differential PCM encoding and decoding circuit based on equation (7). FIG. 2 shows an encoder consisting of an input terminal 1, subtracters 21 and 22, a quantizer 3, an inverse quantizer 4, adders 51 and 52, an adaptive filter 61, a fixed filter 62, and an output terminal 7,
Input terminal 8, inverse quantizer 9, adders 101, 10
2, adaptive filter 111, fixed filter 112,
It consists of a decoder with an output terminal 12. Fixed filters 62 and 112 have the following transfer functions with the fixed prediction coefficients {a^ _j } used in equation (7). P2(Z)= _k 〓 ⁱ⁼¹ a^ _i Z ^-i (8) Furthermore, the adaptive filters 61 and 111 have the following transmission function. P1(Z)= _M 〓 ⁱ⁼¹ b ^j _i Z ^-i (9) However, each adaptation coefficient is modified as follows,
The second document states that this is modified to minimize the power of the e _j signal. b ^j+1 _i = (1-δ)b ^j _i +ge^ _ji e^ _j (10) Now, when input signal x _j is input from terminal 1,
The subtracter 21 takes the difference from the output x〓 _j of the fixed filter 62 to obtain y _j , which is input to the subtracter 22. The subtracter 22 subtracts the output y〓 _j of the adaptive filter from y _j ,
It is added to quantizer 3. The quantizer 3 quantizes e _j and outputs the code n _j from the output terminal 7, and also adds it to the inverse quantizer 4, which generates a quantized error signal.
get e^ _j . e^ _j is input to the adaptive filter 61 and used for filter calculation at the next sampling time, and the output y〓 _j of the adaptive filter 61 is added by the adder 51 and transmitted as y^ _j to the adder 52. It will be done.
The adder 52 adds y^ _j and _x〓j to reproduce the quantized signal x^ _j of the input signal _xj , which is used for filter operation at the next sampling time. For this reason, the fixed filter 62
If the output of is suitable for the average behavior of the input signal, the amplitude level of the first error signal y _j decreases, and the second error signal e _j is obtained by subtracting the output of the adaptive filter 61 from this signal. becomes an even lower level signal. Generally speaking, the adaptive predictor 6 in FIG. 1 predicts the next input signal value from the reproduced quantized input value, whereas the adaptive filter 61 in FIG.
predicts the next input signal from the error signal, and the adaptive filter 61 shown in FIG. Therefore, the encoder shown in FIG. 2 can perform encoding as good as the encoder shown in FIG. 1 as a whole. Next, the operation of the decoder shown in FIG. 2 will be explained. When the quantization code is input from the input terminal 8, the inverse quantizer 9 reproduces the quantized error signal e^ _j , inputs it to the adaptive filter 111, and uses it for the adaptive filter operation at the next sampling time, and , is added to the output y〓 _j of the adaptive filter 111 by the adder 101 to reproduce y^ _j .
y^ _j is the output x〓 _j of the fixed filter 112 and the adder 102
The encoder side input signal x^ _j which is added and quantized by
, output terminal 12 and fixed filter 112
supplied to The transfer functions P1 (Z) and P2 (Z) of the adaptive filter 111 and fixed filter 112 are expressed by formula (8)
and as shown in equation (9), and the transfer function D(Z) from the output of the inverse quantizer 9 to the output terminal 12 is D(Z)=1+P1(Z)/1−P2(Z) (11) Therefore, it is consistent with Equation (7), and the adaptively moving pole is Z
Since it is not held on a flat surface, stable operation can be expected even if transmission line bit errors occur. However, although this method makes the decoder more resistant to bit errors in the transmission path, the truncation error in equation (7) becomes large when the input signal has characteristics significantly different from the average characteristics of speech. As a result, the encoding efficiency deteriorated. That is, for such an input signal, the level of the first error signal y _j in FIG. 2 was sometimes higher than that of the input signal x _j .
This is clear from the graph of input frequency vs. S/N ratio shown in Figure 6 on page 962 of the second document, and since the average nature of speech is that there are few high frequency components, It can be seen that the S/N ratio has significantly deteriorated. Furthermore, a data modem signal is an important voice band signal, but the characteristics of this signal are significantly different from voice signals, so the fixed prediction coefficients are also significantly different, and the bit error rate is higher than that of the data modem signal. This caused some inconvenience. SUMMARY OF THE INVENTION An object of the present invention is to provide an ADPCM method and a circuit thereof that do not become unstable due to transmission line bit errors and have good encoding characteristics for audio and other signals. In the method of the first invention, the input signal is a signal obtained by digitizing signals obtained from two signal sources each having different statistical properties at each sample time, and on the encoding side, a predicted signal is generated from the input signal. quantizes the residual signal obtained by subtracting the quantized residual signal, transmits the quantization code of the quantized residual signal, and generates a locally decoded signal by adding the quantized residual signal and the predicted signal. , two time-invariant prediction functions
FP1(Z) and FP2(Z) are each determined by the average properties of the two signal sources, and the coefficient α is set from 0 to 1.
When changing between 1/{1−αFP1(Z)
-(1-α)FP2(Z)} is a prediction function that has a pole only within the unit circle on the Z plane, and the two predicted values obtained from the locally decoded signal by the two time-invariant predictions are expressed as coefficient α , (1-α) to obtain the predicted signal, and the FP1
Find the distance between the input and output of (Z) and FP2 (Z), and if the distance obtained by the time-invariant prediction that is multiplied by α is smaller than the other distance, increase α,
If not, decrease α and 0≦α≦1
On the decoding side, the transmitted quantization code is dequantized to obtain a decoding side quantization residual signal, and this quantization residual signal and the decoding side predicted signal are A decoded signal is obtained by adding the signals, and from this decoded signal, the above-mentioned FP1 (Z) and FP2 (Z)
The decoding side prediction signal is generated by weighting the two predicted values obtained by the above based on the same rules as the transmission side. In the method of the first invention, the transfer function P of the predictor is
(Z) is K variables as shown in equation (4) {a ^j _i | i
=1, 2,...k}, but instead use a time-invariant prediction filter FP1(Z), which is determined by the statistical average characteristics of two typical input signals.
Using FP2(Z) P(Z)=αFP1(Z)+(1−α)FP2(Z) (12) used as For example, when dealing with telephone signals etc.
For FP1 (Z), prepare a filter that matches the average audio signal, and for FP2 (Z), prepare a filter that matches the average data modem signal, so that α1 is used for audio input and α0 is used for data signal input. If controlled, both low-pass audio signals and band-pass data signals can be encoded with high fidelity.
ADPCM method can be realized. There are two problems when determining the predictor transfer function as shown in equation (12). First, when using equation (12), the decoder transfer function D(Z) is Therefore, when α changes, the question is whether equation (14) has poles only within the unit circle, and in general, 1/1-FP1(Z), 1/1-FP2(Z) (15) Even if there is a pole inside the unit circle, equation (14) may have a pole outside the unit circle, and if such a pole is outside the unit circle, the decoder will be in an oscillation state due to a transmission line bit error. fall into. Regarding this problem, we set the variation range of α from 0 to 1, find the root locus when α is varied from 0 to 1 in equation (14), and find the root locus that moves outside the unit circle. If so, modify {c _i } or {d _i } so that the root locus is within the unit circle. In this way, even if a transmission line bit error occurs and the internal states of the encoder/decoder become inconsistent and the decoder generates a large error, this error will be absorbed by the decoder whose poles are within the unit circle. Due to the transfer function, it decreases exponentially over time. Furthermore, even if the amount of correction of α differs between the encoder and decoder due to bit errors in the transmission path, in this method, which uses equation (14) in which the poles can move freely on the Z plane, the possible positions of the poles are only on the root locus. Therefore, no significant distortion of the transfer function occurs. The second problem is how to correct α. If the format is as shown in equation (4), the variables are coefficients that are multiplied by the tap output signal, and the tap output signals form an orthogonal basis in the signal space, so each variable can be treated as an independent variable. . For this reason, the formula
As shown in (5), an optimal solution can be obtained even if each variable is operated independently to minimize the energy of the residual signal. However, when using equation (12), there is only one variable, and FP1(Z), which is not an independent basis,
It is determined in relation to both FP2(Z).
Therefore, a more efficient α correction algorithm is required. Regarding this point, the present invention corrects this in the following manner. Time-invariant prediction filter
Since the input signals of FP1 and FP2 are locally decoded signals or decoded signals x^ _j , they are considered to be approximately equal to the input signals to the ADPCM encoder. On the other hand, FP1, FP2
predicts the current input signal x _j from the past K decoded signals by considering each signal to have two different characteristics, and these are designated x ⁽¹⁾ _j and x ⁽²⁾ _j . Therefore, for example, when an input signal compatible with FP1 is input, x ⁽¹⁾ _j gives a good prediction of x _j , and conversely, when an input signal compatible with FP2 is input, x ⁽²⁾ _j is x^ _j
It should give a good prediction. In order to modify α so as to increase the effect of the time-invariant prediction filter that gives a better prediction, if the value of α at time j is α _j , then according to equation (5), α _{j +1} = (1−δ)α _j +gSgn(‖x ⁽²⁾ _j −x^ _j ‖−‖
x ⁽¹⁾ _j −x^ _j ‖) (16) However, δ and g are positive constants smaller than 1, Sgn(・) is a sign function, and ‖・‖ can be the norm. As an _example , if we use ^the absolute value for ^the norm, then _{α j+1 = (}
1−δ) α _j +g {|x ⁽²⁾ _j −x^ _j ||x ⁽¹⁾ _j −x^ _j | then α _j+1 = (
1−δ) α _j +g |x ⁽²⁾ _j −x^ _j |＜|x ⁽¹⁾ _j −x^ _j | When α _j+1 = (1−δ
)α _j −g}(17), which gives the correct result. Also, since the correction method is to increase the coefficient with the smaller norm, Equation (16)
A similar effect can be expected by using the following formula instead of . α _j+1 =1+β _j+1 α _j+1 =1+β _j+1 β _j+1 =(1−δ)β _j +g/2{1−Sgn(‖x ⁽²⁾ _j −x^
_{j ‖} ⁻ _‖ It will be understood that efficient encoding is possible. In the ADPCM method of the second invention, the input signal is
This is a signal obtained by digitizing signals obtained from two signal sources each having different statistical properties at each sample time, and on the encoding side, the first
The residual signal obtained by subtracting the predicted signal and the second predicted signal is quantized, the quantization code of the quantized residual signal is transmitted, and the quantized residual signal and the first predicted signal are transmitted. The locally decoded signal is obtained by adding the signal and the second prediction signal, and the adaptive prediction function AP(Z)
is an adaptive prediction function in which {1+AP(Z)} has no poles and only adaptively moving zero points, and by this adaptive prediction from the quantized residual signal, the first
generate a prediction signal of two time-invariant prediction functions FP1
(Z) and FP2(Z) are each determined by the average properties of the two signal sources, and when the coefficient α is varied between 0 and 1, 1/{1−αFP1(Z)−(1
−α)FP2(Z)} is a prediction function that has a pole only within the unit circle on the Z plane, and the two predicted values obtained from the locally decoded signal by the two time-invariant predictions are expressed by coefficients α, (1 −α) to obtain the second predicted signal, and the FP1
Find the distance between the input and output of (Z) and FP2 (Z), and if the distance obtained by the time-invariant prediction that is multiplied by α is smaller than the other distance, increase α,
If not, decrease α and 0≦α≦1
On the decoding side, the transmitted quantization code is dequantized to obtain a decoding side quantization residual signal, and this quantization residual signal and the first decoding side prediction are A decoded signal is obtained by adding the signal and a second decoding side prediction signal, and the first decoding side prediction signal is obtained from the decoding side residual signal using the same adaptive prediction as the encoding side. and weighting the two predicted values obtained from the decoded signal by the FP1 (Z) and FP2 (Z) based on the same rules as the transmitting side to generate the second decoding side predicted signal. It is characterized by In the method of the second invention, in the conventional method,
Since the decoder transfer function D(Z) is expanded into fixed poles and adaptive zeros as shown in Equation (7), the transmission line bit errors are strengthened, but the nature of the input signal is a time-invariant prediction filter with fixed poles. The disadvantage of causing deterioration when the characteristics are significantly different from the characteristics of It is an expansion and improvement of D(Z), and realizes the following equation corresponding to equation (7). The methods shown in equations (16) and (17) can be applied to correct α _j and b ^j _i , respectively. If this method is applied to telephone signals and {c _i } is selected as the time-invariant filter coefficient for voice, {d _i ) is the time-invariant filter coefficient for data modem, α _j will be approximately 1 when a voice signal is input.
Since α _j is almost 0 when a data modem signal is input, the adaptive zero point only needs to correct the deviation from the statistical average of the voice or data signal, resulting in higher quality encoding. This is promising and can guarantee robustness against transmission path bit errors. In addition, if encoding with the same quality as the conventional method is sufficient, the value of M can be reduced, and ADPCM
The complexity of the method could also be reduced. In the ADPCM method of the third invention, the input signal is
This is a signal obtained by digitizing signals obtained from two signal sources with different statistical properties at each sample time, and on the encoding side, the residual signal obtained by subtracting the predicted signal from the input signal is quantized. ,
When transmitting the quantization code of the quantized residual signal, a first quantization width is determined based on the quantization code transmitted at the previous time and the quantization width used at the previous time. , the second quantization width determined from the time average value of the first quantization width is multiplied by coefficients α and (1−α), respectively, and the sum is set as the quantization width at the current time,
adding the quantized residual signal and the prediction signal to generate a locally decoded signal, and generating two time-invariant prediction functions;
FP1(Z) and FP2(Z) are each determined by the average properties of the two signal sources, and the coefficient α is set from 0 to 1.
When changing between 1/{1−αFP1(Z)
−(1−α)FP2(Z)} is a prediction function that has a pole only within the unit circle on the Z plane, and the two predicted values obtained from the locally decoded signal by the two time-invariant predictions are The predicted signal is obtained by weighting α, (1−α), and the
Find the distance between the input and output of FP1 (Z) and FP2 (Z), and if the distance obtained by the time-invariant prediction that is multiplied by α is smaller than the other distance, increase α, otherwise decreases α, and 0≦
Control is performed so that α≦1, and on the decoding side, the transmitted quantized code is dequantized using the quantization width generated based on the same rules as on the encoding side, and then quantized on the decoding side. Obtain a residual signal, add this quantized residual signal and the decoding side prediction signal to obtain a decoded signal, and from this decoded signal,
It is characterized in that the two predicted values obtained by FP1(Z) and FP2(Z) are weighted based on the same rules as the transmitting side to generate the decoding side predicted signal. The method of the third invention is the method of the first invention with the addition of quantization width control, and is particularly suitable for communication channels where voice and data modem signals are transmitted together. The prediction method is as shown in equations (12), (13), and (16). Assuming that FP1 is a time-invariant prediction filter for voice and FP2 is a time-invariant prediction filter for data modem, FP1 and FP2 are designed to remove redundancy from the input signal, focusing on the predictability of voice and data modem, respectively. However, in addition to this, the present invention attempts to take advantage of this redundancy effect by paying attention to the amplitude distribution of voice and data modems. The amplitude distribution of speech is thought to be Laplace distributed (Prentice-Hall, Inc., 1978, published by
(See Digital Processing of Speech Signals P176) It is not uncommon for signals with amplitudes that are three times the standard deviation value or more to appear. Therefore, in order to perform adaptive quantization, it was necessary to modify the quantization width as shown in equation (2) to cause a significant change in the quantization width as shown in Table 1. On the other hand, the amplitude of a data modem signal does not fluctuate as sharply as a voice signal, and in fact, it is expressed by equation (2).
Correcting the quantization width by using this method will unnecessarily increase or decrease the quantization width for modem signals. For such applications, it is advantageous to smooth the quantization width obtained by equation (2). Therefore, when α _j is close to 1 (when an audio signal is input), Equation (2) is used, and when α _j is close to 0 (when a data modem signal is input), the smoothed quantization of Equation (2) is applied. It is preferable to give width. Therefore, the quantization width ^ΔOPT _{j+1 at time j+1} is given from the smoothed quantization widths ^ΔAV _j+1 and _Δj+1 as follows. △ ^OPT _j+1 = α _j+1 △ _j+1 + (1−α _j+1 ) △ ^AV _j+1 △ _j+1 = (△ ^OPT _j )〓M(n _j ) △ ^AV _j+1 = (1−δ)△ ^AV _j +g△ _j (20) According to the present invention as described above, when an audio signal is input, the audio time-invariant prediction filter FP1 and Equation 2
In addition, when a data modem signal is input, it can be encoded using a time-invariant prediction filter for data FP2 and a smoothed quantization width, so both input signals Higher quality encoding can be expected. Furthermore, since the variable α _j used to control the quantization width can be shared with the switching signal of the predictor, this method is simpler than a method in which only the quantization width is controlled independently, and the effect is considered to be about the same. In the ADPCM method of the fourth invention, the input signal is
This is a signal obtained by digitizing signals obtained from two signal sources each having different statistical properties at each sample time, and on the encoding side, the first
When the residual signal obtained by subtracting the predicted signal and the second predicted signal is quantized and the quantized code of the quantized residual signal is transmitted, the quantized code transmitted at the previous time and the quantized code transmitted at the previous time are α, (1− α) The value obtained by multiplying and adding the coefficients is set as the quantization width at the current time, and the quantized residual signal, the first predicted signal, and the second predicted signal are added to obtain a locally decoded signal, Adaptive prediction function AP(Z)
is an adaptive prediction function in which {1+AP(Z)} has no poles and only adaptively moving zero points, and by this adaptive prediction from the quantized residual signal, the first
generate a prediction signal of two time-invariant prediction functions FP1
(Z) and FP2(Z) are each determined by the average properties of the two signal sources, and when the coefficient α is varied between 0 and 1, 1/{1−αFP1(Z)−(1
−α)FP2(Z)} is a prediction function that has a pole only within the unit circle on the Z plane, and the two predicted values obtained from the locally decoded signal by the two time-invariant predictions are expressed by coefficients α, (1 −α) to obtain the second predicted signal, and the FP1
Find the distance between the input and output of (Z) and FP2 (Z), and if the distance obtained by the time-invariant prediction that is multiplied by α is smaller than the other distance, increase α,
If not, decrease α and 0≦α≦1
On the decoding side, the transmitted quantization code is dequantized using the quantization width generated based on the same rules as on the encoding side, and the decoding side quantization residual signal is obtained. and add this quantized residual signal, a first decoding side prediction signal, and a second decoding side prediction signal to obtain a decoded signal, and
Generate the first decoding side prediction signal from the decoding side residual signal using the same adaptive prediction as the encoding side, and generate the FP1 (Z), FP2 (Z) from the decoded signal.
The second decoding side predicted signal is generated by weighting the two predicted values obtained by the above based on the same rules as the transmitting side. The method of the fourth invention is the method of the second invention with the addition of quantization width control (Equation (20)), and this method is also suitable for communication channels where voice and data modem signal transmissions are mixed. It's a method. In this method, the prediction method is as shown in equation (12), equation (13), and equation (16),
FP1 is a time-invariant prediction filter for voice, FP2 is a time-invariant prediction filter for data modem, and the quantization width follows equation (20). Therefore, when an audio signal is input, the value of α _j becomes close to 1, and the quantization width is calculated using equation (20).
Using a decision method that is advantageous for speech signals with Laplace distribution according to equation (2), and predicting the average characteristics of speech using the time-invariant filter FP1 from equation (16), and predicting the average characteristics and current characteristics using the adaptive filter, Since the correction is performed with the input signal, high-quality encoding can be expected. In addition, when a data modem signal is input, α _j becomes a value close to zero, and the quantization width is selected from Equation (20) as △ ^AV , which is the smoothed quantization width obtained from Equation (2). In addition to being convenient for quantization of small modem signals, FP2 is selected as the time-invariant prediction filter according to equation (20), and the adaptive filter calculates the difference between the average modem signal characteristics and the current input modem signal characteristics. is corrected, enabling high-quality modem signal encoding. Furthermore, since the quantization width control parameter α _j is shared with the time-invariant prediction switching parameter, the complexity of the control unit can be reduced. Next, a circuit for realizing the ADPCM method of the fourth invention will be described. A circuit that realizes the ADPCM method of the fourth invention includes a subtracter that generates a residual signal by subtracting a predicted signal from an input signal input at each sample time, and a circuit that adaptively converts the residual signal from the subtracter into a quantizer that performs quantization with a varying quantization width and generates a quantization code; an inverse quantizer that reproduces a quantized residual signal from the output of the quantizer; and a quantized residual signal of the output of the inverse quantizer. In the predictor of the ADPCM circuit, which consists of an adder that adds the difference signal and the predicted signal to generate a locally decoded signal, and a predictor that generates the predicted signal at the next sampling time from the locally decoded signal of the output of the adder, an adaptive quantum The quantization width determination section is configured to instantaneously adapt the quantization width currently in use to the input signal, multiply the quantization width obtained by a coefficient α, which will be described later, and use this result to apply the instantaneous quantization width to the input signal. The quantization width obtained by adapting the input signal to the quantization width obtained by time smoothing is multiplied by (1-α) and added to the quantization width at the next sampling time. In addition, the configuration of the predictor includes an adaptive filter that adaptively generates a signal to be input at the next sample time from the quantized residual signal, and a predictive signal that is input to the quantized residual signal. In addition, an adder that outputs a locally decoded signal or a decoded signal, a first time-invariant prediction filter that predicts an input signal from the locally decoded signal or decoded signal in accordance with the nature of the input signal that has large level fluctuations, and a second time-invariant prediction filter that predicts an input signal from the locally decoded signal or the decoded signal in accordance with the nature of the input signal that is not intense; a first multiplier that multiplies the output of the second time-invariant prediction filter by a variable α, a second multiplier that multiplies the output of the second time-invariant prediction filter by a variable (1-α), and outputs of the first multiplier and the second multiplier. The variable α, which is commonly used by the quantization width determining unit and the predictor, is controlled by multiplying the value of α by α in the range of 0 to 1. When the distance obtained by the invariant prediction filter is smaller than the other distance, α is increased,
The present invention is characterized by comprising means for controlling α to decrease when the distance obtained by the time-invariant prediction filter multiplied by α is larger than the other distance. The circuit of FIG. 3 directly corresponds to the circuit implementing the ADPCM method of the fourth invention, but if the quantization width control section is removed from the circuit of FIG. Corresponding to the circuit that implements the method, the ADPCM method of the third invention can be achieved by removing the adaptive filter that adaptively estimates the input signal from the quantized residual signal from the circuit part of FIG. Corresponding to the circuit for implementing the ADPCM method of the first invention, it is also easy to correspond to the circuit for implementing the ADPCM method of the first invention by removing the quantization width control section and the adaptive filter section from the circuit part of FIG. be understood. Therefore, below, a circuit implementing the ADPCM method of the fourth invention, which is the most complicated, will be described in detail. This will be explained below with reference to the drawings. FIG. 3 shows input terminal 1, subtracters 21 and 22, quantizer 3, inverse quantizer 4, adders 51 and 52, adaptive filter 61, fixed filters 63 and 64, multipliers 65 and 66, and adder 67. , subtracters 81, 82, norm calculation circuits 83, 84, comparator 85, low-pass filter 86, subtracter 87, multiplier generation circuit 90, multiplier 91, register 92, low-pass filter 93, β-power circuit 94, multiplication It consists of subtractors 95, 96, adder 97, and output terminal 7, and subtracters 21, 22,
Quantizer 3, inverse quantizer 4, adders 51, 52, and adaptive filter 61 are the same as the correspondingly numbered circuits in FIG. The norm calculation circuits 83 and 84 output the absolute value of the input signal to this circuit. The multiplier generating circuit 90 generates the multipliers shown in Table 1 (page 13), and the details will be described later.
The β-th power circuit raises the input signal to the β-th power, and the details of this circuit will be described later. The comparison circuit 85 is a circuit that generates ±g depending on the magnitude of two input signals. Hereinafter, the time-invariant prediction filter 64 is a K-order filter having a prediction coefficient {c _i } suitable for speech, the time-invariant prediction filter 63 is a K-order filter having a prediction coefficient {d _i } suitable for data, and The adaptive filter 61 has a coefficient M whose coefficient at time j is {b ^j _i }
This will be explained as a time-varying filter. The operation at time j when x _j is input to input terminal 1 will be described below. At this time, the outputs of the time-invariant prediction filters 64 and 63 are the past K of the adder 52, respectively.
_k 〓 ⁱ⁼¹ _{c i} x^ _ji , _{k 〓} ⁱ⁼¹ d _i x^ _ji using the output signals is multiplied by the output of the time-invariant filter 64 to obtain α _jk 〓 ⁱ⁼¹ c _i x^ _ji , and in the multiplier 65, the output signal α _j of the low-pass filter 86 is subtracted from 1 by the subtractor 87 to obtain the value ( 1−α _j ) and time-invariant filter 63
By multiplying the output signals of (1−α _j ) _k 〓 ⁱ⁼¹ d _i x^ _ji is obtained. Therefore, the output signal of the adder 67 is α _jk 〓 ⁱ⁼¹ c _i x^ _ji + (1−α _j ) _k 〓 ⁱ⁼¹ d _i x^ _ji , and as described later, α _j is the audio signal Since it is close to 1 when inputting and close to 0 when inputting a data modem signal, the subtracter 21 uses a time-invariant prediction suitable for the average characteristics of the voice/data modem signal depending on whether the input signal x _j is a voice signal or a data modem signal. Subtract the signal from x _j . Therefore, regardless of whether the output signal is an audio signal/data modem signal, the output signal of the subtracter 21 still contains a high-level signal because the input signal fluctuates from the average characteristic over time, and from this signal, the adaptive filter 6
1, the quantized residual signal e^ _j
The signal estimated as _M 〓 ⁱ⁼¹ b ^j _i e^ _ji is subtracted by the subtractor 22.
A subtracted residual signal e _j is generated. This residual signal
e^ _j is encoded by the quantizer 3 using the quantization width Δ ^OPT _j which is the output of the adder 97, and is output from the output terminal 7. Furthermore, the encoded quantized residual signal n _j is used by the inverse quantizer 4 to generate a quantized residual signal e^ _j using the quantization width ^ΔOPT _j that is the output of the adder 97. This quantized residual signal e^ _j is input to the adaptive filter 61 and used to obtain the estimated value at the next sampling time, and the adder 51 adds the current estimated value, and the adder 52 adds the estimated value at the next sampling time. Adder 6
7 to generate a locally decoded signal x^ _j , which is applied to time-invariant prediction filters 63 and 64 and used to obtain the predicted value at the next sample time. Furthermore, the difference between the local decoded signal x^ _j and the output signals x ⁽¹⁾ _j and x ⁽²⁾ _j of the time-invariant prediction filters 64 and 63 at the current time is taken by subtracters 82 and 81, respectively, and x ⁽¹⁾ _j −x^ _j , x ⁽²⁾ _j −x^ _j
The absolute values of these signals are taken by the norm circuits 84 and 85 and become |x ⁽¹⁾ _j −x^ _j |, |x ⁽²⁾ _j −x^ _j |, respectively, which are input to the comparator 85. Ru. In the comparator 85, −1 when |x ⁽¹⁾ _j −x^ _j ||x ^{( 2)} _{j −x} ^ _{j |} |<|x ⁽²⁾ _j −x To generate +1 when ^ _j |, mathematically we use the sign function Sgn to generate Sgn(|x ⁽²⁾ _j −x^ _j |−|x ⁽¹⁾ _j −x^ _j |) You can think of it as a thing. This output is passed to a low pass filter 86. If the low-pass filter is a first-order filter, details will be described later, but the operation of the low-pass filter is to multiply the input signal by the gain g, and to combine this signal with
This can be achieved by multiplying the low-pass filter output signal one sample time before by a coefficient (1-δ) smaller than 1 and adding it.
The output signal α _j+1 of is α _j+1 = (1−δ)α _j +gSgn (|x ⁽²⁾ _j −x^ _j |−|x ⁽¹⁾ _j −x^ _j |), and the formula ( 16). Therefore, if this circuit is used, when inputting an audio signal, α _j = a value close to 1, and when inputting a data modem signal, α _j = a value close to 0, as detailed in the ADPCM method of the first invention. I hope you understand. Further, the quantization width △ ^OPT _j which is the output of the adder 97 used in the quantizer 3 and the inverse quantizer 4 is changed to (△ ^OTP _j )〓 by the β power circuit 94, and the multiplier 91 uses this value and the multiplier Multiplier M obtained by generation circuit 90
(n _j ) is multiplied to become ( ^ΔOPT _j )〓M(n _j ), which is stored in the register 92 as Δ _j+1 at the next sample time. At the current time, the register 92 contains △
Since _j is stored, in the low-pass filter 93,
Similar to the low-pass filter 86, the current output value ^ΔAV _j is used to calculate the output value at the next sample time as ^ΔAV _j+1 =(1−δ) ^ΔAV _j + _gΔj . Therefore, at sample time j+ ₁ , the multiplier 96
The value α _{j +1} △ _j+1 obtained by multiplying α j+1 by ^α _{j+1 and} the value ( 1−α _j+1 )△ ^AV _j+1 are added together in an adder 97, and the output is △ ^OPT _j+1 = α _j+1 △ _j+1 + (1−α _j+1 )△ ^AV _j+ Get ₁ . This agrees with equation (20), and the time-invariant prediction filter 64 is used for voice and the time-invariant prediction filter 63 is used for voice.
is selected for the data modem, the data modem signal can be encoded with high quality as detailed in the ADPCM method of the third invention. Further, an ADPCM decoding circuit according to the present invention is shown in FIG. FIG. 4 shows an ADPCM code input terminal 8, an inverse quantizer 9, an adaptive filter 111, and adders 101 and 10.
2, output terminal 12, time-invariant prediction filter 113,
114, subtractor 122, 121, norm calculation circuit 123, 124, comparator 125, low-pass filter 1
26, subtracter 127, multipliers 115, 116, adder 117, multiplier generation circuit 130, multiplier 13
1, a register 132, a low-pass filter 133, multipliers 136, 137, an adder 138, and a β-power circuit 134. Here, inverse quantizer 9, multiplier circuit 130, multiplier 131, register 1
32, low-pass filter 133, multiplier 136, 13
7. The adder 138 and the β power circuit 134 are the inverse quantizer 4, multiplier circuit 90, multiplier 91, register 92, low-pass filter 93, multipliers 96, 95,
Corresponding to the adder 97 and the β-power circuit 94, respectively, and
Adaptive filter 111 and adder 101,1 in FIG.
02, time-invariant prediction filters 113, 114, subtracters 121, 122, norm circuits 124, 123,
Comparison circuit 125, low-pass filter 126, subtracter 1
27, multipliers 116, 115, and adder 117 are the adaptive filter 61 shown in FIG.
1, norm circuits 84, 83, comparison circuit 85, low-pass filter 86, subtractor 87, multiplier 66, 65,
Corresponds to adder 67. Now, when ADPCM code n _j is input to terminal 8,
The inverse quantizer 9 is the output of the adder 138 △ ^OPT _j
is used to generate a quantized residual signal e^ _j , the output of the adaptive film 111 is added to this signal in the adder 101, and the predicted value by the time-invariant prediction filter, which is the output of the adder 117, is added to the signal in the adder 102. In addition, a decoded signal x^ _j is obtained and outputted to the output terminal 12. The outputs of the time-invariant prediction filters 113 and 114 are the outputs of the low-pass filter 126, α _j , and the outputs of the subtracter 127, (1−α _j ), respectively.
7 and added in the adder 117, the predicted value is obtained by α _jk 〓 ⁱ⁼¹ c _i x^ _ji + (1−α _j ) _k 〓 ⁱ⁼¹ d _i x^ _ji . Therefore, the signal x^ _j obtained at the output terminal 12 is the estimated value _M 〓 ⁱ⁼¹ b ^j _i
By adding e^ _ji and e^ _j , we get the following. x^ _j =e^ _j + _M 〓 ⁱ⁼¹ b ^j _i e^ _ji +α _jk 〓 ⁱ⁼¹ c _i x _ji + (1−α _j ) _k 〓 ⁱ⁼¹ d _i x^ _ji From this e^ When the transfer function from _j to x _j is found by the Z transformation in the above formula, (1−α _jk 〓 ⁱ⁼¹ c _i Z ^-i −(1−α _j ) _k 〓 ⁱ⁼¹ d _i Z ^-i × X(Z)=(1+ _M 〓 ⁱ⁼¹ b ^j _i Z ^-i )E(Z) Yotsute , which matches Equation (19). Therefore, it will be understood that this prediction method is resistant to transmission path bit errors and enables high-quality encoding. α _j is controlled by subtracters 122, 121, norm circuits 124, 12
3. This is done in the comparator circuit 125 and low-pass filter 126, but this part is performed by the subtracter 8 in FIG.
2, 81, norm circuit 84, 83, comparison circuit 8
5. Since it matches the part of the low-pass filter 86, the operation that matches equation (16) as explained in the coding circuit part is performed, and high-quality coding is achieved for both voice and data signals. becomes possible. On the other hand, a multiplier circuit 130 and a multiplication circuit 13 are circuits related to correction of the quantization width △ ^OPT _j used in the inverse quantizer 9.
1, register 132, low-pass filter 133, multipliers 136, 137, adder 138, β-power circuit 13
4 is a multiplier circuit 90 and a multiplier 91 in FIG.
register 92, low-pass filter 93, multiplier 96,
95, the adder 97, and the β-power circuit 94, and as explained in the encoding circuit section, the modification method consistent with equation (20) is performed, so the data modem signal can be encoded with high quality. It will look like this. Next, a method of configuring the multiplier generating circuits 90 and 130 will be described. This circuit is a circuit that uniquely outputs the value M(n _j ) shown in Table 1 when n _j is given.
Therefore, it can be configured with a circuit as shown in FIG. Fifth
The figure is composed of an input terminal 200, an absolute value circuit 201, a read-only memory 202, and an output _{terminal 203} . Store the value of . The absolute value circuit simply omits the polarity bit when the ADPCM code is a polarity absolute value display. Therefore, when n _j is input to the terminal 200, it becomes |n _j | in the absolute value circuit 201, and when the read-only memory 202 is accessed with |n _j |, M(n _j ) is input to the terminal 200.
It can be output from 3. In addition, the β power circuit 94 and the
34 circuits can also be realized with read-only memory. Assuming that the quantization width is represented by 9 bits, it is sufficient to prepare a read-only memory with a capacity of 512 words and store n〓 in the word at address n (0n<512). The configuration of the low-pass filter is as shown in FIG. 6 when the filter order is set to 1st order. Figure 6 shows input terminal 3
00, multiplier 301, adder 302, register 3
03, a multiplier 304, and an output terminal 305.The reason why such a circuit becomes a low-pass filter is explained in Theory published by Prentice Hall, Inc., 1975.
and Applications of Digital Signal
Familiar with Processing. Now, if the coefficient of the multiplier 301 is g, the coefficient of the multiplier 304 is (1-δ), the input at time j is A _j , and the output is B _j , then at the next sample time (j+1), B _j is stored in the register 303 and multiplied by (1-δ) by the multiplier 304. To this result, the input signal A _j+i is multiplied by g in the multiplier 301, and the added value is added to the output terminal 305. to
Since it is output as B _j+1 , it can be expressed by the arithmetic expression B _j+1 = (1-δ)B _j +gA _j+1 . Therefore, this format has been used to explain the smoothing of the quantization width and the smoothing of α _j in the ADPCM circuit. As we have seen above, if the ADPCM circuit of the present invention is used, it can handle signals with widely different characteristics, such as voice signals and data modem signals, while maintaining high resistance to transmission line bit errors. High-quality encoding is possible. For convenience of explanation, the characteristics of the smoothing low-pass filter for quantization and the low-pass filter for smoothing α _j were both set with the same parameters g and (1-δ), so the same characteristics were set. However, in the present invention, the characteristics of both filters are changed.Furthermore, all low-pass filters use first-order cyclic filters, but high-order cyclic filters also have high-order cyclic filters. The following acyclic filter is also included in the present invention because it does not change the essence of the present invention. Further, although the norm circuits 83, 84, 123, and 124 all use an absolute value norm, they may also use a norm other than the absolute value, such as a square norm or a square smooth norm. It is part of the invention.

[Brief explanation of drawings]

Figure 1 is a diagram showing a conventional basic ADPCM circuit, Figure 2 is a diagram of a conventional ADPCM that is a conventional basic ADPCM circuit that is resistant to transmission line bit errors, and Figure 3
Figure 4 shows the ADPCM encoding circuit of the present invention.
Figure 5 shows the ADPCM decoding circuit of the present invention.
The figure shows the circuit used in Figures 3 and 4,
FIG. 6 is a diagram showing the circuit used in FIGS. 3 and 4. In Figure 3, 21, 22, 81, 82, 8
7...Subtractor, 3...Quantizer, 4...Dequantizer, 5
1, 52, 67, 97... Adder, 61... Adaptive filter, 63, 64... Time-invariant prediction filter, 83,
84...Norm circuit, 85...Comparator, 86, 93...
Low-pass filter, 65, 66, 95, 96, 91...
Multiplier, 90... Multiplier circuit, 92... Register, 94
...It is a β-th power circuit. In Figure 4, 122,1
21, 127...Subtractor, 9...Dequantizer, 10
1, 102, 138, 117...Adder, 111...
Adaptive filter, 113, 114... Time-invariant prediction filter, 124, 125... Norm circuit, 123... Comparator, 126, 133... Low-pass filter, 116,
115, 136, 137, 131...multiplier, 13
0...multiplier circuit, 132...register, 134...β power circuit.

Claims (1)

[Claims] 1. The input signal is a signal obtained by digitizing signals obtained from two signal sources each having different statistical properties at each sample time, and on the encoding side, a predicted signal is obtained from the input signal. quantizing the residual signal obtained by subtracting, transmitting a quantization code of the quantized residual signal, and adding the quantized residual signal and the prediction signal to generate a locally decoded signal, Two time-invariant prediction functions FP1(Z), FP2(Z)
are determined by the average properties of the two signal sources, and when the coefficient α is varied between 0 and 1, 1/{1-αFP1(Z)-(1-α)FP2(Z)}
is a prediction function that has a pole only within the unit circle on the Z plane, and the two predicted values obtained from the locally decoded signal by the two time-invariant predictions are expressed as coefficients α, (1−
The predicted signal is obtained by weighting with α), and the distance between the input and output of FP1 (Z) and FP2 (Z) is determined respectively, and the distance obtained by the time-invariant prediction that is multiplied by α is calculated. is smaller than the other distance, α is increased, otherwise α is decreased, and controlled so that 0≦α≦1. On the decoding side, the transmitted quantization code is reversed. Quantize to obtain a decoding side quantized residual signal, add this quantized residual signal and a decoding side prediction signal to obtain a decoded signal, and from this decoded signal,
A time-invariant prediction combination type ADPCM characterized in that the two predicted values obtained by FP1(Z) and FP2(Z) are weighted based on the same rules as the transmitting side to generate the decoding side predicted signal. Encoding/decoding method. 2. The input signal is a signal obtained by digitizing signals obtained from two signal sources with different statistical properties at each sample time, and on the encoding side, the first predicted signal and the second predicted signal are extracted from the input signal. A residual signal obtained by subtracting the predicted signal is quantized, a quantization code of the quantized residual signal is transmitted, and the quantized residual signal, the first predicted signal, and the second predicted signal are quantized.
The adaptive prediction function AP(Z) is an adaptive prediction function in which {1+AP(Z)} has no poles and only adaptively moving zero points. The first prediction signal is generated from the converted residual signal by this adaptive prediction, and the two time-invariant prediction functions FP1 (Z) and FP2 (Z) are
determined by the average properties of two signal sources, and when the coefficient α is varied between 0 and 1,
−αFP1(Z)−(1−α)FP2(Z)} is a prediction function that has a pole only within the unit circle on the Z plane, and the two time-invariant predictions obtained from the locally decoded signal by the two time-invariant predictions are The second predicted signal is obtained by weighting the predicted value with coefficients α and (1-α), and the distances between the input and output of FP1 (Z) and FP2 (Z) are determined, and α is If the distance obtained by the time-invariant prediction of the one to be multiplied is smaller than the other distance, increase α, otherwise decrease α,
and 0≦α≦1. On the decoding side, the transmitted quantization code is dequantized to obtain a decoding side quantization residual signal, and this quantization residual signal and the A decoded signal is obtained by adding the first decoding side prediction signal and the second decoding side prediction signal, and the first decoding side prediction signal is added using the same adaptive prediction as the encoding side from the decoding side residual signal. A decoding side prediction signal is generated, and the FP1 (Z) and FP2 are calculated from the decoded signal.
A time-invariant predictive combination type ADPCM encoding decoding characterized in that the two predicted values obtained by (Z) are weighted based on the same rules as the transmitting side to generate the second decoding side predicted signal. method. 3. The input signal is a signal obtained by digitizing signals obtained from two signal sources with different statistical properties at each sample time, and on the encoding side, the residual signal obtained by subtracting the predicted signal from the input signal is When quantizing the difference signal and transmitting the quantization code of the quantized residual signal, the quantization code is determined based on the quantization code transmitted at the previous time and the quantization width used at the previous time. The quantization width at the current time is the value obtained by multiplying the quantization width of 1 and the second quantization width determined by the time average value of the first quantization width by coefficients α and (1-α), respectively. and adding the quantized residual signal and the prediction signal to generate a locally decoded signal, 2
The two time-invariant prediction functions FP1(Z) and FP2(Z) are each determined by the average properties of the two signal sources, and when the coefficient α is varied between 0 and 1,
1/{1-αFP1(Z)-(1-α)FP2(Z)} is Z
It is a prediction function that has a pole only within a unit circle on a plane, and the two predicted values obtained from the locally decoded signal by the two time-invariant predictions are calculated by the coefficients α, (1
- α) to obtain the predicted signal, and also calculate the distances between the input and output of FP1 (Z) and FP2 (Z), respectively, and obtain the time-invariant prediction of the one multiplied by α. When the distance is smaller than the other distance, α is increased, otherwise α is decreased, and the control is performed so that 0≦α≦1. On the decoding side, the transmitted quantization code is , obtain a decoding side quantized residual signal by inverse quantization using a quantization width generated based on the same rules as the encoding side, and add this quantized residual signal and the decoding side predicted signal. A decoded signal is obtained, and the two predicted values obtained from the decoded signal by FP1 (Z) and FP2 (Z) are weighted based on the same rules as the transmitting side to generate the decoding side predicted signal. A time-invariant predictive combination type characterized by
ADPCM encoding/decoding method. 4. The input signal is a signal obtained by digitizing signals obtained from two signal sources each having different statistical properties at each sample time, and on the encoding side, the first predicted signal and the second predicted signal are extracted from the input signal. When quantizing the residual signal obtained by subtracting the predicted signal and transmitting the quantization code of the quantized residual signal,
A first quantization width determined based on the quantization code transmitted at the previous time and the quantization width used at the previous time, and a second quantization determined from the time average value of the first quantization width. The value obtained by multiplying the quantization width by coefficients α and (1−α), respectively, is set as the quantization width at the current time, and the quantized residual signal, the first predicted signal, and the second predicted signal are The adaptive prediction function AP(Z) is an adaptive prediction function in which {1+AP(Z)} has no poles and only adaptively moving zero points, and the above-mentioned quantized residual The first prediction signal is generated from the difference signal by this adaptive prediction, and the two time-invariant prediction functions FP1 (Z) and FP2 (Z) are
determined by the average properties of two signal sources, and when the coefficient α is varied between 0 and 1,
−αFP1(Z)−(1−α)FP2(Z)} is a prediction function that has a pole only within the unit circle on the Z plane, and the two time-invariant predictions obtained from the locally decoded signal by the two time-invariant predictions are The second predicted signal is obtained by weighting the predicted value with coefficients α and (1-α), and the distances between the input and output of FP1 (Z) and FP2 (Z) are determined, and α is If the distance obtained by the time-invariant prediction of the one to be multiplied is smaller than the other distance, increase α, otherwise decrease α,
and 0≦α≦1, and on the decoding side, the transmitted quantization code is dequantized using the quantization width generated based on the same rules as on the encoding side and decoded. Obtain the side quantized residual signal,
A decoded signal is obtained by adding this quantized residual signal, a first decoding side prediction signal, and a second decoding side prediction signal, and the same adaptation as the encoding side is obtained from the decoding side residual signal. generating the first decoding side prediction signal using prediction; and generating the FP1 from the decoded signal.
A time-invariant prediction combination characterized in that the second predicted signal on the decoding side is generated by weighting the two predicted values obtained by (Z) and FP2(Z) based on the same rules as on the transmitting side. Type ADPCM encoding/decoding method.