JPS5941596B2 - Adaptive linear prediction device - Google Patents

Abstract

PURPOSE:To enable the analysis of the synchronism of the sound source, by increasing the degree of band compression, through performing the control of the amount of attenuation for the expected coefficient, in analysis or the like for audio signals.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an adaptive linear prediction device that is used for audio signal analysis, band compression, etc., and that sequentially and adaptively obtains prediction coefficients.

In recent years, linear prediction methods have been attracting attention as an effective means for speech analysis and band compression.

This is due to the fact that the audio signal can be predicted to some extent using past signals, that is, it has redundancy. In this case, the determined prediction coefficient is a parameter representing the characteristics of the audio signal, and since the original signal can be reproduced based on the prediction coefficient and the prediction error, band compression is also possible. For example, in DPCM or ΔM, the current signal is estimated based on the signal of one sample past, and only the difference component (predicted signal difference) is transmitted. At this time, the prediction coefficient is a fixed value. You can imagine that if you use multiple sample values instead of just one sample in the past and select the optimal value for the prediction coefficient based on audio, you will be able to make more accurate predictions, and the prediction error that needs to be transmitted will be smaller. It's not difficult. Regarding linear prediction, for example, J. Maldloul's explanatory paper ('“Linear Prediction■ATut0
ria1Review', Proceedingso
ftheIEEE, Vol. 63, A6.4, pp, 5
61-580, April 1975) or other papers.

According to these, the audio signal sequence is X.

, x, . . . Xj-0, the audio signal at time j can be approximated by minutes ΣaiXj-1 (1) 1=1.

However, a (i=1, 2,...N) is a prediction coefficient.
Here, one way to find the optimal prediction coefficients that minimizes the square mean of the difference between the predicted value (and the true value father j) is to calculate the autocorrelation coefficients ψo, φ, . . . of the audio signal x. ...φN
, and solve the linear equation.

However, this method requires an extremely large amount of complicated calculations to obtain the autocorrelation and solve the equations, which poses a major difficulty when actually building the device. Another method for determining prediction coefficients is to sequentially modify the prediction coefficients.

In other words, if the prediction error at time j is e, the prediction coefficient A is modified as follows.

Here γ is the correction gain, F. g is any non-decreasing function. By repeating equation (4), al approaches the optimal value. An example of voice band compression using the linear prediction method is the method of solving equation (2) by Atal et al.
．． R. SehrOederl“'AdaptivePr
editiveCOdingOfSpeechSig
nals. ．． ``BellSyst.Tech.J.,V
Ol, 48, Pp. l973-1986, 0ct. 19
70) or by the sequential method as shown in equation (4) (J
,D. GibsOn. S. K. JOnesandJ. I
Melsal″SequentiallyAdapti
vePredictiOnandCOdingOfSp
eachSigrrals, “IEEETrans.C
OmmllniCatlOnS. Vol. COM-22
, 應11, Pp. l789-1793, NOv. l97
4).

In either case, the number of prediction coefficients is 10 or less. The former has a large band compression ratio, but has the disadvantage that the device becomes extremely complex. Although the latter has a smaller compression ratio than the former, it is considerably simpler in terms of circuitry because there is no need to solve equations such as equation (2), and there is no need to transmit prediction coefficients. The present invention relates to the latter. Speech signals are composed of vowels and consonants, and vowels originate from changes in airflow due to periodic vibrations of the vocal cords.

Because of this periodic sound source, the speech prediction error shown in FIG. 1X also becomes a periodic signal as shown in FIG. 1F. In other words, parts that are difficult to predict appear periodically. The limit of audio band compression by linear prediction, that is, the power ratio (compression ratio) between the original audio signal and the prediction error signal when performing optimal linear prediction is 14d.
It will be about B. To further increase the compression ratio, there is a method of increasing the number of prediction coefficients, or the above-mentioned At method that measures the period of the sound source in advance and removes the periodicity of the sound source as shown in Figure 1e.
Possible methods include al. However, when increasing the number of prediction coefficients using the sequential method, generally speaking, as the number of prediction coefficients increases, the ability of the prediction coefficients to follow changes in the audio deteriorates, and in order to prevent deterioration of the S/N ratio, the accuracy of calculation must be increased. There are many problems, such as that the prediction coefficient is not unique, and that the solution to the prediction coefficient is not unique. On the other hand, the method of Atal et al. requires a correlation coefficient to be determined in order to determine the periodicity, and the detected period must be transmitted, resulting in an extremely complex circuit.

Furthermore, in speech analysis, complex calculations such as obtaining autocorrelation and solving equations are required in order to obtain the period of the sound source. In order to solve the above problems, an adaptive cascade prediction method has been proposed in which a commonly used linear predictor and an adaptive linear predictor having a large number of prediction coefficients are connected in cascade. (Japanese Patent Application No. 51-99116) In this prediction method, the short-time spectrum of the audio signal is estimated by a normally used adaptive linear predictor with a relatively small number of prediction coefficients, and another adaptive linear predictor with a large number of prediction coefficients is used. This method uses a linear predictor to estimate the periodicity of a sound source and remove redundancy. Next, a conventional adaptive cascade prediction method will be explained. FIG. 1 is a waveform diagram showing an audio signal and a prediction error, and FIG. 2 is a block diagram showing an embodiment of Jumei's adaptive linear prediction device. In this conventional embodiment, a quantizer and a two-stage linear predictor are cascaded to form a loop on the transmitting side, and the calculated prediction coefficients are automatically attenuated in consideration of transmission errors. It is configured. That is, on the transmitting side, at time j, the original audio signal sequence Xj is applied from the input terminal 10, and the first subtraction circuit 60 on the transmitting side calculates the difference from the predicted value, that is, the Δfirst prediction error Fj-Xj-Xj.

The predicted value father j is generated in the first predictor 20 by the following equation based on the past input audio signal in the first predictor 30. However, X・−・ is the audio signal SJ− reproduced on the transmitting side
1 sample, and N has a value of about 10.

hze is successively modified as shown in the following equation. However, k is a value sufficiently smaller than 1 and represents the degree of attenuation of the prediction coefficient h.
The closer to zero, the smaller the amount of attenuation.

ψ1 and ψ2 are arbitrary non-decreasing functions. Fj is the first prediction error reproduced on the transmitting side. First prediction error f・
is input to the second subtractor 61 and the J second prediction error Ej-
Fj-η is found.

The second predicted value ♀j is obtained by the second predictor 30 using the following equation. Here, Fj-1 is the first
is the prediction error.

or is modified according to the following formula: However, l is a value sufficiently smaller than 1. Further, φ3 and φ4 are arbitrary non-decreasing functions. The second prediction error e· is quantized by a quantizer 70 and sent out as a transmission signal Jej'.

The transmission signal Ej' is inputted to an inverse quantizer 71 having characteristics inverse to that of the quantizer RO, and becomes signal j. Fj-j+j is obtained in the first adder 32 and becomes the new input signal of the second predictor 30. Also, the first adder 22 adds
The father j is obtained and becomes the new input signal of the first predictor 20.

On the other hand, on the receiving side, the received signal e' is converted by the inverse quantization device 72 into 嘗・, and in the first adder J, the sum j'' with the output of the first predictor 40 on the receiving side is j''1 j+η・
is obtained.

Fj'' is obtained by the first predictor 40 on the receiving side according to the following equation.Furthermore, the second adder 51 obtains the estimated value X1'ichiryo''jubunr of the original audio signal. In the predictor 50, it is determined by the following equation.The correction of d? and H1 is performed by the same algorithm as the correction of p?, 1.11, and h! on the transmitting side.

As is clear from the above explanation, all the linear predictors used in conventional embodiments only perform a product-sum operation and correction of prediction coefficients to obtain a predicted value, and can be configured relatively easily in terms of circuitry and have a large Bandwidth compression rate can be expected.

Furthermore, even if a malfunction occurs in transmission or calculation, the resulting error in the prediction coefficients will naturally disappear.
However, a predictor with a large number of prediction coefficients has the disadvantage of being noisy.

Generally speaking, the output noise of a transversal filter as shown in equation (7) is approximately equal to the number of taps (M-L).
is proportional to. Therefore, if noise is included in the predicted signal (j), the prediction error (e) will not become smaller than a certain level, so the effect of band compression cannot be expected beyond a certain level.

SUMMARY OF THE INVENTION An object of the present invention is to provide an adaptive prediction device that has an adaptive predictor consisting of a large number of prediction coefficients and is capable of obtaining a higher audio signal compression rate. The goal lies in the realization of Another object of the present invention is to realize a speech analysis device that can accurately determine even the period of a sound source. According to the present invention, there is a first adaptive linear predictor that sequentially corrects prediction coefficients using prediction errors; In an adaptive linear prediction device that is cascade-connected with a second adaptive predictor that sequentially corrects prediction coefficients of
The adaptive linear predictor is characterized by having means for attenuating prediction coefficients other than those in the vicinity of the prediction coefficient having the maximum value among the plurality of prediction coefficients of the second adaptive linear predictor at a faster rate than other prediction coefficients. A prediction device is obtained. As shown in the above explanation, an adaptive predictor with a large number of prediction coefficients can even estimate the periodicity of the sound source, but increasing the number of coefficients also increases noise and weakens the effect of band compression. It has its drawbacks.

One way to suppress the increase in noise is to make the prediction coefficients of the predictor impulsive. In other words, only one or a numerical value out of a large number of submeasurement coefficients is set to a large value, and the remaining coefficients are all set to zero or values close to zero. In this way, the validity of the impulse ratio of the prediction coefficient can be derived from the fact that it reduces noise and from the periodicity of the speech source. In fact, in the Atal et al. partition mentioned above, only one prediction coefficient has a non-zero value. In order to obtain such ideal prediction coefficients, the present invention uses a method of increasing the attenuation effect of coefficients other than those in the vicinity of the prediction coefficient having the maximum value among the prediction coefficients.

In other words, the attenuation amount 1 in equation (8), which is adopted as an error countermeasure, is controlled as described below. At the start of the adaptive operation, all the prediction coefficients Pi (M≦1≦L) have values close to zero, and at that time, the attenuation amount 1 has a relatively small constant value. The adaptive operation continues and prediction coefficients are generated, and if it exceeds a certain predetermined value θ, for example 0.5, the one with the largest value is selected from among the prediction coefficients that exceed θ, and the attenuation amount is determined for the prediction coefficients in the vicinity. Set 1 to a relatively small constant value, and set the attenuation amount 1 of other prediction coefficients to larger values (8)
The correction shown in the formula is performed to strengthen the tendency of the prediction coefficient to approach zero. Next, the present invention will be explained in detail with reference to the drawings.

FIG. 3 shows an embodiment according to the present invention, mainly showing a second predictor 30 having a large number of prediction coefficients. The output η of the inverse quantizer 71 is sent to the second adder 32 and the coefficient correction unit 3.
04 is input. Further, the output Fj- of the second adder 32
1 (1-1 to N) are stored in the signal storage unit 301 configured using a shift register by the switch 302, and the 1
The signals appear sequentially at the output terminal of the signal storage unit 301 during the sampling period. The prediction coefficients pl (1-1 to N) are stored in the coefficient storage unit 303, and the prediction coefficients pl (1-1 to N) are
The coefficients are sequentially read out, modified, and input again.
The outputs of the signal storage section 301 and the coefficient storage section 303 are applied to the multiplier 3.
05 and the integrator 306 to execute equation (7).
The output η-1 (1-1 to N) of the signal storage unit 301, the prediction error j, the prediction coefficient, ! (1-1 to N) is the coefficient 304
is applied to 1j+1 and formula (8) is executed to obtain a new Pi(
1-1 to N) are output.

The coefficient correction unit 304 corrects the prediction coefficient based on equation (8), detects the maximum value of the prediction coefficient, and controls the attenuation amount 1.
Baby. ．． Coefficient Pi(g)- created by the coefficient correction unit 304
1 to N) that exceeds a certain constant value θ (for example, 0.5), the largest prediction coefficient among them is found and its position is stored. When modifying a coefficient near the prediction coefficient having the maximum value, the attenuation amount 1 is set to a small value, and when modifying other prediction coefficients, the attenuation amount 1 is set to a larger value. However, in order to create a new prediction coefficient when the period of the sound source changes, the attenuation amount 1 must not be made too large. As mentioned above, by controlling the attenuation amount 1, the second
The prediction coefficients generated in the predictor 30 become impulse-like, reducing noise at the predictor output and improving prediction accuracy.

Further, in the present invention, since only the attenuation amount of the prediction coefficient is controlled, it is possible to sufficiently follow fluctuations in the period of the sound source. In this embodiment, a shift register was used as the signal storage section 301 and the coefficient storage section 303, but a random access memory (RAM)
may also be used.

In that case, switch 302 is not necessary. The order of the first predictor 20 and the second predictor 30 in this embodiment is adopted for convenience of explanation, and better characteristics can often be obtained by connecting them in the opposite direction. Furthermore, voice analysis can be performed by using only the transmitter of this embodiment. That is, the first predictor 20 in this embodiment represents information regarding the short-time spectrum of the sound, and the second predictor 30 represents the period of the sound source. As described above, according to the present invention, it is possible to obtain a predictive encoding device that achieves an extremely high degree of band compression, and also to obtain a speech analysis device that can analyze even the period of the sound source.

[Brief Description of the Drawings] Fig. 1 is a waveform diagram showing an audio signal and a prediction error, Fig. 2 is an embodiment of a conventional adaptive linear prediction device, and Fig. 3 is an embodiment according to the present invention. The linear predictor 30 is mainly shown. In the figure, 10 is a transmitting side input terminal, 11 is a transmitting side output terminal, 12 is a receiving side input terminal, 13 is a receiving side output terminal,
14 is a communication channel; in the transmission section, 20 is a first predictor, 22 is a first adder, 30 is a second predictor, 32 is a second adder, 60 is a first subtracter, 61 is a second subtracter, 70 is a quantizer, 71 is an inverse quantizer, and in the receiving section, 40 is a first predictor, 41 is a first adder, 50 is a second predictor, 51 is a second adder, 72
is an inverse quantizer, 301 is a signal storage section, 302 is a switch, 303 is a coefficient storage section, 304 is a coefficient correction section, 3
0 is a multiplier, and 306 is an integrator.

Claims (1)

[Claims] 1 A first adaptive linear predictor that sequentially corrects prediction coefficients using prediction errors and a tenth adaptive linear predictor that has a larger number of prediction coefficients and uses its own prediction word difference to correct the plurality of prediction coefficients. In an adaptive linear prediction device that is cascade-connected with a second adaptive linear predictor that sequentially corrects the An adaptive linear prediction device comprising means for attenuating a prediction coefficient at a faster rate than other prediction coefficients.