JPS63234733A - Echo canceller - Google Patents

Abstract

PURPOSE:To attain the high performance double talk detection by calculating the mutual correlation coefficient of a transmission signal and an approximate echo replica signal to judge a double talk. CONSTITUTION:An adaptive digital filter 200 forms the approximate facsimile replica of an echo generated by a two-wire/four-wire converting part from a receiving signal x(n), a subtracter 110 subtracts the replica from a transmission signal y(n) and an echo is decreased. The filter 200 executes the adaptive action so that an expected value E can be close to zero unlimitedly to an error e(n) between the signal y(n) and the replica based on a formula I. On the other hand, a double talk detecting device 200 calculates the mutual correlation coefficient of both signals of the left side of a formula II from the signal y(n) and the replica signal, when the coefficient is decreased together with the time, a double talk is decided, a high performance double talk is detected, the adaptive action of a filter 100 is stopped and the deterioration of the echo cancelling quantity is suppressed.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to an echo canceller that cancels echo signals (hereinafter referred to as echoes) in long distance telephone communication lines, etc. 6. Related to an echo canceller equipped with a double talk (hereinafter referred to as double talk) detection circuit for detecting double talk (hereinafter referred to as "double talk") , a voice switch, an echo canceller, etc. Of these three echo control means, the first two have the disadvantage of causing problems such as disconnection at the beginning of the line and clicking noise when turning the line on and off. 6 echo canceller is
It was developed to compensate for these drawbacks. An echo canceller synthesizes an approximate replica (hereinafter referred to as a replica) of an echo based on the characteristics of the echo path obtained by measuring the echo path, and cancels the actual echo with this replica. It does not have a structure that causes trouble such as noise or click noise.

However, with echo cancellers, if a disturbance signal is superimposed on the echo and input to the transmission input terminal, the measurement accuracy of the echo path will decrease, and the amount of cancellation will also be affected by the level ratio between the echo at the transmission input terminal and other signals. Basically, it has the drawback that the value decreases to a value determined by . Therefore,
In order to prevent the amount of cancellation from decreasing in this superimposed state, a double talk detection circuit has conventionally been used.

A conventional echo canceller including a double talk detection circuit will be explained with reference to FIG.

In the same figure, the received signal from the 4-wire side is received at the receiving input terminal 1.
Added to 0. This signal is sent out from the receiving output terminal 20 to a simulated echo path 300. The echo component then reverberates into the transmission input terminal 30 and tries to return from the transmission output terminal 40 to the 4th line side. Here, the adaptive digital filter (ADF) 100 creates a predicted value of the echo component from the received input signal component, and the subtracter 110 attempts to cancel the echo component. In addition, adaptive control is performed using the erased residual signal component so that the erased residual signal approaches zero. On the other hand, when pseudo-expressed external noise (voice of near-end speaker) 400 is added to the transmission input terminal 30,
Since the adaptive control diverges in the wrong direction.

Double talk detector 200 detects this and operates to stop adaptive control. The double talk detector shown in the figure is very commonly used, and compares the received power and the transmitted power of an echo canceller, and determines that double talk is occurring, for example, if the transmitted power exceeds the received power.

[Problem that the invention seeks to solve]

According to the above conventional technology, when double talk occurs,
Obviously, if the extraneous noise (i.e. the near-end speaker's voice) level is low compared to the far-end speaker's voice, the double talk detector will not work and the false adaptation operation will continue, thus resulting in echo cancellation. There was a problem that the ability decreased.

In order to solve this kind of problem, we estimate the cancellation residual echo level, compare this estimated cancellation residual echo level with the actual cancellation residual echo level measurement, and if the latter exceeds the former, it is determined that there is double talk. A method to do this is proposed in the following document.

``A Method and Experimental Results for Detecting Superimposed Speech in an Adaptive Echo Canceller,'' 1985 National Conference of the Institute of Electronics and Communication Engineers, Nα2343゜The purpose of the present invention is to solve the above-mentioned problems in double talk detection in a different way from the above-mentioned literature. The aim is to solve this problem by using a new approach, and to obtain a high-performance double talk detector and, by extension, a high-performance echo canceller.

[Means for solving problems]

In order to achieve the above object, the present invention includes an adaptive digital filter that identifies characteristics of an echo path based on input and output signals of the echo path and synthesizes an approximate echo replica signal, and subtracts the echo replica from the actual echo component. In the configured echo canceller, the transmitted input signal y(n) (represented as a sample value, where n is a time number)
The cross-correlation coefficient ρ between and the approximate echo replica signal 9(n)
y9(n)+(E(・) here represents expected value calculation)
If ρy9(n) increases with time, it is determined that there is no double talk, and if ρy9(n) decreases with time, it is determined that there is double talk and the adaptive operation of the adaptive digital filter is performed. Configure it to stop immediately.

[Effect]

As is well known, an adaptive digital filter calculates a replica 9(n) of the actual input signal y(n),
And error signal e (n) = y (n) − y (n)
...For (2), E(e(n)'') is adapted to approach zero as much as possible.In fact, if y(n) does not include external noise, E(e(n)'') )2)→0 operation proceeds smoothly.

This means that the waveform of replica 9 (n) is infinitely close to the waveform of signal y (n), so y (n)? (n) cross-correlation function p y9
(n) = E (y (n) ・9 (n) ) should be increasing. However, this measure is inappropriate for a signal whose level fluctuates over a wide dynamic range, such as voice, and a cross-correlation coefficient ρyP(n) normalized by the power of the signal expressed by equation (1) is suitable. The cross-correlation coefficient ρy9(n) is infinitely the value 1
However, when double talk occurs, the signal y (n)
Since a component uncorrelated with 9(n) will be included in ρy9(n), ρy9(n) will decrease with good sensitivity.

The echo canceller of the present invention is configured to capture this moment and stop adaptive control of the adaptive digital filter. In other words, when an external signal occurs, the cross-correlation coefficient ρy9
(n) has the property of decreasing from the value 1 with good sensitivity,
The detection is instantaneous. Therefore, the erroneous correction operation of the adaptive digital filter is quickly stopped, and as a result, the deterioration of the amount of echo cancellation is minimized.

〔Example〕

An embodiment of the present invention will be described below with reference to FIGS. 1, 2, and 4.

FIG. 1 is an overall block diagram of an echo canceller according to an embodiment of the present invention. In FIG. 1, 10 represents a received signal input terminal, 20 represents a received signal output terminal, 30 represents a transmitted signal input terminal, and 40 represents a transmitted signal output terminal. Furthermore, 100 is an adaptive digital filter, 110 is a subtracter,
200 is a double talk detector which is a feature of the present invention.

Now, the call signal x(n) from the far end is input to the receiving input terminal IO.
The signal x(n) is applied to an adaptive digital filter 100 that synthesizes echo replicas, and at the same time is sent out from a reception output terminal 20 toward the near-end speaker. The echo component y (n) of this signal x (n) is the echo bus (
(omitted in FIG. 1) and returns to the transmission input terminal 30.

On the other hand, the adaptive digital filter 100 uses the signal x (n)
Create an echo replica 9 (n) from and subtractor 110
Echo cancellation is performed as e (n)=y (n)-9(n). The canceled residual echo component is sent to the transmission output terminal 40.
At the same time, it becomes a drive signal for adaptive control of an adaptive digital filter. The content of the adaptive control is a well-known learning identification method, and the algorithm is shown below.

e (n) =y (n) 9 (n)
(4) Here, N is the number of taps of the appropriate digital filter.

hn(μ) is the μth tap coefficient value at time n, and α is a constant value called a correction coefficient.

Double talk detector 200 has two signals y(n) and 9
(n) and always performs the calculation of equation (1).

When a decrease in the correlation coefficient is detected, it is determined that there is double talk and the correction coefficient α is controlled. That is, when there is no double talk, α=1, for example, and when it is determined that there is double talk, α=O and the coefficient correction of the above equation (5) is stopped.

The second @ is a detailed configuration diagram of the double talk detector 200. In FIG. 2, 201.202.203 are multipliers,
211, 212, and 213 are low-pass digital filters with equal characteristics, 204 and 205 are multipliers, 206 is a divider, and 207 is a multiplier. Further, 221 is a means for performing an averaging operation, 222 is a thinning switch means, and 223
represents a means to select and hold the larger value,
A dotted line portion 220 that combines 221, 222, and 223 is a means for estimating the correlation coefficient by averaging it. Furthermore, 230 is a comparison means, and 240 is a means for controlling the correction coefficient α according to the result of the comparison means 230. Further, 250 is means for monitoring signal power.

Multipliers 201, 202, and 203 each have y'' (n)
Create p y (n)・y (n), y2(n),
The outputs of these multipliers 201, 202, and 203 are output as expected values by time averaging operations by filters 211, 212, and 213 provided at subsequent stages, respectively. In other words, the filters 211, 212, and 213 perform the expected value calculation E(·) of equation (1). And multiplier 204.
205 and the divider 206, the result of the square calculation of equation (1) is obtained. In addition, in this example.

In order to avoid square root calculations, the output of the divider 206 is again set as the correlation coefficient ρy9(n).

Correlation coefficient ρy9 at time n calculated in this way
(n) is input to means 220 for averaging and estimating the correlation coefficient, and the estimated correlation coefficient py9 is calculated. That is, the averaging means 221 averages over a relatively long time window, the switching means 222 thins out the results, the results in the previous time window and the results in the current time window are compared, and the larger value is selected and held. Py is held by means 223 for
Get 9. Average value means 221 and thinning switch 22
2 is specifically composed of a circuit such as an accumulator. In this way, the average estimated value p in the time window of the correlation coefficient ρy9(n), which is approaching the value 1 every moment.
y9 is accumulated in the holding means 223. On the other hand, the instantaneous value ρy9(n) is amplified by β=1+i (ε is a small positive number) times by the multiplier 207 and compared with the above-mentioned yy by the comparing means 230.

W(n) = p y9(n) (1+β) py
9-(6) If W(n)≧O, it is determined that there is no double talk, and the means 240 for controlling the correction coefficient α is turned on. In this case, α=1. Note that the above β times means that α
This is to ensure that α=1, or to ensure that α=1 against slow line fluctuations.

Now, when double talk occurs, the instantaneous value ρy9(n) starts to decrease. The result of equation (6) is negative with good sensitivity,
As a result, α is controlled to be 0. Since the correlation coefficient disturbed by the double talk becomes much lower than 1, its average value decreases and is discarded from the holding means 223.

By the way, if there is no received input signal (if the far-end speaker is not speaking), it is natural that there will be an echo.
(n) and echo replica 9 (n) are zero, and (
1) The operation of the expression becomes mathematically indeterminate. In such cases, it is better to avoid calculations. The monitoring means 250 performs this by comparing the two powers σy2=E (y(n)") cr p" = E (9(n)") with appropriate thresholds and determining whether either If it is determined that there is no signal type, the calculations after division are disabled.Therefore, the previous value of α remains as it is.

FIG. 4 is a diagram showing measurement data in this example. LR represents received signal power, LN represents external noise power,
The upper row of the measured values is the residual power before double talk, and the lower row is the residual power after double talk. As an echo loss condition, the hybrid return loss is 6 db. According to FIG. 4, it can be seen that almost no deterioration is observed. This value is.

In the conventional example shown in FIG. 3, deterioration of about 10 db could not be avoided, so the effect of the present invention is significant.

〔Effect of the invention〕

According to the present invention, it is possible to quickly and accurately detect the occurrence of double talk in an echo canceller including an adaptive digital filter, and it is possible to immediately stop the coefficient correction work of the adaptive digital filter, thereby preventing erroneous correction work. Alternatively, there is an effect that the coefficient divergence can be prevented and the echo cancellation amount can be prevented from decreasing.

[Brief explanation of the drawing]

FIG. 1 is an overall block diagram of an echo canceller according to an embodiment of the present invention, FIG. 2 is a detailed diagram of the double talk detection circuit shown in FIG. 1, and FIG. 3 is a block diagram of a conventional echo canceller. , FIG. 4 is a comparison diagram of erase residual power before and after double talk in the embodiment shown in FIG. 1. 10...Reception input terminal, 20...Reception output terminal, 3
0... Transmission input terminal, 40... Transmission output terminal, 10
0...Adaptive digital filter, 110...Subtractor, 200...Double talk detector, 201.202.
203.204.207... Multiplier, 206... Divider, 211.212.213... Low pass digital filter, 220... Correlation coefficient estimation means, 230...
... Comparison means, 240 ... Correction coefficient control means, 250
...Power monitoring means. Representative Patent Attorney Tadashi Akimoto Minoru 1 [Figure 1 2 δ lI 2 Figure 3 Figure 4

Claims (1)

[Claims] Signal y containing echo components generated in the 1, 2-wire and 4-wire conversion section
(n) (where n is the time point number), its replica ■(
n) and an adaptive digital filter means for synthesizing the error e
(n)=y(n)−■(n);
In an echo canceller configured to adaptively control the adaptive digital filter means using the error e(n),
means for calculating a cross-correlation coefficient ρy (n) between the signal y(n) and the replica (n);
(n) over a long time window and successively replace the average value with a larger value to obtain an estimated value @ρ@y; and the cross-correlation coefficient ρy(n) and the estimated value @ρ@ Means for controlling adaptive control stop of the adaptive digital filter according to the sign or negative of W(n)=ρy■(n)・β−@ρ@y■ (here β is a positive constant) compared with y■ and is configured to disable the adaptive control stop means of the adaptive digital filter and maintain the previous control state when the cross-correlation coefficient ρy(n) cannot be calculated. Echo canceller.