CN109089004B - A Correlation Entropy-Induced Set-Member Adaptive Echo Cancellation Method - Google Patents

Abstract

A set-member adaptive echo cancellation method based on correlation entropy, the steps are as follows: A. Remote signal acquisition, sampling the far-end signal transmitted from the far-end, to obtain the discrete signal of the far-end input signal at the current time n. Value x(n), its filter input signal vector is x(n)=[x(n), x(n-1),...,x(n-L+1)] ^T ; B, echo signal Estimate, pass the input signal vector x(n) of the current moment n through the adaptive filter, and its output value y(n) is the estimated value of the echo signal; C, echo cancellation, use the near-end microphone to sample to obtain the current echo with echo The near-end signal d(n) at time n minus the estimated value y(n) of the echo signal; D, the filter tap weight coefficient is updated, and the tap weight vector w(n+1) of the filter at the next time n+1 is calculated ), w(n+1)=w(n)+μ(n)U(n)( ^UT (n)U(n)) ^-1 E(n)-C(n); E, let n= n+1, repeat the above process of A, B, C, D, and E until the call ends. This method can obtain faster convergence speed and low steady-state error, and the effect of echo cancellation is good.

Description

A Correlation Entropy-Induced Set-Member Adaptive Echo Cancellation Method

technical field

The invention belongs to the technical field of echo cancellation of communication systems.

technical background

As a branch of information processing, adaptive signal processing technology has developed rapidly in recent years and has been widely used in the field of communication. Echo phenomenon in communication system refers to the reflection of sound or signal back to the signal source after delay or deformation. Since the echo phenomenon will seriously affect the quality of people's calls, how to eliminate the echo has become the focus of people's attention. The communication echo can be adaptively eliminated by the system identification model: the identified system is the echo channel, and the output of the system identification is the estimation of the echo signal. The echo can be eliminated by subtracting the speech signal containing the echo signal and the estimation of the echo signal. . Adaptive echo cancellation technology has the advantages of low cost, fast convergence speed and small echo residual, so it has attracted the attention of many researchers, and it is also considered to be the most promising echo cancellation technology in the field of communication.

Echo channels all have sparse characteristics, that is, most of the coefficients of the channel (system) are close to zero or equal to zero, and only a few coefficients have large amplitudes. In this case, traditional adaptive algorithms cannot achieve satisfactory results.

Document 1 "Set-membership affine projection algorithm" (Werner, S., and Diniz, P.S.R., IEEE Signal Process. Lett., 2001, 8, (8), pp. 231–235 Electronics Letters 52.17 (2016): 1461-1463.) Combining the set membership theory with the affine projection method, a set membership affine projection algorithm is proposed. Set membership filter is a kind of error-based recursive step size estimation algorithm, which seeks the step size set that produces bounded filter output error, improves the inherent contradiction between the convergence speed and steady-state error of the fixed-step adaptive algorithm, and can It is guaranteed that the filter has fast convergence speed and low steady-state error. However, when the system is a sparse echo channel, the tap weight vector w(n) at the current moment n of the system accounts for most of the items close to zero or zero, and is interfered by the background noise, so that the update value of the tap weight vector w(n+1 ) is too large, resulting in a large steady-state error, and the effect of echo cancellation needs to be improved.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an adaptive echo cancellation method based on set membership induced by correlation entropy, which has small steady-state error, fast convergence speed and good echo cancellation effect.

The technical solution adopted by the present invention to achieve the object of the invention is a kind of adaptive echo cancellation method based on the set membership induced by correlation entropy, the steps of which are:

A. Remote signal acquisition

Sampling the signal from the far end to obtain the discrete value x(n) of the far-end input signal at the current time n; (n-1),...,x(n-L+1), the input vector x(n) of the current moment n constituting the adaptive filter, x(n)=[x(n),x(n -1),...,x(n-L+1)] ^T , where T represents the transpose operation, and L=512 represents the number of filter taps;

B. Echo signal estimation

Pass the input signal vector x(n) of the current time n through the adaptive filter to obtain the output value of the current time n of the adaptive filter, that is, the estimated value of the echo signal y(n)

y(n)= ^xT (n)w(n)

where w(n) is the tap weight vector of the current moment n of the adaptive filter, w(n)=[w ₁ (n),w ₂ (n),...,w _L-1 (n)] ^T , the initial value of w(n) is zero vector;

C, echo cancellation

Sampling the near-end microphone to obtain the near-end signal d(n) of the current time n with echo, subtract the estimated value y(n) of the echo signal from it, and obtain the error signal e(n) of the current time n, and then send it back For the remote end, e(n)=d(n)-y(n);

D, filter tap weight vector update

D1. Calculate the affine projection matrix of the input signal

The current time n and the input vectors x(n), x(n-1),...,x(n-P+1) of the current time n and the previous P-1 times constitute the input signal affine projection matrix U of the current time n (n), U(n)=[x(n),x(n-1),...,x(n-P+1)]; where P represents the affine projection order, the value range is 2 ~9;

D2. Calculate the error signal vector

The current moment n and the error signals e(n), e(n-1),...,e(n-P+1) of the current moment n and the previous P-1 moments constitute the current moment n error signal vector E(n), E(n)=[e(n),e(n-1),...,e(n-P+1)] ^T ;

D3. Calculate the step factor

The step factor μ(n) of the current time n is calculated by the following formula:

Where γ represents the error threshold parameter, the value range is 0.0001～1;

D4. Calculate the relevant entropy induction factor

From the tap weight vector w(n) of the current time n of the adaptive filter, the relevant entropy factor C(n) of the current time n is calculated:

Among them, ρ is the control factor, the value range is 10 ^-8 ~ 1; σ is the kernel width, the value range is 0.001 ~ 2; exp(·) represents the exponential operation.

D5, filter tap weight vector update

The filter tap weight vector w(n+1) at the next moment n+1 is obtained from the following formula:

w(n+1)=w(n)+μ(n)U(n)( ^UT (n)U(n)) ^-1 E(n)-C(n)

E. Let n=n+1, and repeat the processes of A, B, C, and D until the call ends.

Compared with the prior art, the beneficial effects of the present invention are:

作为减除项(步长调整项)。当为稀疏系统的回声信道时，系统当前时刻n的抽头权向量w(n)中接近零或为零的项通常占大多数，由相关熵诱导因子表达式可以看出，当w(n)越接近零，减除项的相关熵诱导因子越大，与已有的加法项(步长项)抵消调整后，抽头权向量更新值w(n+1)更接近零；实现接近零或为零的抽头权向量的适当小步长更新，从而能够有效的降低背景噪声对抽头权向量更新的影响，稳态误差小，回声消除效果好；相反，当前时刻n的抽头权向量w(n)中存在的非零项，作为减除项的相关熵诱导因子的值小，对加法项(步长项)的抵消小，能充分利用当前时刻非零抽头权向量w(n)中的有用信息，实现大步长的快速更新抽头权向量更新值w(n+1)；加快收敛速度，并减少了稳态误差。The relevant entropy induction factor C(n) is added to the update formula of the filter tap weight vector,

as a subtraction item (step adjustment item). When it is an echo channel of a sparse system, the items close to zero or zero in the tap weight vector w(n) of the current time n of the system usually account for the majority. It can be seen from the expression of the relevant entropy induction factor that when w(n) The closer it is to zero, the greater the related entropy induction factor of the subtraction term, and after offset adjustment with the existing addition term (step term), the updated value of the tap weight vector w(n+1) is closer to zero; The appropriate small step size update of the zero tap weight vector can effectively reduce the influence of background noise on the update of the tap weight vector, the steady-state error is small, and the echo cancellation effect is good; on the contrary, the tap weight vector w(n) at the current moment n The non-zero items existing in , the value of the relevant entropy induction factor as the subtraction item is small, and the offset of the addition item (step term) is small, which can make full use of the useful information in the non-zero tap weight vector w(n) at the current moment , to achieve fast update of tap weight vector update value w(n+1) with large step size; speed up convergence and reduce steady-state error.

The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

Description of drawings

Fig. 1 is the normalized steady-state offset curve obtained by the simulation experiment of Document 1 and the method of the present invention.

Detailed ways:

Example

A specific embodiment of the present invention is an adaptive echo cancellation method based on set membership induced by correlation entropy, the steps of which are:

A. Remote signal acquisition

Sampling the signal from the far end to obtain the discrete value x(n) of the far-end input signal at the current time n; (n-1),...,x(n-L+1), the input vector of the current moment n constituting the adaptive filter, x(n), x(n)=[x(n),x( n-1),...,x(n-L+1)] ^T , where T represents the transpose operation, and L=512 represents the number of filter taps;

B. Echo signal estimation

Pass the input signal vector x(n) of the current time n through the adaptive filter to obtain the output value of the current time n of the adaptive filter, that is, the estimated value of the echo signal y(n)

y(n)= ^xT (n)w(n)

where w(n) is the tap weight vector of the current moment n of the adaptive filter, w(n)=[w ₁ (n),w ₂ (n),...,w _L-1 (n)] ^T , the initial value of w(n) is zero vector;

C, echo cancellation

Sampling the near-end microphone to obtain the near-end signal d(n) of the current time n with echo, subtract the estimated value y(n) of the echo signal from it, and obtain the error signal e(n) of the current time n, and then send it back For the remote end, e(n)=d(n)-y(n);

D, filter tap weight vector update

D1. Calculate the affine projection matrix of the input signal

The current time n and the input vectors x(n), x(n-1),...,x(n-P+1) of the current time n and the previous P-1 times constitute the input signal affine projection matrix U of the current time n (n), U(n)=[x(n),x(n-1),...,x(n-P+1)]; where P represents the affine projection order, the value range is 2 ~9;

D2. Calculate the error signal vector

The current moment n and the error signals e(n), e(n-1),...,e(n-P+1) of the current moment n and the previous P-1 moments constitute the current moment n error signal vector E(n), E(n)=[e(n),e(n-1),...,e(n-P+1)] ^T ;

D3. Calculate the step factor

The step factor μ(n) of the current time n is calculated by the following formula:

Where γ represents the error threshold parameter, the value range is 0.0001～1;

D4. Calculate the relevant entropy induction factor

From the tap weight vector w(n) of the current time n of the adaptive filter, the relevant entropy factor C(n) of the current time n is calculated:

Among them, ρ is the control factor, the value range is 10 ^-8 ~ 1; σ is the kernel width, the value range is 0.001 ~ 2; exp(·) represents the exponential operation.

D5, filter tap weight vector update

The filter tap weight vector w(n+1) at the next moment n+1 is obtained from the following formula:

w(n+1)=w(n)+μ(n)U(n)( ^UT (n)U(n)) ^-1 E(n)-C(n)

E. Let n=n+1, and repeat the processes of A, B, C, and D until the call ends.

Simulation:

In order to verify the effectiveness of the present invention, simulation experiments are carried out, and the method of the reference is compared with the present invention.

The far-end signal x(n) of the simulation experiment is a colored signal, which is generated by Gaussian white noise through a first-order autoregressive process T(z)=1/(1-0.95z ^-1 ), the sampling frequency is 8000Hz, and the number of sampling points is 5000. The echo channel impulse response is obtained in a quiet and closed room with a width of 3.75m, a height of 2.5m, a length of 6.25m, a temperature of 20°C, and a humidity of 50%. The length of the impulse response is the number of filter taps L=32. The background noise of the experiment is Gaussian white noise v(n), and the signal-to-noise ratio is 30dB.

The above-mentioned far-end signal and the corresponding near-end signal are used for echo cancellation using the method of the present invention and the method in Document 1. The optimal parameters of the two methods are listed in Table 1.

Table 1 Approximate values of optimal parameters for the two experimental methods

Document 1 γ=0.07, P=4 this invention γ=0.07, P=4, σ=0.35, ρ=6×10^-6

The simulation experiments were run independently for 50 times to obtain the simulation results. FIG. 1 is a normalized steady-state offset curve diagram of the method of Document 1 and the method of the present invention.

It can be seen from Fig. 1 that in the sparse system, the steady-state error of Document 1 is about -58dB, while the steady-state error of the method of the present invention is about -66dB. Better echo cancellation.

Claims (1)

1. A self-adaptive echo cancellation method based on collective member induced by correlation entropy comprises the following steps:

A. remote signal acquisition

Sampling a signal transmitted from a far end to obtain a discrete value x (n) of a far end input signal at the current moment n; discrete values x (n), x (n-1),. and x (n-L +1) of an input signal at a current time n and L-1 times before the current time n are combined into an input vector x (n) of the current time n of the adaptive filter, x (n) ([ x (n)), x (n-1),. and x (n-L +1)]^TWhere T represents the transpose operation, and L512 represents the number of filter taps;

B. echo signal estimation

The input signal vector x (n) of the current time n passes through an adaptive filter to obtain the output value of the current time n of the adaptive filter, namely the estimated value y (n) of the echo signal

y(n)＝x^T(n)w(n)

Where w (n) is the tap weight vector at the current time n of the adaptive filter, w (n) ═ w₁(n),w₂(n),...,w_L-1(n)]^TThe initial value of w (n) is a zero vector;

C. echo cancellation

Sampling a near-end microphone to obtain a near-end signal d (n) with echo at the current time n, subtracting an estimated value y (n) of the echo signal from the near-end microphone to obtain an error signal e (n) at the current time n, and sending the error signal e (n) back to a far end, wherein e (n) d (n) -y (n);

D. filter tap weight vector update

D1 calculating affine projection matrix of input signal

Forming an input vector x (n), x (n-1),. and x (n-P +1) of a current time n and previous P-1 times into an input signal affine projection matrix U (n) of the current time n, wherein U (n) ([ x (n), x (n-1),. and x (n-P +1) ]; wherein P represents an affine projection order, and the value range is 2-9;

d2 calculating error signal vector

Forming an error signal vector E (n), E (n-1),. and e (n-P +1) of the current time n and the previous P-1 times into an error signal vector E (n) of the current time n, E (n) ([ e (n)), e (n-1),. and e (n-P +1)]^T；

D3 calculating step factor

The step factor μ (n) at the current time n is calculated by:

wherein gamma represents an error threshold parameter, and the value range is 0.0001-1;

d4 calculating relevant entropy inducing factors

Calculating a related entropy factor C (n) of the current time n by using a tap weight vector w (n) of the current time n of the adaptive filter:

wherein rho is a control factor and has a value range of 10^-8-1; sigma is the width of the nucleus, and the range is 0.001-2; exp (·) denotes exponential operation;

d5, filter tap weight vector update

The filter tap weight vector w (n +1) for the next time instant n +1 is derived from:

w(n+1)＝w(n)+μ(n)U(n)(U^T(n)U(n))^-1E(n)-C(n)

E. let n be n +1, repeat the above A, B, C, D procedure until the call is over.

Images