CN106161820A - A kind of interchannel decorrelation method for stereo acoustic echo canceler - Google Patents

Abstract

The invention provides a kind of interchannel decorrelation method for stereo acoustic echo canceler, including: step 1) far-end microphone pickup acoustic reference signal x₁(n) and x₂(n)；Frequency spectrum X is obtained by Fourier transformation₁(k) and X₂(k)；Step 2) to frequency spectrum X₁(k) and X₂K () carries out high frequency decorrelative transformation；Time domain reference signal x after being processed_1d(n) and x_2d(n)；Step 3) to time domain reference signal x₁N () carries out low frequency decorrelative transformation, construct three sinusoidal signals；Step 4) by time domain addition, above three sinusoidal signal is joined x_1dIn (n)；Obtain x₁′(n)；x₂' (n) is x_2d(n)；Step 5) to step 4) the signal x that obtains₁' (n) and x₂' (n) carries out adaptive channel identification by adaptive algorithm, it is achieved echo cancellor.The method of the present invention effectively reduces the Inter-channel Correlation of boombox signal, improves the imbalance performance of stereo echo canceller system.

Description

A kind of interchannel decorrelation method for stereo acoustic echo canceler

Technical field

The present invention relates to acoustic echo technology for eliminating field, particularly to a kind of for stereo acoustic echo canceler Interchannel decorrelation method.

Background technology

In hands-free communication system, there is acoustic echo due to the closed loop transmission of acoustical signal.Such issues that at hands The scenes such as machine communication, TeleConference Bridge, video conferencing system generally exist.Acoustic echo can reduce system Communication quality, it is therefore desirable to eliminated, general employing adaptive channel identification algorithm (list of references [1]: Benesty J,Kellermann K et al.A perspective on stereophonic echo cancellation.Berlin,Heidelberg: Springer-Verlag Berlin Heidelberg, 2011:10.) realize the elimination of echo.In stereo communication system, There is strong Inter-channel Correlation in the boombox signal in near-end room, this causes adaptive channel identification to be calculated The convergency value of method is not unique, so that stereo acoustic echo eliminates the relatively big (list of references of wave filter imbalance of system [2]: Benesty J, Morgan D R et al.A better understanding and an improved solution to the specific problems of stereophonic acoustic echo cancellation.IEEE Trans.Speech and Audio Processing,1998,6(2):156-165.)。

In order to improve imbalance performance, need in advance two passage Near end speaker signal to be carried out decorrelation pretreatment；With Time, in order to ensure communication quality, respective handling should not introduce obvious distortion.Have some the most both at home and abroad can use In the interchannel decorrelation method that stereo acoustic echo eliminates；Principle be in primary signal add a part other Signal or directly primary signal is converted；The most representational is classical non-linear transform algorithm (ginseng Examine document [2]), the method adds the nonlinear transformation composition of a part of primary signal in primary signal.These tradition The defect of method is that quality distortion is big, and decorrelation effect is the strongest, causes quality distortion the biggest.In order to go phase Closing and obtain more preferable compromise between effect and quality distortion, in recent years, domestic and international researcher starts with psychoacoustics Effect carries out interchannel decorrelative transformation, to obtain less quality distortion；Auditory masking effect on psychoacoustics Should (list of references [3]: Valin J M.Perceptually-motivated nonlinear channel decorrelation for stereo acoustic echo cancellation,In Proc.Joint Workshop Hands-Free Speech Communication and Microphone Arrays, Trento, May 2008:188-191.) and fundamental frequency loss effect (list of references [4]: Cecchi S, Romoli L et al.A combined psychoacoustic approach for stereo acoustic echo cancellation.IEEE Trans.Audio,Speech and Language Processing,2011, 19 (6): 1530-1539.) it is used for carrying out low frequency decorrelative transformation.

Compared with adaptive channel discrimination method in early days, method based on psychoacoustics effect can be less at quality distortion Obtain more preferable decorrelation effect simultaneously.But require higher occasion for eliminating echo, current based on psychology sound The method learning effect can not meet higher quality distortion requirement, and the calculating process of the method is complex.

Summary of the invention

It is an object of the invention to overcome the method being currently based on psychoacoustics effect above-mentioned present on elimination echo Defect, it is proposed that a kind of interchannel decorrelation method for stereo acoustic echo canceler, the method can be effective Reduce the Inter-channel Correlation of boombox signal, improve the imbalance performance of stereo echo canceller system, it Ensure that less quality distortion, and realize simple, computation complexity is low.

In order to realize foregoing invention purpose, the present invention proposes a kind of interchannel for stereo acoustic echo canceler and goes Correlation technique, described method includes:

Step 1) far-end microphone pick-up of acoustic signals, obtain stereo reference signal x₁(n) and x₂(n)；Simultaneously to x₁(n) And x₂N () carries out Fourier transformation, obtain frequency spectrum X₁(k) and X₂(k)；

Step 2) to frequency spectrum X₁(k) and X₂K () carries out high frequency decorrelative transformation；Time domain reference signal after being processed x_1d(n) and x_2d(n)；

Step 3) to three-dimensional acoustic reference signal x₁N () carries out low frequency decorrelative transformation, construct three sinusoidal signals: s₁(n)、s₂(n) and s₃(n)；

Step 4) pass through time domain addition by described three sinusoidal signals s₁(n)、s₂(n) and s₃N () joins x_1dIn (n)； Obtain x '₁(n)；x′₂N () is x_2d(n)；Described x '₁(n) and x '₂N () is stereo after Whole frequency band decorrelative transformation Signal；

Step 5) to step 4) the signal x ' that obtains₁(n) and x '₂N () carries out adaptive channel by adaptive algorithm and distinguishes Know, it is achieved echo cancellor.

In technique scheme, described step 2) particularly as follows:

Simultaneously to frequency spectrum X₁(k) and X₂K () carries out high frequency decorrelative transformation:

$\left\{\begin{array}{c}{X}_{1d}\left(k\right)=X_1\left(k\right)\·[(1-\mathrm{\β}\left(k\right))+\mathrm{\β}\left(k\right)\·e^{\mathrm{i\Θ}}]\\ X_{2d}\left(k\right)=X_2\left(k\right)\·[(1-\mathrm{\β}\left(k\right))+\mathrm{\β}\left(k\right)\·e^{-\mathrm{i\Θ}}]\end{array}---\left(5\right)\right.$

Wherein, β (k) is nonlinear weight coefficient, it is contemplated that the pass of the phase sensitive following frequency of human auditory system System, β (k) value is:

$\mathrm{\&beta;}\left(k\right)=\left\{\begin{array}{cc}0.5,& f\left(k\right)\&le;2\mathrm{kHz}\\ 0.5+0.4\&CenterDot;\&CenterDot;\frac{f\left(k\right)-2000}{f_s/2-2000},& 2\mathrm{kHz}<f\left(k\right)<f_s/2\end{array}\right.---\left(6\right)$

Wherein, f (k) is the linear frequency value corresponding to discrete frequency indices k,NFFT is The length of Fourier transformation；f_sSample frequency for signal；β (k) is at f_sMaximum is taken at/2；

Judge that analog frequency f (k) corresponding to k is more than f₀Whether set up, described f₀It is high frequency or low for division signals The marginal value of frequency；If a determination be made that negative, Θ value is set to 0, otherwise, Θ value is set to sinusoidal change The value changed；

To frequency spectrum X_1d(k) and X_2dK () carries out inverse Fourier transform, obtain the time domain reference after high frequency decorrelative transformation Signal x_1d(n) and x_2d(n)。

In technique scheme, described f₀Value be: f₀=1kHz.

In technique scheme, described step 3) specifically include:

Step 3-1) calculate reference signal x₁The weighted auto-correlation function ψ of (n)_w(τ)；

x₁N the average magnitude difference function φ (τ) of () is defined as:

φ (τ)=E{ | x₁(n)-x₁(n+τ)|} (7)

Wherein, τ is time delay, and E{} represents and seeks desired operation；

x₁N the auto-correlation function of () is calculated as follows:

ψ (τ)=E{x₁(n)x₁(n+τ)}. (8)

Weighted auto-correlation function is:

ψ_w(τ)=ψ (τ)/[φ (τ)+ε], (9)

Wherein ε is to prevent quotient from dissipating the fixed value introduced, ε=1；

Step 3-2) calculate fundamental frequency cycles estimated value τ₀；

τ₀={ τ: ψ_w(τ)=max (ψ_w(τ))；f_s/350≤τ≤f_s/60} (10)

Wherein, max () expression takes maximum；

Step 3-3) recalculate fundamental frequency estimated value f₀；

To with maximum ψ_w(τ₀Adjacent 3 ψ centered by)_w(τ) value carries out interpolation, then searches again for maximum and obtains Precision higher fundamental frequency cycles estimated value T₀, then fundamental frequency value is:

f₀=1/T₀ (11)

Step 3-4) retrieval reference signal frequency spectrum X₁K (), carries out amplitude weighting and obtains the amplitude of three low order harmonics compositions A_iAnd first phase

k₀Represent the frequency indices corresponding to fundamental frequency, then two, three harmonics correspond respectively to 2k₀And 3k₀；Utilize three Individual frequency indices value retrieval X₁K (), respectively obtains fundamental frequency and the amplitude of two, three harmonics and first phase, carries out amplitude and add Power obtains:

Wherein, A₁、And σ₁For fundamental frequency amplitude, first phase and the amplitude weighting factor, A₂、And σ₂Humorous for secondary Amplitude, first phase and the amplitude weighting factor, the A of frequency₃、And σ₃For adding for the amplitude of three harmonics, first phase and amplitude Weight factor；Angle{ } represent plural number taken phase function, σ_i(i=1,2,3) is in [0,1] interior value, it is preferred that σ₁=0.1, σ₂=0.2, σ₃=0.3；

Step 3-5) construct three sinusoidal signals s₁(n)、s₂(n) and s₃(n):

Wherein, f_i=i f₀, i=1,2,3.

In technique scheme, described adaptive algorithm is: least fibre method, normalized least fibre method or frequency Territory least fibre method.

It is an advantage of the current invention that:

1, the method for the present invention utilizes the nonlinear Phase Modulation of time-varying to tradition nonlinear transformation decorrelation at high band Method improves, thus obtains more preferable high frequency decorrelation effect；

2, the method for the present invention composes advantageous effect based on psychoacoustics, Implantation Energy near three low order harmonics compositions More weak sinusoidal signal, to carry out low frequency decorrelative transformation, which ensure that less quality distortion；

3, the method for the present invention utilizes time domain addition to inject in primary signal by sinusoidal signal, makes algorithm have relatively low Operand；

4, the energy size of the sinusoidal signal that the method for the present invention is injected by regulation is lost with tonequality in decorrelation effect Compromise in various degree is carried out between Zhen, therefore can be according to the requirement of concrete application scenarios by sine when reality is applied The energy of signal is arranged to different value, such as, for those occasions less demanding to tonequality, can increase and be injected The energy of sinusoidal signal to obtain more preferable decorrelation effect.

Accompanying drawing explanation

Fig. 1 is that stereo acoustic echo eliminates model general illustration；

Fig. 2 is the present invention schematic flow sheet for the interchannel decorrelation method of stereo acoustic echo canceler；

Fig. 3 is the interchannel phase obtained of the method using the present invention in example 1 and existing non-linear transform algorithm Responsibility number comparison diagram；

Fig. 4 is the imbalance performance obtained of method and the existing non-linear transform algorithm using the present invention in example 1 Comparison diagram；

Fig. 5 is the voice quality obtained of method and the existing non-linear transform algorithm using the present invention in example 1 The comparison diagram of perception assessment.

Detailed description of the invention

As it is shown in figure 1, in hands-free communication system, the acoustical signal of far-end speaker is by the stereomicrophone of far-end Device receives, and obtains stereo reference signal x of two passages₁(n) and x₂(n).Stereo reference signal x₁(n) and x₂(n) Through decorrelation pretreatment, the signal x after being processed₁' (n) and x₂'(n).Then by line transmission to near-end, x₁' (n) and x₂' (n) again by near-end stereophonic loud-speaker system play.Two microphone channel signals of near-end are respectively d₁(n) and d₂(n), the signal that i.e. near-end microphone receives, it is divided into two parts: a part is x₁' (n) and x₂' (n) by After Near end speaker plays back, its acoustical signal is propagated in proximal void, reflection in near-end room and dissipating After penetrating, arrive the signal of near-end microphone；Another part is the background noise in near-end scene and voice, i.e. Near end signal.Near-end microphone signal d₁(n) and d₂N the Part I signal in () can be returned to again far by line transmission End, and reset by far-end loudspeaker in far-end scene.So, the speaker in far-end scene just hears The sound of a part oneself, here it is need the acoustic echo eliminated.

If the signal vector corresponding to two channel speakers signals is respectively as follows:

x₁' (n)=[x₁′(n-L+1) x₁′(n-L+2) ... x₁′(n)]

x₂' (n)=[x₂′(n-L+1) x₂′(n-L+2) ... x₂′(n)]

The acoustic echo presented in far-end scene mainly due to caused by the acoustic propagation in near-end scene, because of This utilizes adaptive channel identification to carry out acoustic echo elimination in near-end scene.h₁₂(n) and h₂₂N () represents near respectively In end scene, two loudspeaker channel are to the room impulse response of left microphone,WithIt is adaptive the most respectively Answer passage identification algorithm to h₁₂(n) and h₂₂The estimation of (n)；h₂₁(n) and h₁₁N () represents respectively in near-end scene and raises for two Sound device passage to the room impulse response of right microphone,WithIt is adaptive channel identification algorithm the most respectively To h₂₁(n) and h₁₁The estimation of (n).

Adaptive algorithm is passed through

$e_1\left(n\right)=d_1\left(n\right)-[{\hat{h}}_{21}\left(n\right)*{x_2}^{\&prime;}\left(n\right)+{\hat{h}}_{11}\left(n\right)*{x_1}^{\&prime;}\left(n\right)]---\left(1\right)$

$e_2\left(n\right)=d_2\left(n\right)-[{\hat{h}}_{22}\left(n\right)*{x_2}^{\&prime;}\left(n\right)+{\hat{h}}_{12}\left(n\right)*{x_1}^{\&prime;}\left(n\right)]---\left(2\right)$

Obtain the residue signal e after echo cancellor₁(n) and e₂(n)；e₁(n) and e₂(n) stereo by far-end scene Speaker is play, and so can eliminate the acoustic echo occurred in far-end scene.

Wave filter imbalance is defined as:

$\mathrm{\&eta;}\left(n\right)=20\&CenterDot;\lg \frac{\underset{i=1}{\overset{2}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{2}{\mathrm{\&Sigma;}}}{\left|\right|h_{\mathrm{ij}}\left(n\right)-{\hat{h}}_{\mathrm{ij}}\left(n\right)\left|\right|}_2}{\underset{i=1}{\overset{2}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{2}{\mathrm{\&Sigma;}}}{\left|\right|h_{\mathrm{ij}}\left(n\right)\left|\right|}_2}---\left(3\right)$

Wherein, | | | |₂Represent 2 norms.

If there is no decorrelation pretreatment, bigger wave filter imbalance value η (n) will be there is.It is to say, self adaptation is led to There is bigger wave filter estimation difference in road identification algorithm.Especially physical pulse response changes when, greatly Wave filter imbalance will cause adaptive tracing to slow even cannot accurately following the tracks of it to change, so that echo disappears Except effect severe exacerbation.Its reason is, if not having decorrelation pretreatment, then

$\left\{\begin{array}{c}{x_1}^{\&prime;}\left(n\right)=x_1\left(n\right)\\ {x_2}^{\&prime;}\left(n\right)=x_2\left(n\right)\end{array}\right.---\left(4\right)$

Due to stereo reference signal x₁(n) and x₂N () comes from the same speaker in far-end scene, between them There is the strongest Inter-channel Correlation.This strong correlation will make the convergency value of adaptive channel identification algorithm the most only One, the sef-adapting filter converged to during actual applicationMay and be not equal to desired by us The physics transmission function h obtained_ij(n), i=1,2；J=1,2, thus cause big wave filter to be lacked of proper care.

In order to solve this problem, improve imbalance performance, generally in reference signal x₁(n) and x₂N () is delivered to near-end field Before scape, it is carried out decorrelation pretreatment, thus effectively reduce wave filter imbalance, improve stereo acoustic echo and disappear Except performance.

The present invention will be further described in detail with specific embodiment below in conjunction with the accompanying drawings.

As in figure 2 it is shown, a kind of interchannel decorrelation method in stereo acoustic echo eliminates, concrete steps As follows:

Step 1) far-end microphone pick-up of acoustic signals, obtain stereo reference signal x₁(n) and x₂(n)；Simultaneously to x₁(n) And x₂N () carries out Fourier transformation (FFT), obtain frequency spectrum X₁(k) and X₂(k)；

Step 2) to frequency spectrum X₁(k) and X₂K () carries out high frequency decorrelative transformation；Time domain reference signal after being processed x_1d(n) and x_2d(n)；Specifically include:

Simultaneously to frequency spectrum X₁(k) and X₂K () carries out high frequency decorrelative transformation:

$\left\{\begin{array}{c}{X}_{1d}\left(k\right)=X_1\left(k\right)\&CenterDot;[(1-\mathrm{\&beta;}\left(k\right))+\mathrm{\&beta;}\left(k\right)\&CenterDot;e^{\mathrm{i\&Theta;}}]\\ X_{2d}\left(k\right)=X_2\left(k\right)\&CenterDot;[(1-\mathrm{\&beta;}\left(k\right))+\mathrm{\&beta;}\left(k\right)\&CenterDot;e^{-\mathrm{i\&Theta;}}]\end{array}---\left(5\right)\right.$

Wherein, β (k) is nonlinear weight coefficient, it is contemplated that the pass of the phase sensitive following frequency of human auditory system System, β (k) value is:

$\mathrm{\&beta;}\left(k\right)=\left\{\begin{array}{cc}0.5,& f\left(k\right)\&le;2\mathrm{kHz}\\ 0.5+0.4\&CenterDot;\frac{f\left(k\right)-2000}{f_s/2-2000},& 2\mathrm{kHz}<f\left(k\right)<f_s/2\end{array}\right.---\left(6\right)$

Wherein, f (k) is the linear frequency value corresponding to discrete frequency indices k,NFFT is The length of Fourier transformation；f_sSample frequency for signal；β (k) is at f_sMaximum is taken at/2；

Judge that analog frequency f (k) corresponding to k is more than f₀Whether set up, described f₀For division signals be high frequency still The marginal value of low frequency；Preferably, f₀=1kHz；If a determination be made that negative, Θ value is set to 0, otherwise, Θ value is set to the value of sinusoidal variations；

To frequency spectrum X_1d(k) and X_2dK () carries out inverse Fourier transform (IFFT), obtain after high frequency decorrelative transformation time Territory reference signal x_1d(n) and x_2d(n)。

Step 3) to three-dimensional acoustic reference signal x₁N () carries out low frequency decorrelative transformation, construct three sinusoidal signals； Specifically include:

Step 3-1) calculate stereo reference signal x₁The weighted auto-correlation function ψ of (n)_w(τ)；

x₁N average magnitude difference function (AMDF) φ (τ) of () is defined as:

φ (τ)=E{ | x₁(n)-x₁(n+τ)|} (7)

Wherein, τ is time delay, and E{} represents and seeks desired operation；

x₁N the auto-correlation function of () is calculated as follows:

ψ (τ)=E{x₁(n)x₁(n+τ)}. (8)

Weighted auto-correlation function is:

ψ_w(τ)=ψ (τ)/[φ (τ)+ε], (9)

Wherein ε is to prevent quotient from dissipating the fixed value introduced, preferred value ε=1.

Step 3-2) calculate fundamental frequency cycles estimated value τ₀；

Weighted auto-correlation function ψ_w(τ) maximum ψ_w,maxCorresponding time delay number τ₀It is the estimation of fundamental frequency cycles Value；The span of fundamental frequency is defined to [60,350] Hz, and this is equivalent to limit τ₀Hunting zone, it may be assumed that

τ₀={ τ: ψ_w(τ)=max (ψ_w(τ))；f_s/350≤τ≤f_s/60} (10)

Wherein, max () expression takes maximum.

Step 3-3) recalculate fundamental frequency estimated value f₀；

In order to improve the accuracy of detection of fundamental frequency, further to maximum ψ_w(τ₀Adjacent 3 ψ centered by)_w(τ) value Carry out interpolation, then search again for maximum and obtain precision higher fundamental frequency cycles estimated value T₀, then fundamental frequency value is:

f₀=1/T₀ (11)

Step 3-4) retrieval reference signal frequency spectrum X₁K (), carries out amplitude weighting and obtains the amplitude of three low order harmonics compositions A_iAnd first phase

k₀Represent the frequency indices corresponding to fundamental frequency, then two, three harmonics correspond respectively to 2k₀And 3k₀；Utilize three Individual frequency indices value retrieval X₁K (), respectively obtains fundamental frequency and the amplitude of two, three harmonics and first phase, carries out amplitude and add Power obtains:

Wherein, A₁、And σ₁For fundamental frequency amplitude, first phase and the amplitude weighting factor, A₂、And σ₂Humorous for secondary Amplitude, first phase and the amplitude weighting factor, the A of frequency₃、And σ₃For adding for the amplitude of three harmonics, first phase and amplitude Weight factor；Angle{ } represent plural number taken phase function, σ_i(i=1,2,3) is in [0,1] interior value, it is preferred that σ₁=0.1, σ₂=0.2, σ₃=0.3.

Step 3-5) construct three sinusoidal signals s₁(n)、s₂(n) and s₃(n):

Wherein, f_i=i f₀, i=1,2,3.

Step 4) pass through time domain addition by described three sinusoidal signals s₁(n)、s₂(n) and s₃N () joins x_1dIn (n)； Obtain x '₁(n)；x′₂N () is x_2d(n)；x′₁(n) and x '₂N () is the stereophonic signal after Whole frequency band decorrelative transformation:

$\left\{\begin{array}{c}{x}_1^{\&prime;}\left(n\right)=x_{1d}\left(n\right)+s_1\left(n\right)+s_2\left(n\right)+s_3\left(n\right)\\ x_2^{\&prime;}\left(n\right)=x_{2d}\left(n\right)\end{array}\right.---\left(14\right)$

Step 5) to step 4) the signal x ' that obtains₁(n) and x '₂N () carries out adaptive by normalized least fibre method Answer passage identification, it is achieved echo cancellor；

Within the n-th sampling period, utilize the sef-adapting filter in the n-th sampling periodTo x '₁(n) and x′₂N () filters, obtain the echo signal y at microphone_j(n), it may be assumed that

$y_j\left(n\right)={\mathrm{\&Sigma;}}_{i=1}^2{\hat{h}}_{\mathrm{ij}}\left(n\right)*{x_i}^{\&prime;}\left(n\right)---\left(15\right)$

Again at microphone signal d_jN echo signal is eliminated by (), obtain residue signal e_j(n):

e_j(n)=d_j(n)-y_j(n) (16)

Then sef-adapting filter is updated, obtains the sef-adapting filter in (n+1)th sampling period ${\hat{h}}_{\mathrm{ij}}(n+1):$

${\hat{h}}_{\mathrm{ij}}(n+1)={\hat{h}}_{\mathrm{ij}}\left(n\right)+\frac{\mathrm{\&mu;}}{{\mathrm{\&Sigma;}}_{i=1}^2{\mathrm{\&Sigma;}}_{m=n-L+1}^n{\left|{x_i}^{\&prime;}\left(m\right)\right|}^2}e\left(n\right){x_i}^{\&prime;}\left(n\right)---\left(17\right)$

Wherein, x_i' (n)=[x_i′(n-L+1) x_i′(n-L+2) ... x_i′(n)]。

For the next sampling period, then continue to carry out adaptive tracing from formula (17) to formula (19), until signal ended.

Described normalized least fibre method (NLMS) can be other adaptive algorithm: least fibre method (LMS) Or frequency domain least fibre method (FBLMS).

Below in conjunction with example, the present invention will be further described.

Example 1: near-end room a size of 4 × 3 × 3m³, two Near end speaker lay respectively at 1,2,1.2}m and {3,2,1.2}m；Investigation is positioned at that { the near-end microphone at 1.8,1,1.2}m, the acoustical signal it picked up is regarded as closely End microphone signal d (n).The white Gaussian noise simulation near end signal using signal to noise ratio to be 30dB.Sample rate takes 16 KHz, utilizes NLMS algorithm to carry out adaptive channel identification.By the method for the present invention and traditional nonlinear transformation Decorrelation method (NLT) carries out performance comparison.

The interchannel coherence factor of two channel signals x1 (n) and x2 (n) (Inter-Channel Coherence Coefficient, ICCC) it is defined as:

$\mathrm{ICCC}\left(k\right)=\frac{\left|E\right\{X_1^{*}\left(k\right)\&CenterDot;X_2\left(k\right)\left\}\right|}{\sqrt{E\left\{{\left|X_1\left(k\right)\right|}^2\right\}\&CenterDot;E\left\{{\left|X_2\left(k\right)\right|}^2\right\}}}---\left(18\right)$

ICCC is in [0,1] interior value, and its value is the least, and inter-channel coherence is the most weak, and decorrelation effect is the best.Such as Fig. 4 Shown in, in the result of three methods, the ICCC value of two signals after being processed by the method for the present invention is Little.

In imbalance learning curve, reach stable state the shortest, represent that convergence rate is the fastest；Its value is the least, identification essence Spend the highest.As Fig. 4 shows, in the result of three methods, the signal after being processed by the method for the present invention is lost Adjust best performance.

Perceptual evaluation of speech quality (Perceptual Evaluation of Speech Quality, PESQ) is to multiple human ear Auditory properties models, by subjectivity tonequality objective quantification.PESQ is value between-0.5 to 4.5, and its value is the biggest, table Show that distortion is the least.As it is shown in figure 5, in the result of three methods, after being processed by the method for the present invention The PESQ value of signal is bigger than the PESQ value of signal after NLT process.

Claims (5)

1., for an interchannel decorrelation method for stereo acoustic echo canceler, described method includes:

Step 1) far-end microphone pick-up of acoustic signals, obtain stereo reference signal x₁(n) and x₂(n)；Simultaneously to x₁(n) And x₂N () carries out Fourier transformation, obtain frequency spectrum X₁(k) and X₂(k)；

Step 2) to frequency spectrum X₁(k) and X₂K () carries out high frequency decorrelative transformation；Time domain reference signal after being processed x_1d(n) and x_2d(n)；

Step 3) to three-dimensional acoustic reference signal x₁N () carries out low frequency decorrelative transformation, construct three sinusoidal signals: s₁(n)、s₂(n) and s₃(n)；

Step 4) pass through time domain addition by described three sinusoidal signals s₁(n)、s₂(n) and s₃N () joins x_1dIn (n)； Obtain x₁′(n)；x₂' (n) is x_2d(n)；x₁' (n) and x₂' (n) is the stereophonic signal after Whole frequency band decorrelative transformation；

Step 5) to step 4) the signal x that obtains₁' (n) and x₂' (n) carries out adaptive channel by adaptive algorithm and distinguishes Know, it is achieved echo cancellor.

Interchannel decorrelation method for stereo acoustic echo canceler the most according to claim 1, it is special Levy and be, described step 2) particularly as follows:

Simultaneously to frequency spectrum X₁(k) and X₂K () carries out high frequency decorrelative transformation:

$\left\{\begin{array}{c}{X}_{1d}\left(k\right)=X_1\left(k\right)\&CenterDot;[(1-\mathrm{\&beta;}\left(k\right))+\mathrm{\&beta;}\left(k\right)\&CenterDot;e^{\mathrm{i\&Theta;}}]\\ X_{2d}\left(k\right)=X_2\left(k\right)\&CenterDot;[(1-\mathrm{\&beta;}\left(k\right))+\mathrm{\&beta;}\left(k\right)\&CenterDot;e^{\mathrm{i\&Theta;}}]\end{array}\right.---\left(5\right)$

Wherein, β (k) is nonlinear weight coefficient, it is contemplated that the pass of the phase sensitive following frequency of human auditory system System, β (k) value is:

$\mathrm{\&beta;}\left(k\right)=\left\{\begin{array}{cc}0.5,& f\left(k\right)\&le;2\mathrm{kHz}\\ 0.5+0.4\&CenterDot;\frac{f\left(k\right)-2000}{f_s/2-2000},& 2\mathrm{kHz}<f\left(k\right)<f_s/2\end{array}\right.---\left(6\right)$

Wherein, f (k) is the linear frequency value corresponding to discrete frequency indices k,NFFT is The length of Fourier transformation；f_sSample frequency for signal；β (k) is at f_sMaximum is taken at/2；

Judge that analog frequency f (k) corresponding to k is more than f₀Whether set up, described f₀It is high frequency or low for division signals The marginal value of frequency；If a determination be made that negative, Θ value is set to 0, otherwise, Θ value is set to sinusoidal change The value changed；

To frequency spectrum X_1d(k) and X_2dK () carries out inverse Fourier transform, obtain the time domain reference after high frequency decorrelative transformation Signal x_1d(n) and x_2d(n)。

Interchannel decorrelation method for stereo acoustic echo canceler the most according to claim 2, it is special Levy and be, described f₀Value be: f₀=1kHz.

Interchannel decorrelation method for stereo acoustic echo canceler the most according to claim 2, it is special Levy and be, described step 3) specifically include:

Step 3-1) calculate reference signal x₁The weighted auto-correlation function ψ of (n)_w(τ)；

x₁N the average magnitude difference function φ (τ) of () is defined as:

φ (τ)=E{ | x₁(n)-x₁(n+τ)|} (7)

Wherein, τ is time delay, and E{} represents and seeks desired operation；

x₁N the auto-correlation function of () is calculated as follows:

ψ (τ)=E{x₁(n)x₁(n+τ)}. (8)

Weighted auto-correlation function is:

ψ_w(τ)=ψ (τ)/[φ (τ)+ε], (9)

Wherein ε is to prevent quotient from dissipating the fixed value introduced, ε=1；

Step 3-2) calculate fundamental frequency cycles estimated value τ₀；

τ₀={ τ: ψ_w(τ)=max (ψ_w(τ))；f_s/350≤τ≤f_s/60} (10)

Wherein, max () expression takes maximum；

Step 3-3) recalculate fundamental frequency estimated value f₀；

To with maximum ψ_w(τ₀Adjacent 3 ψ centered by)_w(τ) value carries out interpolation, then searches again for maximum and obtains Precision higher fundamental frequency cycles estimated value T₀, then fundamental frequency value is:

f₀=1/T₀ (11)

Step 3-4) retrieval reference signal frequency spectrum X₁K (), carries out amplitude weighting and obtains the amplitude of three low order harmonics compositions A_iAnd first phase

k₀Represent the frequency indices corresponding to fundamental frequency, then two, three harmonics correspond respectively to 2k₀And 3k₀；Utilize three Individual frequency indices value retrieval X₁K (), respectively obtains fundamental frequency and the amplitude of two, three harmonics and first phase, carries out amplitude and add Power obtains:

Wherein, A₁、And σ₁For fundamental frequency amplitude, first phase and the amplitude weighting factor, A₂、And σ₂Humorous for secondary Amplitude, first phase and the amplitude weighting factor, the A of frequency₃、And σ₃For adding for the amplitude of three harmonics, first phase and amplitude Weight factor；Angle{ } represent plural number taken phase function, σ_i(i=1,2,3) is in [0,1] interior value, it is preferred that σ₁=0.1, σ₂=0.2, σ₃=0.3；

Step 3-5) construct three sinusoidal signals s₁(n)、s₂(n) and s₃(n):

Wherein, f_i=i f₀, i=1,2,3.

Interchannel decorrelation method for stereo acoustic echo canceler the most according to claim 4, it is special Levying and be, described adaptive algorithm is: least fibre method, normalized least fibre method or frequency domain least fibre method.