JPS5917839B2 - Adaptive linear prediction device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an adaptive linear prediction device that is used for audio signal analysis and band compression 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 its past signals. 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. It is not hard to imagine that if a plurality of past sample values are used, more accurate predictions can be made and the prediction error to be transmitted is smaller. At this time, the weighting coefficient of each sample is called a prediction coefficient. For linear prediction, for example, J, Ma
khoul's explanatory paper ('Linear Predi
ction■ATutoichiria1Review''3
Proceedings of the IEEE, Vol.
63, Haruka 4, pp, 561-58O, 5April19
75) or other papers. According to these, the audio signal sequence is X.

, X1, ...Xj-1, the audio signal at time j is Λ
N '0Xj=Σ ・・・・・・・・・・・・・・・・・・
(1) It can be approximated by i■1aiXj-i.

However, ai (i■1, 2,...N) is a prediction coefficient.
ΛHere, the root mean square of the difference between the predicted value Xj and the true value Xj j5
One method for finding the optimal prediction coefficient that minimizes is the autocorrelation coefficient ψ of the audio signal Xj.

, ψ1, ψN, and ψoψ10゜゜゜゜゜゜゜゜゜゜゜ψN−1ψ1ψ1ψ0 ψ2 ・i2) ψN−1゜ψo ψN However, the method 25, which involves solving this linear equation, requires an extremely large amount of calculations to find the autocorrelation. This is a major difficulty when actually building the device.

Another method for determining prediction coefficients is to sequentially modify the prediction coefficients.

In other words, the blade prediction error ej at time j is Λ ej = xj - xj (3
), then the prediction coefficient a)(i■1, 2,...
N) is 96a?゛''a4fγ・f(Xj-i)・g(
ej), (i=1, 2, . . . N)・a4).

Here γ is a modified gain, and f and g are arbitrary non-decreasing functions. By repeating equation (4), a! 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.
．． Alternatively, a method using a sequential method as shown in equation (4) (J.

In either case, the number of prediction coefficients is about 10. 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 obtain correlation coefficients or to transmit prediction coefficients. The present invention therefore 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 prediction error is still a periodic signal. In other words, parts that are difficult to predict appear periodically. The limit of speech band compression by linear prediction is determined by the above-mentioned sound source, and the power ratio (compression ratio) between the original speech signal and the prediction error signal when performing optimal linear prediction is 14 dB when the number of prediction coefficients is around 10. It will be about. To further increase the compression ratio, it is possible to use a method of increasing the number of prediction coefficients, or a method such as that of Atal mentioned above, which measures the period of the sound source in advance and removes the periodicity of the sound source. 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, At
In the method such as al, the circuit becomes extremely complicated because a correlation coefficient must be determined to determine the periodicity, and the detected period must be transmitted.

Also, in voice analysis, in order to determine the period of a sound source, a large amount of calculation is required, such as calculating autocorrelation.The purpose of the present invention is to provide a device that performs voice analysis including the period of the sound source using a simple circuit. be.

Another object of the present invention is to provide a prediction device that enables even greater band compression without complicating the circuit.

According to the present invention, there is provided a first adaptive linear predictor that sequentially corrects its own prediction coefficient using a prediction error, and a second adaptive linear predictor, which are connected in cascade. An adaptive linear prediction device for speech analysis is obtained.

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

FIG. 1 is a diagram explaining the principle of the present invention, and FIG. 2 is a diagram showing the waveforms of each signal. Audio signal series X. ,Xl,...
...Xj is given from the transmission side input terminal 10 and applied to the first predictor 20 having a relatively small number of prediction coefficients. The first predictor 20 obtains an estimated value of Xj based on the input signal up to time j-1.

Here, h1 is the i-th child measurement coefficient of the first predictor at time j. Generally, the input signal should be sampled at 8kHz, and a value of about 10 should be used for N.The input subtracter 21 can detect the true voice? The first prediction error Fj=Xj-Xj is obtained by subtracting the pre-cooked direct publication from issue Xj.

When the original audio signal is represented by x in FIG. 2, the first prediction error is as shown in f in the same figure. In other words, the prediction error increases with the period of the audio source. The first prediction error is applied to a second predictor 30, which outputs an estimate 1-v of the first prediction error F,.

However, L is 10-20. A value of about 100 is selected for M.
Further, p1 is the i-th child measurement coefficient at time j.
7. The prediction coefficients where i is smaller than L in the second predictor 30 are used in the first predictor 20, and the periodicity of the sound source is 7. ′:′:π groan 2;:′I′? 1, Seed::ru. Ej becomes an extremely small signal as shown in FIG. 2e. Here, the first predictor 20 is successively corrected from the input signal sequence and the first prediction error. An example of this is the aforementioned Gibs
On et al., and is generally indicated by .

However, γ is a modified gain, ψ1 and ψ2 are arbitrary non-decreasing numerical numbers, and 1 is a value from 1 to N. However, in order to improve the ability to follow changes in voice, a relatively fast algorithm is required for equation (7). The second predictor 30 is successively corrected from the first prediction error and the second prediction error. Generally expressed, it becomes. Here, ψ3 and ψ4 are arbitrary non-decreasing functions, 1 is the value of LM, and δ is a modified gain. In the second predictor 30, the number of prediction coefficients is quite large, but rapid fluctuations in the period of the speech source are rare. Therefore high modification speeds are not required. Furthermore, high accuracy is not required when erasing periodicity. For these reasons, although the second predictor 30 has a large number of prediction coefficients, it is possible to employ a simple correction algorithm with low calculation accuracy.

A voice analysis device can be implemented only on the transmitting side.

In this case, the prediction coefficients of the first predictor 20 can be converted into information about speech formants or vocal tracts. Also the second
The period of the speech source is obtained from the prediction coefficient of the predictor 30.
These are described in the paper by MakhOul mentioned above. On the other hand, when used for band compression, the second prediction error Ej sent from the transmitting side output terminal is given to the receiving side through the transmission channel 14.

The first adder 41 on the receiving side combines the received signal e.
is the output signal of the predictor 40. Add outside 1 to make FS. The first predictor 40 on the receiving side makes the second prediction on the transmitting side: 'I
,′. ′(′;″,″(, え？ あ き“) Also, the second predictor 50 on the receiving side generates the predicted signal X/j, and the second adder 51 reproduces the original signal X5. However, if there is no error, x・,=Xj,)1=money j.

Furthermore, the first predictor 50 on the receiving side is
) has a correction algorithm for the formula. As is clear from the above description, there is no need to transmit prediction coefficients when sequential linear prediction is performed.

Therefore, in the conventional type, f in FIG. 2 must be transmitted, whereas in the present invention, by transmitting e in FIG. 2, it is possible to significantly reduce the number of transmission bits. Although the principle operation of the present invention has been shown in the above explanation, in reality, there are some differences in the configuration because errors may occur in transmission and calculation, quantization must be performed, etc. The third example is an example according to the present invention, in which the transmitting side predictor and quantizer form a loop. That is, on the transmitting side, at Ji time, the original audio signal sequence Xj is applied from the input terminal S0, and the first subtraction circuit 60 on the transmitting side calculates the difference from the measured value Xj, that is, the first prediction error Fj=Xj-Xj. △Predicted value Xj
is generated in the first predictor 20 by the first predictor 30 based on the past input audio signal using the following equation.

However, Xj-1 is an audio signal sample reproduced on the transmitting side, and N is a value of about 10.

h} is modified as shown in the following equation. However, k is a value sufficiently smaller than 1, and ψ1 and ψ2 are arbitrary non-decreasing functions.

7j is the first prediction error reproduced on the transmitting side.

The first prediction error Fj is
-;=Z Now 1 is obtained in the second predictor 30 by the following equation. Here, Fj-1 is sent! This is the first prediction error reproduced on the side. Further, P} is corrected to 0 according to the following equation. However, t is a value sufficiently smaller than 1. Further, ψ3 and ψ4 are arbitrary non-decreasing functions. The second prediction error Ej is quantized by a quantizer 70 and sent out as a transmission signal Ej. The transmitted signal Ej'' is input to the inverse quantizer 71 which has the inverse characteristic of the quantizer 0, and is given as 7;
O. Yo. This results in a poor input signal.

Also, the first adder 22 obtains 'Xj=endj+maj, which becomes a new input signal to the first predictor 20. On the other hand, on the receiving side, the received signal e is reversed △
It is converted into e by the J quantizer 72 and the first
Addition of j?゛I? : "廿== It is determined by the measuring instrument 40 according to the following formula.

5 shouts {゜? '． 'V engineering 1 layer! ÷Zr weight It is determined by the following equation in the predictor 50 of No. 2.

．． 151. ．． The modification of p~ and h'ν is p! at the transmitter side. 1.11 and h! This is done using the same algorithm as the modification of .

As is clear from the above description, the linear predictors used in this embodiment are relatively simple in terms of circuitry, as they only perform sum-of-products operations and correction of prediction coefficients to obtain predicted values.

Furthermore, two predictors are used on each of the transmitting side and the receiving side, but if the algorithms used by both predictors are similar, the two predictors can be realized by using one arithmetic circuit in a time-sharing manner. Moreover, there is no problem even if the order of the two predictors is switched. As described above, according to the present invention, it is possible to obtain a linear predictive code decoding device that has a small and simple circuit configuration and has a high band compression ratio.

In the above explanation, we only talked about bandwidth compression, but
By removing only the transmitter shown in FIG. 1, it can also be used as a speech analysis device.

In other words, a predictor having a decimal number of prediction coefficients is required to provide information regarding the formant of the voice or the vocal tract, and other predictors are capable of extracting information about the sound source, that is, the pitch frequency. In other words, according to the present invention, linear prediction, which is an extremely effective speech analysis device, can be obtained using a simple circuit.

[Brief explanation of the drawing]

FIG. 1 is a diagram explaining the principle of the present invention, FIG. 2 is a waveform diagram of each part of FIG. 1, and FIG. 3 is a block diagram showing an embodiment of the present invention.

Claims (1)

[Claims] 1. It has a first adaptive linear predictor that sequentially corrects its own prediction coefficients and a second adaptive linear predictor that sequentially corrects its own prediction coefficients, and the two are continuously connected. An adaptive linear prediction device characterized by: