CN104835503A - Improved GSC self-adaptive speech enhancement method - Google Patents

Abstract

The invention relates to an improved GSC self-adaptive speech enhancement method. An improved GSC self-adaptive speech enhancement system is used, a weight coefficient of a post filter is estimated based on the signal received by a microphone, incoherent noises in the signal are removed by using the post wiener filter, a self-adaptive blocking matrix is used for replacing a fixed blocking matrix in a conventional GSC structure so as to better block target signals and reduce cancellation of the target signals, the iterative mode of a self-adaptive algorithm is improved, the convergence rate and steady-state misadjustment signals and filter noise interference signals are balanced, and based on the characteristics of easy noise removal, small amount of calculation and the like of a spectral subtraction method, the possible residual noise signals can be further removed by using an improved spectral subtraction method. According to the invention, the method is capable of self-adaptive speech enhancement of signals at any angle and has a certain robustness for speech signal enhancement under a random environment, and the subsequent spectral subtraction method can be used for further removing the residual noises and improving the denoising capability of the entire system.

Description

A kind of improvement GSC adaptive voice Enhancement Method

Technical field

The invention belongs to signal transacting field, particularly Array Signal Processing, auto adapted filtering and speech enhan-cement.

Technical background

Voice are that human information is propagated and the important carrier of emotion communication, are also that the most Chong of human information interchange Yaos the mode of the most You Xiao ﹑ most convenient of ﹑.In the voice communication of reality, voice are always inevitably subject to the interference of extraneous environmental noise, comprise the interference of noise that transmission medium introduces, inside communication equipment electrical noise and even other talker.These disturb voice signal the primary speech signal of impure that microphones is arrived, but are subject to the Noisy Speech Signal of noise pollution, cause many speech processing system performances sharply to worsen.Therefore, in order to obtain primary speech signal pure as far as possible from Noisy Speech Signal, just need to carry out speech enhan-cement.Compare with single microphone, the basis that microphone array is listed in time-frequency domain adds a spatial domain, makes full use of the spatial domain of voice signal, time domain and frequency domain information, there is the feature such as high spatial resolution, high signal gain and stronger antijamming capability simultaneously.It can make up the deficiency that single microphone exists in noise processed, voice extraction and isolation etc.

Microphone array technical foundation is introduced auto adapted filtering and can control beam direction neatly, real-time follow-up targeted voice signal.Auto adapted filtering only needs little or basic without any need for the priori about signal and noise statistics, only just can the statistical property of estimated signal and noise in real time according to observation information.In adaptive filter algorithm, least mean square algorithm (LMS) is the most typical, although comparatively recurrent least square method (RLS) speed of convergence is slow for this algorithm, this algorithm is simple, is convenient to real-time implementation.Filtered voice signal also can be difficult to the noise of removal with some ground unrests or other, these noises can utilize spectrum-subtraction to remove further.Spectrum-subtraction is as a kind of conventional sound enhancement method, and because it is simple and convenient, the advantages such as calculated amount is little make it be popular in single-channel voice strengthens, but can produce ear-piercing " music " noise after wherethrough reason, and more difficult elimination.

Summary of the invention

The present invention is based on traditional Microphone Array Speech enhancing technology, propose a kind of adaptive voice Enhancement Method of novel broad sense Sidelobe Canceller.The method utilizes self-adaptation blocking matrix cascade adaptive Canceller to estimate the residual noise in S filter output signal, and output signal with S filter and offset, then a follow-up spectrum-subtraction removes the residual noise that may leave further, thus reaches the object strengthening echo signal and restraint speckle undesired signal.

The improvement GSC adaptive voice that the present invention adopts strengthens system chart (as shown in Figure 1), according to microphones to Signal estimation obtain the weight coefficient of postfilter, then rearmounted S filter is utilized to remove noncoherent noise in signal, utilize the fixing blocking matrix that self-adaptation blocking matrix replaces in traditional GSC structure, better obstruction falls echo signal, reduce echo signal to offset, improve the iterative manner of adaptive algorithm in adaptive cancellation device, speed of convergence and steady output rate signal are taken into account, maximum filtering noise undesired signal, noise is removed simple and convenient according to spectrum-subtraction, the characteristics such as calculated amount is little, the present invention utilizes the spectrum-subtraction of improvement to remove the remaining noise signal of possibility further.

The present invention for achieving the above object, adopts following technical scheme:

A kind of improvement GSC adaptive voice Enhancement Method, concrete steps are as follows:

Step one: Delay Estima-tion and compensation are carried out to Noisy Speech Signal X (t) that microphone array receives, the signal in each microphone channel is made to be consistent in time, wherein X (t)=A (θ) S (t)+N (t), θ is the arrival bearing of echo signal, the array manifold that A (θ) is echo signal, S (t) is targeted voice signal, and N (t) is directional interference noise or random noise;

Step 2: X (t) now processes respectively through upper and lower branch road, upper branch road, through fixed beam former, playing and preliminary suppress interfering noise signal and strengthen the effect of echo signal, exporting as having certain target voice reference signal y strengthened _c(n), wherein y _c(n)=A ^tx (n), A are fixed weighting coefficient, A=[a ₀, a ₁..., a _(M-1)] ^t, A meets

Step 3: utilize rearmounted S filter to remove y further _cnoncoherent noise remaining in (n), the voice signal y after being enhanced _a(n);

Step 4: X (t) also carries out relevant treatment through lower branch road while upper branch road process, the self-adaptation blocking matrix of lower branch road can well tracking target signal, and self-adaptation obstruction is carried out to it, make the echo signal of leakage reach minimum or do not have completely;

Step 5: the automatic adaptation FIR Canceller of improvement estimates the estimated value of noise signal in upper branch output signal according to U (n), noise signal estimated value is:

$y_1\left(n\right)=\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{w}_k^T\left(n\right)u_k\left(n\right)=W_K^T\left(n\right)U\left(n\right)$

y ₂(n)＝y _a(n)-y ₁(n)

$W_K(n+1)=W_K\left(n\right)+\frac{y_2\left(n\right)}{{\left|\right|U\left(n\right)\left|\right|}^2}U\left(n\right)$

Wherein, W _k=[w ₁(n), w ₂(n) ... w _k(n)] ^t, U (n)=[u ₁(n), u ₂(n) ... u _k(n)] ^t, y ₂n () is for obtaining comparatively pure voice signal after process, improvement adaptive cancellation device is the same with adaptive algorithm weight coefficient update mode in self-adaptation blocking matrix has all taken into account the steady output rate of renewal speed and signal, in two large modules, adaptive application enhances the stability of whole system process random signal, and the system that simultaneously too increases does not mate with actual arrival direction the robustness bringing error to the anticipated orientation of echo signal;

Step 6: in order to remove the residual noise that may exist further, at y ₂n () be continued access spectrum-subtraction afterwards.

Wherein, rearmounted S filter is utilized to remove y further in step 3 _cnoncoherent noise remaining in (n), the voice signal y after being enhanced _an the disposal route of () is: the autocorrelation spectrum S estimating Received signal strength and echo signal according to X (t) _xx(k) and S _ssk () also obtains filter coefficient W _o(k), k=0,1,2 ... K-1, wherein, k is frequency domain sequence number, and K is the length of Short Time Fourier Transform, uses W _ok () is to y _cn () carries out frequency domain filtering, obtain y _acarry out after (k) inverse Fourier transform be enhanced after voice signal y _a(n).

Wherein, the method for carrying out self-adaptation obstruction in step 4 is: m channel adaptive algorithm is:

$y_{\mathrm{am}}\left(n\right)=\underset{i=0}{\overset{N-1}{\mathrm{\Σ}}}{w}_{\mathrm{mm}}\left(i\right)y_a(n-i)={W_{\mathrm{mm}}}^T\left(n\right)Y_a\left(n\right)$

u _m(n)＝x _m(n)-y _am(n)

$W_{\mathrm{mm}}(n+1)=W_{\mathrm{mm}}\left(n\right)+\frac{u_m\left(n\right)}{{\left|\right|Y_a\left(n\right)\left|\right|}^2}{Y}_a\left(n\right)$

Wherein, m=1,2 ... M, M are microphone number, and N is wave filter betweenness, u _m(n) for respective channel offset after signal, mostly be the combination of Noise and Interference signal.

Further, to y in step 6 ₂n () afterwards continued access spectrum-subtraction improves: spectrum-subtraction is on the basis of short-term stationarity supposition, Fourier transform is carried out to Noisy Speech Signal and carries out overlapping sub-frame processing, the noise power estimating to obtain is deducted with every frame signal power, and utilize the Amplitude Ratio of people's ear to voice signal more responsive, and this characteristic of phase-unsensitive to voice, can by the phase place that replace clean speech signal containing the phase place of noisy voice signal, then to its carry out inverse Fourier transform can be enhanced after voice signal.When utilizing spectrum-subtraction to carry out denoising to Noisy Speech Signal, the present invention utilizes low-pass first order filter to carry out the power of estimating noise:

${\left|D_{i,w}\left(\mathrm{\ω}\right)\right|}^2=\left\{\begin{array}{cc}\mathrm{\σ}{\left|Y_{i,w}\left(\mathrm{\ω}\right)\right|}^2& i=1\\ {\left|D_{i-1,w}\left(\mathrm{\ω}\right)\right|}^2+(1-\mathrm{\σ}){\left|Y_{i,w}\left(\mathrm{\ω}\right)\right|}^2& 2\≤i\≤K\end{array}\right.$

Wherein 0< σ <1, K are noisy speech totalframes.The every frame clean speech power obtained is:

Wherein, α >1, β <<1, retains in noise segment the effect that certain noise can obtain good noise reduction and suppression pure tone noise, reduces the generation of " music " noise, improves auditory effect.

Beneficial effect:

The microphone array adaptive voice Enhancement Method that the present invention proposes, on the basis of traditional generalized sidelobe canceller structure, introduce rearmounted S filter, self-adaptation blocking matrix, improvement adaptive cancellation device and improve spectrum-subtraction, compared to traditional GSC, the inventive method can carry out adaptive voice enhancing for signal at any angle, have certain robustness to the voice signals enhancement under random environment, and follow-up spectrum-subtraction can remove residual noise raising entire system noise removal capability further.

Accompanying drawing explanation

The speech-enhancement system structured flowchart that Fig. 1, the present invention adopt;

Fig. 2, traditional generalized sidelobe canceller (GSC) system chart;

Fig. 3, fixed beam former system chart;

Fig. 4, rearmounted S filter system chart;

Fig. 5, spectrum-subtraction system chart;

Fig. 6, clean speech signal (orientation angle is 10 degree);

Fig. 7, directional interference signal (orientation angle is 30 degree);

The voice signal (Noisy Speech Signal) that Fig. 8, microphone array receive;

The voice signal obtained after Fig. 9, traditional GSC process;

The voice signal obtained after Figure 10, the inventive method process;

The beam pattern obtained after Figure 11, the inventive method process;

The directional diagram obtained after Figure 12, the inventive method process.

Embodiment

For making the object, technical solutions and advantages of the present invention clearly, below in conjunction with the drawings and specific embodiments, illustrate the present invention further.One provided by the invention improves GSC adaptive voice Enhancement Method, as shown in Figure 1, specifically comprises following operation steps:

Step one: Delay Estima-tion and compensation are carried out to Noisy Speech Signal X (t)=A (θ) S (the t)+N (t) that microphone array receives, the signal in each microphone channel is made to be consistent in time, the x in Fig. 1 ₁, x ₂, x ₃... .x _nfor the Noisy Speech Signal after alignment, wherein, θ is the arrival bearing of echo signal, the array manifold that A (θ) is echo signal, and S (t) is targeted voice signal, and N (t) is directional interference noise or random noise;

Step 2: X (t) now processes respectively through upper and lower branch road, upper branch road is through fixed beam former (concrete structure as shown in Figure 3), play and preliminary suppress interfering noise signal and strengthen the effect of echo signal, export as there is certain target voice reference signal strengthened, y _c(n)=A ^tx (n), A are fixed weighting coefficient, A=[a ₀, a ₁..., a _(M-1)] ^t, A meets

Step 3: utilize rearmounted S filter (concrete structure as shown in Figure 4) to remove y further _cn noncoherent noise remaining in (), obtains y _a(n).Disposal route is for estimating the autocorrelation spectrum S of Received signal strength and echo signal according to X (t) _xx(k) and S _ssk () also obtains filter coefficient W _o(k), k=0,1,2 ... K-1, wherein, k is frequency domain sequence number, and K is the length of Short Time Fourier Transform, uses W _ok () is to y _cn () carries out frequency domain filtering, obtain y _acarry out after (k) inverse Fourier transform be enhanced after voice signal y _a(n).

Step 4: X (t) also carries out relevant treatment through lower branch road while upper branch road process, the self-adaptation blocking matrix of lower branch road can well tracking target signal, and self-adaptation obstruction is carried out to it, make the echo signal of leakage reach minimum or do not have completely, m (m=1,2 ... M, M are microphone number) individual channel adaptive algorithm is:

$y_{\mathrm{am}}\left(n\right)=\underset{i=0}{\overset{N-1}{\mathrm{\&Sigma;}}}{w}_{\mathrm{mm}}\left(i\right)y_a(n-i)={W_{\mathrm{mm}}}^T\left(n\right)Y_a\left(n\right)$

u _m(n)＝x _m(n)-y _am(n)

$W_{\mathrm{mm}}(n+1)=W_{\mathrm{mm}}\left(n\right)+\frac{u_m\left(n\right)}{{\left|\right|Y_a\left(n\right)\left|\right|}^2}{Y}_a\left(n\right)$

Wherein, N is wave filter betweenness, u _m(n) for respective channel offset after signal, mostly be the combination of Noise and Interference signal.

Step 5: the automatic adaptation FIR Canceller of improvement estimates the estimated value of noise signal in upper branch output signal according to U (n), noise signal estimated value is:

$y_1\left(n\right)=\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}{w}_k^T\left(n\right)u_k\left(n\right)=W_K^T\left(n\right)U\left(n\right)$

y ₂(n)＝y _a(n)-y ₁(n)

$W_K(n+1)=W_K\left(n\right)+\frac{y_2\left(n\right)}{{\left|\right|U\left(n\right)\left|\right|}^2}U\left(n\right)$

Wherein, W _k=[w ₁(n), w ₂(n) ... w _k(n)] ^t, U (n)=[u ₁(n), u ₂(n) ... u _k(n)] ^t.

Y ₂n () is for obtaining comparatively pure voice signal after process, improvement adaptive cancellation device is the same with adaptive algorithm weight coefficient update mode in self-adaptation blocking matrix has all taken into account the steady output rate of renewal speed and signal, in two large modules, adaptive application enhances the stability of whole system process random signal, and the system that simultaneously too increases does not mate with actual arrival direction the robustness bringing error to the anticipated orientation of echo signal.

Step 6: in order to remove the residual noise that may exist further, at y ₂n () be continued access spectrum-subtraction (specific implementation block diagram as shown in Figure 5) afterwards.Spectrum-subtraction is on the basis of short-term stationarity supposition, Fourier transform is carried out to Noisy Speech Signal and carries out overlapping sub-frame processing, the noise power estimating to obtain is deducted with every frame signal power, and utilize the Amplitude Ratio of people's ear to voice signal more responsive, and this characteristic of phase-unsensitive to voice, can by the phase place that replace clean speech signal containing the phase place of noisy voice signal, then to its carry out inverse Fourier transform can be enhanced after voice signal.When utilizing spectrum-subtraction to carry out denoising to Noisy Speech Signal, the present invention utilizes low-pass first order filter to carry out the power of estimating noise:

${\left|D_{i,w}\left(\mathrm{\&omega;}\right)\right|}^2=\left\{\begin{array}{cc}\mathrm{\&sigma;}{\left|Y_{i,w}\left(\mathrm{\&omega;}\right)\right|}^2& i=1\\ {\left|D_{i-1,w}\left(\mathrm{\&omega;}\right)\right|}^2+(1-\mathrm{\&sigma;}){\left|Y_{i,w}\left(\mathrm{\&omega;}\right)\right|}^2& 2\&le;i\&le;K\end{array}\right.$

Wherein 0< σ <1, K are noisy speech totalframes.The every frame clean speech power obtained is:

Wherein, α >1, β <<1.Retain in noise segment the effect that certain noise can obtain good noise reduction and suppression pure tone noise, reduce the generation of " music " noise, improve auditory effect.

Experimental result as shown in figs 6-12, the voice signal that microphone array receives is that Noisy Speech Signal can draw from Fig. 6 and Fig. 8 contrast, Fig. 9 is the voice signal contrast and Figure 10 (the inventive method result figure) that obtain after traditional GSC structure (concrete structure block diagram as shown in Figure 2) process, obviously can see that the inventive method removes the better effects if of noise, Figure 11, Figure 12 is the directional diagram that obtains of the inventive method process Noisy Speech Signal and beam pattern respectively, can see that the echo signal in 10 degree of directions is well accepted and directional interference noise (arrival bearing is 30 degree) obtains very large suppression.Comprehensively obtain the inventive method better to Noisy Speech Signal treatment effect compared to traditional GSC, and have certain adaptability to random environment, range of application is wider.

The above; it is only preferred embodiment of the present invention; not any pro forma restriction is done to the present invention; any those skilled in the art; do not departing within the scope of technical solution of the present invention; according to technical spirit of the present invention, any simple amendment that above embodiment is done, equivalently replace and improve, within the protection domain all still belonging to technical solution of the present invention.

Claims (4)

1. improve a GSC adaptive voice Enhancement Method, it is characterized in that: concrete steps are as follows:

Step one: Delay Estima-tion and compensation are carried out to Noisy Speech Signal X (t) that microphone array receives, the signal in each microphone channel is made to be consistent in time, wherein X (t)=A (θ) S (t)+N (t), θ is the arrival bearing of echo signal, the array manifold that A (θ) is echo signal, S (t) is targeted voice signal, and N (t) is directional interference noise or random noise;

Step 2: X (t) now processes respectively through upper and lower branch road, upper branch road, through fixed beam former, playing and preliminary suppress interfering noise signal and strengthen the effect of echo signal, exporting as having certain target voice reference signal y strengthened _c(n), wherein y _c(n)=A ^tx (n), A are fixed weighting coefficient, A=[a ₀, a ₁..., a _(M-1)] ^t, A meets

Step 3: utilize rearmounted S filter to remove y further _cnoncoherent noise remaining in (n), the voice signal y after being enhanced _a(n);

Step 4: X (t) also carries out relevant treatment through lower branch road while upper branch road process, the self-adaptation blocking matrix of lower branch road can well tracking target signal, and self-adaptation obstruction is carried out to it, make the echo signal of leakage reach minimum or do not have completely;

Step 5: the automatic adaptation FIR Canceller of improvement estimates the estimated value of noise signal in upper branch output signal according to U (n), noise signal estimated value is:

$y_1\left(n\right)=\underset{k=1}{\overset{K}{\mathrm{\&Sigma;}}}{w}_k^T\left(n\right)u_k\left(n\right)=W_K^T\left(n\right)U\left(n\right)$

y ₂(n)＝y _a(n)-y ₁(n)

$W_K(n+1)=W_K\left(n\right)+\frac{y_2\left(n\right)}{{\left|\right|U\left(n\right)\left|\right|}^2}U\left(n\right)$

Wherein, W _k=[w ₁(n), w ₂(n) ... w _k(n)] ^t, U (n)=[u ₁(n), u ₂(n) ... u _k(n)] ^t, y ₂n () is for obtaining comparatively pure voice signal after process, improvement adaptive cancellation device is the same with adaptive algorithm weight coefficient update mode in self-adaptation blocking matrix has all taken into account the steady output rate of renewal speed and signal, in two large modules, adaptive application enhances the stability of whole system process random signal, and the system that simultaneously too increases does not mate with actual arrival direction the robustness bringing error to the anticipated orientation of echo signal;

Step 6: in order to remove the residual noise that may exist further, at y ₂n () be continued access spectrum-subtraction afterwards.

2. improvement GSC adaptive voice Enhancement Method according to claim 1, is characterized in that: utilize rearmounted S filter to remove y further in step 3 _cnoncoherent noise remaining in (n), the voice signal y after being enhanced _an the disposal route of () is: the autocorrelation spectrum S estimating Received signal strength and echo signal according to X (t) _xx(k) and S _ssk () also obtains filter coefficient W _o(k), k=0,1,2 ... K-1, wherein, k is frequency domain sequence number, and K is the length of Short Time Fourier Transform, uses W _ok () is to y _cn () carries out frequency domain filtering, obtain y _acarry out after (k) inverse Fourier transform be enhanced after voice signal y _a(n).

3. improvement GSC adaptive voice Enhancement Method according to claim 1, is characterized in that: the method for carrying out self-adaptation obstruction in step 4 is: m channel adaptive algorithm is:

$y_{\mathrm{am}}\left(n\right)=\underset{i=0}{\overset{N-1}{\mathrm{\&Sigma;}}}{w}_{\mathrm{mm}}\left(i\right)y_a(n-i)={W_{\mathrm{mm}}}^T\left(n\right)Y_a\left(n\right)$

u _m(n)＝x _m(n)-y _am(n)

$W_{\mathrm{mm}}(n+1)=W_{\mathrm{mm}}\left(n\right)+\frac{u_m\left(n\right)}{{\left|\right|Y_a\left(n\right)\left|\right|}^2}{Y}_a\left(n\right)$

Wherein, m=1,2 ... M, M are microphone number, and N is wave filter betweenness, u _m(n) for respective channel offset after signal, mostly be the combination of Noise and Interference signal.

4. improvement GSC adaptive voice Enhancement Method according to claim 1, is characterized in that: to y in step 6 ₂n () afterwards continued access spectrum-subtraction improves: spectrum-subtraction is on the basis of short-term stationarity supposition, Fourier transform is carried out to Noisy Speech Signal and carries out overlapping sub-frame processing, the noise power estimating to obtain is deducted with every frame signal power, and utilize the Amplitude Ratio of people's ear to voice signal more responsive, and this characteristic of phase-unsensitive to voice, the phase place of clean speech signal can be replaced by the phase place containing noisy voice signal, then to its carry out inverse Fourier transform can be enhanced after voice signal, when utilizing spectrum-subtraction to carry out denoising to Noisy Speech Signal, the present invention utilizes low-pass first order filter to carry out the power of estimating noise:

${\left|D_{i,w}\left(\mathrm{\&omega;}\right)\right|}^2=\left\{\begin{array}{cc}\mathrm{\&sigma;}{\left|Y_{i,w}\left(\mathrm{\&omega;}\right)\right|}^2& i=1\\ {\left|D_{i-1,w}\left(\mathrm{\&omega;}\right)\right|}^2+(1-\mathrm{\&sigma;}){\left|Y_{i,w}\left(\mathrm{\&omega;}\right)\right|}^{\text{2}}& 2\&le;i\&le;K\end{array}\right.$

Wherein 0< σ <1, K are noisy speech totalframes, and the every frame clean speech power obtained is:

Wherein, α >1, β <<1, retains in noise segment the effect that certain noise can obtain good noise reduction and suppression pure tone noise, reduces the generation of " music " noise, improves auditory effect.