CN1937432A - Sound echo cancellation processing method based on optimized parameter predication - Google Patents

Abstract

The invention discloses an acoustic echo cancellation processing method based on optimal parameter prediction, which utilizes a robust adaptive filtering parameter iteration method, and combines the correlation parameters of the signal in the system and the change trend of the iteration parameter itself Iterative parameters are used for control. The control method is different from the general system. It is not simply to set the iterative parameters, but to reset some filter coefficient values reasonably, so as to effectively track the echo path under the premise of ensuring the stable operation of the system. jump. This method has a moderate amount of computation and is fully applicable to the current DSP processors, which has great practical value.

Description

Acoustic echo cancellation processing method based on the optimized parameter prediction

One, technical field

The present invention relates to acoustic echo processing method in a kind of communication system, particularly a kind of acoustic echo cancellation processing method of effectively predicting based on optimized parameter.

Two, background technology

In recent years, along with the continuous development of mechanics of communication, and people are for the continuous increase of exchange way demand quickly and easily, and video conferencing system and voip technology have obtained application more and more widely.The existence that it is exactly acoustic echo that this class communication system has a very significant problem, promptly far-end speaker's sound can be transmitted back to far-end by the coupling between near-end speaker system and the microphone, has seriously influenced the speech quality of communication two party.

For addressing this problem, suppress acoustic echo effectively, existing lot of documents is discussed, most processing method all is based on classical NLMS algorithm structure, in conjunction with a state machine communication state is judged, when decision-making system is in both-end and speaks state and idle condition, stop the convergence of filter; When the judgement system was single-ended speaking, filter was normally restrained; When judgement is the echo path saltus step, adjusts the filter iterative parameter filter is restrained again.Such processing method has a very big problem to be: the performance of system is the operational process of dependent status machine too, in case erroneous judgement occurs, particularly both-end is spoken erroneous judgement during for the echo path saltus step, can cause the instability of system, even causes the appearance of uttering long and high-pitched sounds.

Three, summary of the invention

1, goal of the invention: the present invention provides a kind of system stability, can effectively suppress the acoustic echo cancellation processing method based on the optimized parameter prediction of acoustic echo.

2, technical scheme: acoustic echo cancellation processing method of the present invention may further comprise the steps:

(1) system initialization obtains reference signal and raw echo.Comprise given a series of parameter about filter, comprise filter order, filter initial value, correlation parameter initial value, iterative parameter initial value or the like, because the present invention can adjust according to actual conditions each parameter, so the influence that is provided with systematic function of initial parameter is not obvious especially, this also is one of advantage of the present invention.

(2) after obtaining reference signal and raw echo, utilize sef-adapting filter estimated echo signal, and subtract each other with it with raw echo and to obtain the residual echo signal, notice that the residual echo had here both comprised the echo signal of not filtering, also comprised near-end background noise and possible near-end voice.

(3) estimated value with true residual echo and actual speech output is optimized the filter iterative parameter, and judges that whether current this optimal value constantly descends than previous moment, if descend, is designated as condition one.

(4) if condition one satisfies, then estimate the correlation of residual echo signal and echo estimated signal, and judge that whether current this correlation constantly rises than previous moment, if rise, is designated as condition two with the correlation calculations formula.

(5) if condition one and condition two satisfy simultaneously, judge further then whether the correlation of current time is enough big, if enough big, then to part filter parameter assignment again, the size of assignment is proportional to the echo path gain.

(6) a series of relevant parameters of filter of following that obtain according to top deduction carry out the renewal of filter coefficient, and enter next cycle period.

The invention provides a kind of new acoustic echo cancellation processing method, this method no longer clearly makes a distinction adaptive-filtering module and state and module, but utilize a kind of adaptive iteration parameter setting method of robustness, in conjunction with effectively filter initial value design and replacement, reach the purpose of the acoustic echo counteracting that realizes stability and high efficiency.

The parameter iteration formula of general NLMS algorithm is as follows:

$g(k+1)=g\left(k\right)+\frac{\mathrm{\σ}}{{\left|\right|x\left(k\right)\left|\right|}^2}e\left(k\right)x\left(k\right)$

Wherein x represents reference signal, and e represents error signal, and g represents filter coefficient.Though it is regular that iterative parameter has been undertaken by the power of reference signal, can avoid when reference signal is excessive system to disperse, iterative formula still has constant σ to exist.The system that how to guarantee under any circumstance can both guarantee stable and convergence fast, and this depends on the selection of σ to a great extent.In general, a fixing σ can not meet the demands certainly, how to allow σ optimize and revise (this means that σ has become variations per hour σ (k)) according to system running state, and many these problems of relevant documents are arranged.For the echo cancelltion system, reliable theoretical method is according to as follows: the optimal value of σ (k) is determined by formula (2):

${\mathrm{\σ}}_{\mathrm{opt}}\left(k\right)=\frac{E\[{\mathrm{\ϵ}}^2\left(k\right)\]}{E\[e^2\left(k\right)\]}---\left(2\right)$

Wherein, ε (k) represents real residual echo, and e (k) then represents actual residual echo (comprising true residual echo and near-end background noise and possible near-end speech).The denominator of formula (2) obviously is the known quantity of system, but how molecule accurately estimates it is a problem.For steady pumping signal in short-term:

$E\[{\mathrm{\ϵ}}^2\left(k\right)\]=\frac{1}{N}{\left|\right|g\left(k\right)-h\left(k\right)\left|\right|}^2{\left|\right|x\left(k\right)\left|\right|}^2---\left(3\right)$

Wherein g represents adaptive filter coefficient, and h is an echo path impulse response coefficient, and N is a filter order.Supposition on more rational engineering be each parameter of filter the disturbance quantity of iterative process middle distance optimal value be statistics uniformly.This means, if known a part of echo path transfer function optimal value, just could be according to the disturbance quantity of this part all filter coefficient of information inference.A method in the practicality is: artificially increase by one section time-delay to reference signal in the echo cancelltion system, the pairing echo path impulse response of this section time-delay is zero certainly so, sef-adapting filter parameter correspondingly itself is exactly a disturbance quantity, and formula (3) can obtain by following formula is approximate:

$E\[{\mathrm{\ϵ}}^2\left(k\right)\]=\frac{1}{N_p}{\left|\right|g_p\left(k\right)\left|\right|}^2{x\left(k\right)\left|\right|}^2---\left(4\right)$

G wherein _p(k) adaptive filter coefficient of expression delay number correspondence, N _pThen represent time-delay length.Formula (3) also has an approximation method as follows: if sef-adapting filter exponent number long enough, then in general conference system, the parameter optimal value of sef-adapting filter end approaches zero, so we also can approach formula (3) by following formula:

$E\[{\mathrm{\ϵ}}^2\left(k\right)\]=\frac{1}{N_{\mathrm{ifc}}}{\left|\right|g_{\mathrm{ifc}}\left(k\right)\left|\right|}^2{\left|\right|x\left(k\right)\left|\right|}^2---\left(5\right)$

G wherein _Ifc(k) parameter value of expression filter end, N _IfcRepresent corresponding exponent number.

Said method can be so that filter coefficient be in course of adjustment according to the parameter convergence situation, near-end background noise and near-end voice are reasonably regulated auto-adaptive parameter, when the generation both-end is spoken, because it is big that the denominator in the formula (2) obviously becomes, make the filter iterative parameter reduce rapidly like this, avoided dispersing of filter effectively.

For following the trail of the situation of change of echo path, the present invention does following processing: the first, and the correlation parameter R between error signal e and the echo estimated signal y _EyThe second, not only use relevance threshold, and will utilize the variation tendency of correlation.Because when echo path changed, correlation rose certainly, and when both-end was spoken, correlation descended, the correlation rise and fall present random characteristic under remaining situation.Like this, reduce and correlation rising generation simultaneously as long as judge the filter iterative parameter, can assert that just variation has taken place echo path, for the accuracy of judging, also need to add the condition (here relevance threshold just assist parameter, optional leeway big) of correlation greater than certain threshold value.In case after determining that echo path takes place, just need to increase iterative parameter, the method for increase is consistent with the initial convergence method, promptly reset the value (time-delay corresponding parameters or the last some rank of filter parameter) of above-mentioned specific filter parameter.For the setting of filter parameter, can utilize existing R _Ey, with reference signal power to normalization just can obtain reflecting and the parameter of echo path gain size by this parameter the filter segment parameter is set again.

3, beneficial effect: different with existing method, the present invention no longer to system single-ended speak, both-end is spoken, states such as background noise is excessive, echo path saltus step, free time are done clear and definite differentiation, but utilizing a kind of adaptive approach of robustness, the assurance system is stable when excessive and near-end has speaker's sound in near-end noise; When the echo path saltus step, the present invention has utilized the correlation information of signal in the change information of filter iteration step length and the system simultaneously, can realize the tracking of echo path exactly; In addition, when the iterative parameter of sef-adapting filter is set, utilize the gain of echo path to estimate, guaranteed the stability and high efficiency of system.

Four, description of drawings

Accompanying drawing is a theory diagram of the present invention.

Five, embodiment

The present invention is described in detail below by example:

As shown in the figure, system initialization comprises given a series of parameter about filter, comprise filter order, filter initial value, correlation parameter initial value, iterative parameter initial value or the like, because algorithm structure of the present invention all can be adjusted according to actual conditions each parameter, so the influence that is provided with systematic function of initial parameter is not obvious especially, this also is one of advantage of the present invention.

After obtaining reference signal and raw echo, utilize sef-adapting filter estimated echo signal, and subtract each other with it with raw echo and to obtain the residual echo signal, notice that the residual echo had here both comprised the echo signal of not filtering, also comprised near-end background noise and possible near-end voice.

The optimal value of the method estimation filter iterative parameter of being introduced with aforementioned formula (4), (5), and judge that whether current this optimal value constantly descends than previous moment, if descend, is designated as condition one.

Estimate the correlation (desirable multiply each other summation and normalization calculate in the practicality) of residual echo signal and echo estimated signal at 1024 with the correlation calculations formula, and judge whether current this correlation constantly rises than previous moment, if rise, be designated as condition two.

If condition one and condition two satisfy simultaneously, and the correlation of current time is also enough big, then to part filter parameter assignment again.The determination methods whether condition one, two satisfies simultaneously has a variety of, can select to be provided with the mode of indexed variable statistics summation in the practicality, and the increase of operand is very limited like this; Whether enough big condition is subsidiary conditions to correlation in addition, and the selection of threshold value has very big space, in the practicality desirable 0.5.

Again the assignment of filter segment parameter need be used the estimation of echo path gain, succinct on calculating, the echo path gain can be used aforesaid correlation calculations intermediate object program, i.e. 1024 the summed result that multiplies each other, with the Energy Estimation of reference signal to consolidation just can obtain reflecting the actual parameter A of echo path gain.The setting and the A/10 relation in direct ratio of practical median filter parameter.

The top deduction of last foundation obtains the relevant parameter of a series of very filters and carries out the renewal of filter coefficient, enters next cycle period again.

By aforesaid description as can be known the operand of this new method increase seldom, the DSP that is fit to very much system realizes.The present invention has realized above-mentioned algorithm structure on ADSP21161N EZ-LITE plate, after tested, acoustic echo bucking-out system good stability based on this algorithm structure, filter divergence in the time of preventing effectively that excessive and both-end is spoken in near-end background noise, and can track the saltus step phenomenon of echo path effectively.

Claims (1)

1. An acoustic echo cancellation processing method based on optimal parameter prediction, characterized in that the method comprises the following steps: (1) system initialization, obtain the reference signal and the original echo about the filter; (2) Utilize the adaptive filter to estimate the echo signal, and subtract it from the original echo to obtain the residual echo including the unfiltered echo signal, the near-end background noise, and the possible near-end speech residual echo signal; (3) Optimize the filter iteration parameters with the estimated value of the real residual echo and the actual voice output, and judge whether the optimal value at the current moment is lower than that at the previous moment, if it is lower, record it as condition one; (4) If condition one is satisfied, then use the correlation calculation formula to estimate the correlation value of the residual echo signal and the echo estimation signal, and judge whether the correlation value at the current moment is higher than that at the previous moment, and if it rises, record it as condition two; (5) If condition one and condition two are satisfied at the same time, it is further judged whether the correlation value at the current moment is large enough, and if it is large enough, some filter parameters are reassigned, and the size of the assignment is proportional to the echo path gain; (6) Obtain a series of parameters related to the filter according to the above inference to update the filter coefficients, and enter the next cycle.

Images