CN101826328A - Echo offset method in embedded wireless visual doorbell - Google Patents

Abstract

The invention discloses an echo offset method in an embedded wireless visual doorbell, comprising the following steps: 1) initializing; 2) obtaining the far-end signal, the real echo path, the near-end voice signal, the background noise, the near-end microphone input signal and the error signal of the visual doorbell at n moments and an estimated echo copy signal output by a filter; 3) judging whether double-end conversation is carried out, if no, directly calculating the error signal and outputting; and if yes, updating a self-adaption filter coefficient, and then calculating the error signal; 4) whitening the far-end signal and the error signal, and calculating an echo path mismatch value; and 5) judging that the delta H (n) obtained by calculation exceeds the echo path mismatch threshold value, arranging a sign for generating double-end conversation, and judging whether treatment is finished; if no, returning to step 2), and if yes, ending treatment. The echo offset method has small calculation cost and can effectively improve convergence performance.

Description

Echo cancelling method in the embedded wireless visual doorbell

Technical field

The present invention relates to the radio visual doorbell field, the echo cancelling method in especially a kind of embedded wireless visual doorbell.

Background technology

Voice are the piths in the multimedia communication terminal service.Voice transfer has been gone through century more than one to the multimedia communication of integrated voice, data, video again from initial " fool " phone digital telephone, wireless telephone finally, and voice transfer all has a problem all the time, is exactly the voice echoes.Voice echo problem is perplexing the user always, and it has had a strong impact on the QoS (Quality of Service) of speech transmission service.Therefore cancelling out echo has just become one of main design problem, also is a considerable part in the global design of voice solution.The voice echo mainly comprises acoustic echo and electricity echo, and along with the digitizing and the development of semiconductor of communicating by letter, the electricity echo can be well controlled on physical device, yet acoustic echo still has a strong impact on the QoS of speech business.Acoustic echo is meant that the sound of loudspeaker broadcast when the side of being received hears, is also picked up by microphone by multiple path.

In the last few years at voice echo problem, the develop rapidly of various cancelling out echo technology, conclude and get up probably to have three kinds: the one, the network damped method, the 2nd, the voice echo suppresses method, the 3rd, modern auto-adaptive filtering technique.First method is to insert an attenuator on both transmit and receive direction respectively, so the decay of the attenuation ratio speech of echo is big 2 times.Yet the speech decay increases along with the increase of transmission range, may cause electrical speech level to drop to the stage that can't catch at last.Second method is by opening transmit path or insert the purpose that bigger decay reaches the control echo on transmit path.Ideally, echo suppressor should be opened transmit path and close transmit path when remote subscriber receives when remote subscriber is spoken, but this is difficult to accomplish.The third method is to estimate the characteristic parameter of echo path, produces the echo path of a simulation, draws the simulation echo signal, deducts this signal from received signal, realizes cancelling out echo.This technology relies on adaptive algorithm, and the quality of adaptive algorithm has determined the quality of cancelling out echo effect.The adaptive algorithm that guarantees speech business transmission robustness has very complicated calculating degree and long convergence time usually.

Using maximum in the echo neutralizer auto adapted filtering is finite impulse response (FIR) wave filter, and for the automatic adaptation FIR system because simple and calculation cost is little, lowest mean square (LMS) and normalization minimum mean-square (NLMS) algorithm application are maximum.Though yet these two kinds of algorithm calculation costs are less relatively, its convergence is too slow.Recursive least-squares (RLS) algorithm convergence is fast, but its calculation cost is too big.Echo neutralizer also comprises the double talk detection device, and it can guarantee speech quality when dual end communication takes place.Usually the DTD algorithm has three classes: based on the detection of energy, based on the detection of correlativity with based on the detection of echo path.Detection algorithm based on energy is simple, and complexity is low, but under the situation of low signal-to-noise ratio the False Rate height; It is comparatively accurate to judge based on the detection of correlativity, but its computation complexity is too high; Detection complexity based on echo path is low slightly, but it is judged by accident when echo path change easily.Wherein echo path impulse response variance detection algorithm has lower computation complexity, and the susceptibility when having avoided for echo path change by the adjusting of converging factor.But owing to of the fluctuation of its testing process based on filter coefficient, when sef-adapting filter is not restrained, the as easy as rolling off a log erroneous judgement by accident.And its detection threshold is too small, is difficult in the reality choosing.As seen, realize that the little while of a kind of calculation cost guarantees that again the cancelling out echo technology of its constringency performance is very necessary.

Summary of the invention

The deficiency of, constringency performance difference big for the calculation cost of the cancelling out echo DTD algorithm that overcomes existing embedded wireless visual doorbell the invention provides that a kind of calculation cost is little, the echo cancelling method in the embedded wireless visual doorbell that can effectively improve constringency performance.

The technical solution adopted for the present invention to solve the technical problems is:

Echo cancelling method in a kind of embedded wireless visual doorbell, described echo cancelling method may further comprise the steps:

1), initialization: filter coefficient is c (n), and its vector is C (n), is expressed as [c (n) c (n-1) c (n-2) ... c (n-L+1)] T; Vector length L is a filter length, and the iterations M of definite wave filter and adjustment step factor μ, smoothing factor thresholding α, β, echo path mismatch threshold value; 2), obtain n constantly the remote signaling of radio visual doorbell be x (n), its vector is X (n), is expressed as [x (n) x (n-1) x (n-2) ... x (n-L+1)] T; True echo path is h (n), and its vector is H (n), is expressed as [h (n) h (n-1) h (n-2) ... h (n-L+1)] T; Near-end voice signals is z (n), and its vector is Z (n), is expressed as [z (n) z (n-1) z (n-2) ... z (n-L+1)] T; Ground unrest is w (n), and its vector is W (n), is expressed as [w (n) w (n-1) w (n-2) ... w (n-L+1)] T; The near-end microphone input signal is y (n), and its vector is Y (n), is expressed as [y (n) y (n-1) y (n-2) ... y (n-L+1)] T; Error signal is e (n), and wave filter is output as the echo copy signal s (n) that estimates, and near-end microphone input signal y (n) equals ground unrest w (n), near end signal z (n) and echo signal x (n) * h (n) sum;

3), judge whether to take place dual end communication, as not, then directly error signal and output, wherein:

Wave filter output: s (n)=C ^T(n) x (n) (1)

Estimated echo error: e (n)=y (n)-s (n) (3)

In this way, then upgrade adaptive filter coefficient, error signal again, wherein:

Filter coefficient is adjusted: C (n+1)=C (n)+μ e (n) X (n) (2)

Carry out i time interative computation constantly at n, the scope of i is [0～M-1], and iteration is M time altogether.The FIANLMS formula that obtains through derivation is as shown in Equation (4):

$C(n+1)=C\left(n\right)+\frac{\mathrm{\μ}}{X^T\left(n\right)X\left(n\right)}X\left(n\right)\underset{i=0}{\overset{M-1}{\mathrm{\Σ}}}{e}_i\left(n\right)$

$=C\left(n\right)+\frac{\mathrm{\μ}}{X^T\left(n\right)X\left(n\right)}X\left(n\right)\frac{[1-{(1-\mathrm{\μ})}^L]}{2\mathrm{\μ}}e\left(n\right)$

$=C\left(n\right)+\frac{[1-{(1-\mathrm{\μ})}^M]}{X^T\left(n\right)X\left(n\right)}e\left(\text{n}\right)X\left(n\right)---\left(4\right)$

Wherein $e_i\left(n\right)=y\left(n\right)-C_i^T\left(n\right)X\left(n\right);$

4), the albefaction remote signaling, obtain x _w(n), the albefaction error signal obtains e _w(n); The remote signaling after the calculating albefaction and the covariance of error signal, and calculate echo path mismatch value Δ H (n), computing formula is as follows:

$\mathrm{\ΔH}\left(n\right)=R_{\mathrm{xx}}^{-1}{R}_{\mathrm{xe}}=R_{x_w{e}_w}/{\mathrm{\σ}}_{x_w}^2---\left(5\right)$

Wherein ${\mathrm{\σ}}_{x_w}^2\left(n\right)={\mathrm{\α\σ}}_{x_w}^2(n-1)+(1-\mathrm{\α})x_w^2\left(n\right)---\left(6\right)$

$R_{x_e{e}_w}\left(n\right)={\mathrm{\βR}}_{x_w{e}_w}(n-1)+(1-\mathrm{\β})x_w\left(n\right)e_w\left(n\right)---\left(7\right)$

Smoothing factor thresholding α, β are between 0.9～1;

5), judge whether the Δ H (n) calculate surpasses the echo path mismatch threshold value? if surpass, then put the dual end communication sign takes place, and judge whether end process, if not, turn back to step 2), if, end process then; If do not surpass, directly judge whether end process, if not, turn back to step 2), if, end process then.

Technical conceive of the present invention is: its sef-adapting filter adopts a kind of calculation cost little and can improve the algorithm of constringency performance: iteratively faster is adjusted NLMS (FIANLMS) algorithm, its double talk detection adopts a kind of echo path mismatch variance DTD algorithm, and this algorithm can effectively improve the detection performance.The cancelling out echo module is in voice compression module in the whole software system before, after the voice decompression module, shielded the influence of compress speech and decompression.This cancelling out echo module has obtained effect preferably in the embedded wireless visual doorbell system.

Beneficial effect of the present invention mainly shows: save calculation cost, constringency performance that easier acquisition is good and module consistance are better.

Description of drawings

Fig. 1 is the echo neutralizer structural drawing.

Fig. 2 is the hierarchy chart of echo neutralizer software module at system software.

Fig. 3 is the cancelling out echo program flow diagram.

Embodiment

Below in conjunction with accompanying drawing the present invention is further described.

With reference to Fig. 1～Fig. 3, the echo cancelling method in a kind of embedded wireless visual doorbell, described echo cancelling method may further comprise the steps:

1), initialization: filter coefficient is c (n), and its vector is C (n), is expressed as [c (n) c (n-1) c (n-2) ... c (n-L+1)] T; Vector length L is a filter length, and the iterations M of definite wave filter and adjustment step factor μ, smoothing factor thresholding α, β, echo path mismatch threshold value;

2), obtain n constantly the remote signaling of radio visual doorbell be x (n), its vector is X (n), is expressed as [x (n) x (n-1) x (n-2) ... x (n-L+1)] T; True echo path is h (n), and its vector is H (n), is expressed as [h (n) h (n-1) h (n-2) ... h (n-L+1)] T; Near-end voice signals is z (n), and its vector is Z (n), is expressed as [z (n) z (n-1) z (n-2) ... z (n-L+1)] T; Ground unrest is w (n), and its vector is W (n), is expressed as [w (n) w (n-1) w (n-2) ... w (n-L+1)] T; The near-end microphone input signal is y (n), and its vector is Y (n), is expressed as [y (n) y (n-1) y (n-2) ... y (n-L+1)] T; Error signal is e (n), and wave filter is output as the echo copy signal s (n) that estimates, and near-end microphone input signal y (n) equals ground unrest w (n), near end signal z (n) and echo signal x (n) * h (n) sum;

3), judge whether to take place dual end communication, as not, then directly error signal and output, wherein:

Wave filter output: s (n)=C ^T(n) x (n) (1)

Estimated echo error: e (n)=y (n)-s (n) (3)

In this way, then upgrade adaptive filter coefficient, error signal again, wherein:

Filter coefficient is adjusted: C (n+1)=C (n)+μ e (n) X (n) (2)

Carry out i time interative computation constantly at n, the scope of i is [0～M-1], and iteration is M time altogether.The FIANLMS formula that obtains through derivation is as shown in Equation (4):

$C(n+1)=C\left(n\right)+\frac{\mathrm{\μ}}{X^T\left(n\right)X\left(n\right)}X\left(n\right)\underset{i=0}{\overset{M-1}{\mathrm{\Σ}}}{e}_i\left(n\right)$

$=C\left(n\right)+\frac{\mathrm{\μ}}{X^T\left(n\right)X\left(n\right)}X\left(n\right)\frac{[1-{(1-\mathrm{\μ})}^L]}{2\mathrm{\μ}}e\left(n\right)$

$=C\left(n\right)+\frac{[1-{(1-\mathrm{\μ})}^M]}{X^T\left(n\right)X\left(n\right)}e\left(\text{n}\right)X\left(n\right)---\left(4\right)$

Wherein $e_i\left(n\right)=y\left(n\right)-C_i^T\left(n\right)X\left(n\right);$

4), the albefaction remote signaling, obtain x _w(n), the albefaction error signal obtains e _w(n); The remote signaling after the calculating albefaction and the covariance of error signal, and calculate echo path mismatch value Δ H (n), computing formula is as follows:

$\mathrm{\ΔH}\left(n\right)=R_{\mathrm{xx}}^{-1}{R}_{\mathrm{xe}}=R_{x_w{e}_w}/{\mathrm{\σ}}_{x_w}^2---\left(5\right)$

Wherein ${\mathrm{\σ}}_{x_w}^2\left(n\right)={\mathrm{\α\σ}}_{x_w}^2(n-1)+(1-\mathrm{\α})x_w^2\left(n\right)---\left(6\right)$

$R_{x_e{e}_w}\left(n\right)={\mathrm{\βR}}_{x_w{e}_w}(n-1)+(1-\mathrm{\β})x_w\left(n\right)e_w\left(n\right)---\left(7\right)$

Smoothing factor thresholding α, β are between 0.9～1;

5), judge whether the Δ H (n) calculate surpasses the echo path mismatch threshold value? if surpass, then put the dual end communication sign takes place, and judge whether end process, if not, turn back to step 2), if, end process then; If do not surpass, directly judge whether end process, if not, turn back to step 2), if, end process then.

N moment remote signaling is x (n) among Fig. 1, and its vector is x (n) (can be expressed as [x (n) x (n-1) x (n-2) ... x (n-L+1)] T); True echo path is h (n), and its vector is H (n) (can be expressed as [h (n) h (n-1) h (n-2) ... h (n-L+1)] T); Near-end voice signals is z (n), and its vector is Z (n) (can be expressed as [z (n) z (n-1) z (n-2) ... z (n-L+1)] T); Ground unrest is w (n), and its vector is W (n) (can be expressed as [w (n) w (n-1) w (n-2) ... w (n-L+1)] T); The near-end microphone input signal is y (n), and its vector is Y (n) (can be expressed as [y (n) y (n-1) y (n-2) ... y (n-L+1)] T).Error signal is e (n), adaptive filter coefficient is c (n), its vector is C (n) (can be expressed as [c (n) c (n-1) c (n-2) ... c (n-L+1)] T), and wave filter is output as the echo copy signal s (n) that estimates, and above vector length is that L is a filter length.

Wave filter output: s (n)=C ^T(n) x (n) (1)

Filter coefficient is adjusted: C (n+1)=C (n)+μ e (n) X (n) (2)

Wherein μ is called the adjustment step factor, the amplitude that the control adaptive filter coefficient is adjusted.

Estimated echo error: e (n)=y (n)-s (n) (3)

Carry out i time interative computation constantly at n, the scope of i is [0～M-1], and iteration is M time altogether.The FIANLMS formula that obtains through derivation as shown in Equation (4).

$C(n+1)=C\left(n\right)+\frac{\mathrm{\μ}}{X^T\left(n\right)X\left(n\right)}X\left(n\right)\underset{i=0}{\overset{M-1}{\mathrm{\Σ}}}{e}_i\left(n\right)$

$=C\left(n\right)+\frac{\mathrm{\μ}}{X^T\left(n\right)X\left(n\right)}X\left(n\right)\frac{[1-{(1-\mathrm{\μ})}^L]}{2\mathrm{\μ}}e\left(n\right)$

$=C\left(n\right)+\frac{[1-{(1-\mathrm{\μ})}^M]}{X^T\left(n\right)X\left(n\right)}e\left(\text{n}\right)X\left(n\right)---\left(4\right)$

Wherein $e_i\left(n\right)=y\left(n\right)-C_i^T\left(n\right)X\left(n\right).$

Adaptive filter algorithm can guarantee speech quality when having only single-ended conversation, but when both-end is spoken simultaneously since originally convergent e (n) superpose a near-end voice signals altered a great deal like filter divergence, estimation to echo path can produce bigger error like this, therefore need carry out double talk detection, adopt echo path mismatch variance DTD algorithm here.

Echo path mismatch variance DTD algorithm adopts smooth estimated to come approximate treatment on its relevant normalized basis, has avoided square root calculating, reduces by L multiplying at each sampled point, reduces calculation cost greatly.The echo path mismatch variance DTD formula that obtains through derivation as shown in Equation (5).

Path mismatch: $\mathrm{\ΔH}\left(n\right)=R_{\mathrm{xx}}^{-1}{R}_{\mathrm{xe}}=R_{x_w{e}_w}/{\mathrm{\σ}}_{x_w}^2---\left(5\right)$

Wherein ${\mathrm{\σ}}_{x_w}^2\left(n\right)={\mathrm{\α\σ}}_{x_w}^2(n-1)+(1-\mathrm{\α})x_w^2\left(n\right)---\left(6\right)$

$R_{x_e{e}_w}\left(n\right)={\mathrm{\βR}}_{x_w{e}_w}(n-1)+(1-\mathrm{\β})x_w\left(n\right)e_w\left(n\right)---\left(7\right)$

Smoothing factor thresholding α, β between 0.9～1, x _w(n) and e _w(n) be respectively signal after x (n) and e (n) albefaction.

System voice is the worker that enjoys a double blessing, so phonological component indoor and outdoor machine basically identical.As shown in Figure 2, phonetic entry is carried out cancelling out echo through behind the voice collecting, carries out compress speech again, voice transmitting-receiving control, and voice decompress, and carry out cancelling out echo, speech play again.Cancelling out echo was handled after voice decompress before compress speech, like this can be so that the whole software system level is clearly more demarcated, and the module transplantability is better, and cancelling out echo can debug separately, has shielded the influence of compress speech and decompression.

With reference to Fig. 3, the cancelling out echo module of embedded wireless visual doorbell system is carried out initialization to the data structure of parameter, prewhitening filter response and some speech frames earlier.Iterations in the FIANLMS sef-adapting filter (M) is chosen to be 8, because every frame voice are 1024Bytes, is 2Bytes sampling, so filter length (L) is chosen to be 512, and step factor (μ) gets 0.3; Based on smoothing factor thresholding α=0.93 in echo path mismatch double talk detection (DTD) algorithm, β=0.91, echo path mismatch threshold value Δ H _TH=0.25.Obtain near-end speech frame and far-end speech frame then, carry out FIANLMS according to formula (4) again and calculate.Speech frame buffer gets 3 frames, and voice can compare smoothness, i.e. maximum-delay 3 frame times, general 1.2～1.8 frames just can calculate and finish general 70～80ms (16fps).As judge that the then former adaptive filter coefficient of generation dual end communication is inapplicable, need to upgrade.Then according to formula (3) error signal e (n) and output.Calculate the variance of far-end speech x (n) again, then by prewhitening filter albefaction far-end speech and error signal.Calculate far-end speech after the albefaction and the error signal variance after the albefaction according to formula (6) and formula (7) again, thereby calculate echo path mismatch value according to formula (5).As judge that echo path mismatch value surpasses threshold value (Δ H _TH=0.25) then puts generation dual end communication sign.All resources that when voice call finishes initialization obtained discharge.

Claims (1)

1. the echo cancelling method in the embedded wireless visual doorbell, it is characterized in that: described echo cancelling method may further comprise the steps:

1), initialization: filter coefficient is c (n), and its vector is C (n), is expressed as [c (n) c (n-1) c (n-2) ... c (n-L+1)] T; Vector length L is a filter length, and the iterations M of definite wave filter and adjustment step factor μ, smoothing factor thresholding α, β, echo path mismatch threshold value;

2), obtain n constantly the remote signaling of radio visual doorbell be x (n), its vector is X (n), is expressed as [x (n) x (n-1) x (n-2) ... x (n-L+1)] T; True echo path is h (n), and its vector is H (n), is expressed as [h (n) h (n-1) h (n-2) ... h (n-L+1)] T; Near-end voice signals is z (n), and its vector is Z (n), is expressed as [z (n) z (n-1) z (n-2) ... z (n-L+1)] T; Ground unrest is w (n), and its vector is W (n), is expressed as [w (n) w (n-1) w (n-2) ... w (n-L+1)] T; The near-end microphone input signal is y (n), and its vector is Y (n), is expressed as [y (n) y (n-1) y (n-2) ... y (n-L+1)] T; Error signal is e (n), and wave filter is output as the echo copy signal s (n) that estimates, and near-end microphone input signal y (n) equals ground unrest w (n), near end signal z (n) and echo signal x (n) * h (n) sum;

3), judge whether to take place dual end communication, as not, then directly error signal and output, wherein:

Wave filter output: s (n)=C ^T(n) x (n) (1)

Estimated echo error: e (n)=y (n)-s (n) (3)

In this way, then upgrade adaptive filter coefficient, error signal again, wherein:

Filter coefficient is adjusted: C (n+1)=C (n)+μ e (n) X (n) (2)

Carry out i time interative computation constantly at n, the scope of i is [0～M-1], and iteration is M time altogether.The FIANLMS formula that obtains through derivation is as shown in Equation (4):

$C(n+1)=C\left(n\right)+\frac{\mathrm{\μ}}{X^T\left(n\right)X\left(n\right)}X\left(n\right)\underset{i=0}{\overset{M-1}{\mathrm{\Σ}}}{e}_i\left(n\right)$

$=C\left(n\right)+\frac{\mathrm{\μ}}{X^T\left(n\right)X\left(n\right)}X\left(n\right)\frac{[1-{(1-\mathrm{\μ})}^L]}{2\mathrm{\μ}}e\left(n\right)$

$=C\left(n\right)+\frac{[1+{(1-\mathrm{\μ})}^M]}{X^T\left(n\right)X\left(n\right)}e\left(n\right)X\left(n\right)---\left(4\right)$

Wherein $e_i\left(n\right)=y\left(n\right)-C_i^T\left(n\right)X\left(n\right);$

4), the albefaction remote signaling, obtain x _w(n), the albefaction error signal obtains e _w(n); The remote signaling after the calculating albefaction and the covariance of error signal, and calculate echo path mismatch value Δ H (n),

Computing formula is as follows:

$\mathrm{\ΔH}\left(n\right)=R_{\mathrm{xx}}^{-1}{R}_{\mathrm{xe}}=R_{x_w{e}_w}/{\mathrm{\σ}}_{x_w}^2---\left(5\right)$

Wherein ${\mathrm{\σ}}_{x_w}^2\left(n\right)=\mathrm{\α}{\mathrm{\σ}}_{x_w}^2(n-1)+(1-\mathrm{\α})x_w^2\left(n\right)---\left(6\right)$

$R_{x_w{e}_w}\left(n\right)=\mathrm{\β}{R}_{x_w{e}_w}(n-1)+(1-\mathrm{\β})x_w\left(n\right)e_w\left(n\right)---\left(7\right)$

Smoothing factor thresholding α, β are between 0.9～1;

5), judge whether the Δ H (n) calculate surpasses the echo path mismatch threshold value? if surpass, then put the dual end communication sign takes place, and judge whether end process, if not, turn back to step

2), if, end process then; If do not surpass, directly judge whether end process, if not, turn back to step 2), if, end process then.

Images