JPH0740680B2 - Double talk detection method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a double-talk detection method in a system to which a digital echo canceller for removing a receiver echo caused by 2-wire to 4-wire conversion or the like is applied.

(Prior Art) FIG. 1 shows the configuration of a system to which a digital echo canceller is applied. In FIG. 1, 1 is a digital signal receiving end from a 4-wire line, 2 is a D / A converter, 3 is a hybrid circuit, 4 is a connecting end of a 2-wire line, 5 is a balancing network, and 6 is an A / D. Converter, 7 is subtractor, 8
Is an echo canceller, and 9 is a digital signal transmitting end to the 4-wire line.

Next, the operation of the above system will be described. The digital signal at the digital signal receiving end 1 is called the far end input, x _j
Will be represented by. This far-end input is converted by the D / A converter 2 into an analog signal x (t). A part of x (t) is input to the A / D converter 6 as an echo, and this echo signal is represented by y (t). The path through which this echo passes is called an echo path, and its characteristic is y
An amount called echo path loss, which is a power ratio between (t) and x (t), is defined. The echo path loss is A _ECHO [d
B]. In addition, 2 to A / D converter 6
Input signal from line circuit is connection end 4, hybrid circuit 3
To enter. This signal is called near-end input, and n (t)
Will be represented by. Therefore, in the A / D converter 6, s
An analog signal of (t) = y (t) + n (t) is input,
Output a digital signal. This digital signal output
Let us denote it by s _j . Further, the digital signal at the digital signal transmitting end 9 is called a residual echo signal and is represented by e _j . The echo canceller 8 _successively estimates the impulse response of the echo path based on the residual echo signal e _j and the far-end input sequence vector X _j (x _j x _j + x _jN ).
And at the next sample time, the estimated impulse response and
A pseudo echo is generated from the input sequence vector X _{j + 1} . At this time, in the subtracter 7, the echo is removed by subtracting the pseudo echo from s _{j + 1} to obtain e _{j + 1} .

However, when the estimation is continued in the double-talk state where the near-end input n (t) is large, the estimated impulse response diverges to a value completely different from the actual impulse response, and conversely.
The estimation error is added to s _j . Therefore, it has a function to stop the estimation during double talk. This double talk detection uses the fact that the echo path loss is A _ECHO ≧ 6 (1) when a coil is used in the hybrid circuit, By.

In this way, by performing the calculations of equations (2), (3), and (4), double talk can be detected for an arbitrary echo path using the hybrid coil.

(Problems to be Solved by the Invention) However, the above method has a problem that the divergence of the estimated impulse response cannot be sufficiently suppressed because of the low double-talk detection sensitivity. In addition, the power calculation of the equations (3) and (4) also requires a data memory of 2 m, and there is a problem that the processing time is long when the calculation is executed by software.

The present invention solves such a problem, and an object of the present invention is to provide an excellent double-talk detection method with high speed and high sensitivity with a small amount of calculation and memory.

(Means for Solving the Problems) In order to achieve the above-mentioned object, the present invention detects double talk by using e _j that _greatly changes in response to a change in n (t). S _Pj = S _Pj-1 +2 ^-q (S _j ² -S _Pj-1 ) (6) E _Pj = E _Pj-1 +2 ^-q (e _j ² -E _Pj-1 ) (7) The detection sensitivity is thereby increased. Here, K is the maximum value of the power ratio of y (t) and n (t) when double talk detection is required, and q is an integer coefficient used in equations (6) and (7), and both are appropriate. Can be set to any value. Also, the power calculation of equations (6) and (7) is
The cyclic filter operation is performed by using the power of the signal before echo removal and the power of the signal after echo removal, respectively, and the required memory is 2 data, and multiplication is performed twice. , Subtraction, shift, addition can be performed.

(Operation) The present invention has the following operations by the above method. That is, when the echo canceller is sufficiently converged, the powers of e _j and n (t) are about the same.
When the power of (t) increases, e _j ² also increases, and the double talk can be quickly detected by the equation (5). Further, since the change in the value of the equation (5) with respect to the power change of n (t) is sufficiently large, the integral calculation as in the equations (6) and (7) is applied to the power calculation without significantly deteriorating the detection sensitivity. As a result, the memory and calculation amount required for double-talk detection can be reduced.

(Embodiment) FIG. 2 shows an operation flow of an embodiment of the present invention. x _j is the far-end input signal and s _j is the echo signal y
The weighted signal of (t) and the near-end input signal n (t) is the A / D converted digital signal. X _j is an input signal sequence vector, and H _j is an estimated impulse response at j sample points. When the sample values x _j and s _j are input, the residual echo e _j is calculated and output based on the estimated impulse response H _{j up} to that point. Furthermore, the power of s _j and e _j is calculated, and double talk determination is performed. In the case of double talk, the step gain α is set to zero and the estimation is stopped, but when not double talk, H _j is updated with α = 2 ⁻³ .

Thus, according to this embodiment, when the power of the near-end input n (t) becomes -15 dB or more of the power of the echo signal y (t). Therefore, there is an advantage that high-sensitivity double-talk detection becomes possible. Further, when a near-end input n (t) having the same power as y (t) is added, double talk is detected with a delay of 4 samples, which has the advantage of enabling high-speed detection of double talk.

In the embodiment, K in the equation (5) is set to 15, but it may be increased within a range that does not impede the progress of estimation. In this case, the effect of further increasing the double-talk detection sensitivity can be obtained. Furthermore, although q in the equations (6) and (7) is set to 5, it may be reduced within a range that does not hinder the safety of the transition from double talk to single talk. In this case, the effect of further increasing the double talk detection speed can be obtained.

(Effects of the Invention) As is apparent from the above embodiment, the present invention is one of the powers of respective signals when the power of a signal before echo removal and the power of a signal after echo removal are obtained in double talk detection. Since the cyclic filter operation is performed using the previous value and the respective powers are compared, there is an effect that the amount of operation and memory for power calculation are small and the calculation processing time can be shortened. . Furthermore, the power ratio before and after echo removal changes significantly during single talk and double talk, and double talk is detected by using this ratio, so the speed of double talk detection can be increased. Have an effect.

[Brief description of drawings]

FIG. 1 is a block diagram of a system to which a digital echo canceller according to the present invention is applied, and FIG. 2 is a flowchart of an embodiment of the present invention. 1 ... Digital signal receiving end, 2 ... D / A converter, 3 ...
… Hybrid circuit, 4 …… 2-wire line connection end, 5 ……
Parancing network, 6 ... A / D converter, 7 ...
Subtractor, 8 ... Echo canceller, 9 ... Digital signal transmitting end.

Claims (1)

[Claims] 1. A system to which an echo canceller is applied,
When obtaining the power of the signal before echo removal by the echo canceller and the power of the signal after echo removal, the cyclic filter operation is performed using the value one before the power of each signal, and the power of the signal A double talk detection method characterized by detecting double talk based on a ratio.