Nowadays, the using of microphone array (MA) technology has been common installed in almost acoustic equipment, such as smart phone, voice-controlled device, teleconference system, surveillance equipment, cochlear implant, hearing aids for processing the noisy mixture and enhancing the speech quality, the speech intelligibility and perceptual metric listener. Generalized sidelobe canceller (GSC) is one the most effective beamforming method for extracting the desired target speaker while suppressing the interference, third-party talker and other signal from uncertain direction without speech distortion. However, because of the complex noisy environment, the inaccurate estimation of preferred steering vector, the displacement of MA geometry, the different microphone quality, the error of sampling rate or the rapidly changed environmental factor, the GSC beamformer's performance often corrupted. The existence of speech distortion or musical noise decreases the perceptual metric listener, the speech intelligibility. In this contribution, the author proposed efficient post-filtering for eliminating the noise level with an acceptable satisfied speech quality to enhance the GSC beamformer's output signal. The conducted simulation shows the effectiveness of removing the noise level to 11.7 dB and improving the signal-to-noise (SNR) ratio from 12.5 to 13.4 dB. The author's proposed post-filtering can deal other complicated problems.
Speech enhancement plays an important role in numerous acoustic speech applications, such as, teleconference system, hearing aids, voice - controlled devices, smart-phone, cochlear implant for extracting the desired target speaker while attenuating the different incoming signal from other signals. Nowadays, the utilizing of microphone array technology has been popular, due to its convenience of removing noise and enhancing speech at the same time. MA beamforming often uses the spatial information of preferred steering vector, the designed configuration of MA geometry, the properties of surrounding environment to obtain high directional beampattern towards the sound source. The significant benefit of using MA beamforming is the ability of incorporating single - channel algorithm, pre - processing, post - Filtering for enhancing the digital signal processing system’s evaluation.
Microphone array can be categorized into two groups: the fixed beamformer and adaptive
beam-former. Fixed beamformer often uses the prior spatial diversity to generate constant
beamformer’s coefficients. Delay and sum (DAS) [
GSC beamformer is one of the most useful beamforming, which concerns the steerable beampattern at specified direction and use an adaptive noise canceller to obtain the clean speech data. GSC beamformer outperforms in coherent - diff use noise fi eld and owns the easy computing. Unfortunately, in realistic recording environment, the error of sampling, the displacement of MA, the inaccurate estimation of environmental properties, the error of sampling rate, the different sensitivities of sensors usually decrease the GSC beamformer’s performance. There is numerous research and work attempt dealing this drawback.
Wang J et al [
The numerical results has confirmed the robustness of GSC beamformer in various types of recording environment.
Kim S et al [
In [
In [
Zohourian M et al [
Priyanka S et al [
Jiang Q et al [
However, these works cannot address the realistic recording scenario, because of the undetermined reasons, the imperfect estimation of necessary parameter. In this contribution, the author proposed an efficient post - Filtering for overcoming this drawback for further increasing the speech enhancement and noise reduction. The suggested approach based on the priori information of direction of arrival of interest useful signal to obtain the post - Filtering.
The rest of this contribution is organized as following ways: Section II describes the GSC’s structure in frequency domain, the author’s suggested post - Filtering is presented in section III with using the prior direction of arrival and section IV illustrates the conducted experiment in reallife recording environment, section V concludes the work.
Consider a microphone array comprising microphones, which captures a target desired sound at
specified direction in a noisy reverberant recording scenario. The representation of received array
signals y ( k , n)=[Y 1 ( k , n) … Y M ( k , n)]T can be described in the short-time Fourier transform
(STFT) domain as [
y ( k , n)= x ( k , n)+ v ( k , n) ¿ gr ( k , n) X r ( k , n)+ v ( k , n) where k is the frequency index, n is the current frame, the superscript ❑T denotes the transpose operator, v ( k , n) means the noise signal and x ( k , n) is the target speech component, X r ( k , n) is the clean speech signal at the selected reference microphone, respectively.
The steering vector gr ( k , n) with respect to the r −th microphone can be calculated as the
following equation [
G1 ( k , n) Gr ( k , n) , … ,
GM ( k . n) Gr ( k , n)
T ] (1) (2) (3) (4) (5) where Gi ( k , n) denote the acoustic transfer function from the sound source to the t -th microphone. We can define that gr ( k , n) ≈ gr ( n), based on assumption that the environment is slowly time - varying.
Note that each frequency bin is treated independently, and the author will omit the frequency index k for brevity. The necessary problem of speech enhancement is finding an optimum coefficient obtaining the processed signal, which approximately the original speech component. We recover the desired speaker by the GSC beamformer as:
where W GSC ( n)=[W 1 ( n) , … , W M ( n)]T is a filter length M . The traditional GSC beamforming technique retains the desired talker undistorted while suppressing the background noise, interference and surrounding noise. Therefore, the formulation of W GSC ( n) can be expressed as: ^X ( n)=W GHSC ( n) y ( n)
W GSC ( n)=W q−B W a ( n)
steerable beampattern at specified direction and form a main signal Y s ( n)=W qH ( n) y ( n) for the
gr
desired speech component. A matched filter W q ( n)=
grH gr
robustness under various types of noisy conditions [
The Adaptive Interference Canceller often deploys un constrained adaptive filter W a ( n) to block the residual noise in Y s ( n) with the reference signal u ( n).
In practical recording situation, the robust speech enhancement of GSC beamformer is indispensable due to the microphone mismatches, the displacement of MA geometry, the different microphone quality, the inaccurate estimation of preferred steering vector, the error of sampling rate or the blocking matrix imperfect block the target speech signal, the reference signal contains target speech leakage. Consequently, the remained noise component at GSC beamformer’s output signal or speech distortion seriously degrade the performance of signal processing system.
In this section, the author proposed using an efficient post - Filtering, which based on the prior spatial information of preferred direction of interest useful target talker, the phase difference of two mounted microphone signals for achieving the necessary formulation.
The author’s post - Filtering is given by: thoPF ( n)=
σ 2x ( n) σ 2x ( n)+σ 2v ( n) where σ x ( n), σ v ( n) is the variance of speech and noise, respectively.
Based on prior steering vector gr ( n) and observed covariance matrix of received array signals, the covariance of speech component and noise can be yield as: (6) (7) (8) And σ 2x ( n)=
1 grH ( n) Φ−yy1 ( n) gr ( n) σ 2v ( n)=W GHSC ( n) Φvv ( n) W GSC ( n) where Φ yy ( n), Φvv ( n) is covariance matrix of observed microphone array signals and noise. Φ yy ( n)= E { y H ( n) y ( n)} based on the received array signals, and Φvv ( n) is computed at the frame, in which exists the only noise - frame.
In numerous practical recording situation, the information about the noise is not always available, and calculation of Φvv ( n) still challenge task in almost acoustic equipment.
The author suggested exploiting the spectral masking
GBM ( n)=1−exp ( j ϕinjorm) and
GDS ( n)=0.5+0.5∗exp ( j ϕinjorm) [
The covariance matrix of noise can be determined as:
The advantage of the author’s proposed approach is utilizing the designed MA configuration, the impinging angle of interest speaker relative to the MA to compute post - Filtering for further removing the musical noise, residual noise and increasing the speech quality.
And the enhanced signal can be derived as: Φvv ( n)=
GBM ( n) GDSB ( n)+GBM ( n)
× Φ yy ( n) ^ X thoPF ( n)= ^X ( n) × thoPF ( n) (9) (10)
In the next section, the author illustrates experiment to verify the effectiveness of post Filtering.
In this section, the author illustrates an experiment to verify the advantage of proposed post
Filtering (adpF) in comparison with the conventional GSC beamformer (cGSCbe). An objective
measurement [
For capturing the original microphone array signals, these parameters were used: frequency sampling Fs=16 kHz, Hanning window, nFFT =512, overlap 50 %. The received array signals can be shown in Figure 5 and Figure 6.
With suitable smoothing parameter α =0.1, the GSC beamformer’s output signal can be derived:
Because of the complex and annoying recording environment, the microphone mismatches, the moving head of speaker, the inaccurate estimation of steering vector, GSC beamformer’s performance often corrupted. The musical noise, residual noise still be a challenging problem. In Figure 8, the musical noise seriously affects the output signal. Musical noise occurs due to the heterogeneous environment, microphone mismatches or error of sampling rate at high-frequency band. Therefore, the necessary of post - Filtering to overcome this drawback is an essential core in almost acoustic problems. The efficient author’s suggested technique suppresses musical noise has been illustrated in Figure 9 and Figure 10.
The effectiveness of the author’s proposed method in reducing musical noise, residual background noise to 11.7 dB and improving the perceptual metric listener, speech intelligibility, speech quality from 12.5 to 13.4 dB. The above post - Filtering not only recover the speech component but also removes musical noise, background noise. The advantage of the author’s post Filtering is using the accurate estimation of variance of speech and variance of noise, which was calculated by applying spectral gain. The author’s approach is utilizing the prior information of impinging incident angle of direction of arrival of interest signal to form an additive post - Filtering for alleviating the musical noise, background noise. The presented post - Filtering owns the characteristic of rapid changed environmental factors and exactly computes variance of speech and noise component according to the considered frame. This approach ensures preserving the original clean speech date while suppressing background noise. In the future, the author will continue investigating the described above method for multi-channel processing system.
GSC beamformer is commonly installed in almost acoustic equipment, due to its high spatial diversity, high directivity index and easy implementation. In realistic recording environment, due to the microphone mismatches, the different microphone sensitivities and other undetermined reasons, GSC beamformer’s evaluation usually degrades. The existence of musical noise or unacceptable residual noise corrupts the speech quality of processed signal. In this paper, the author illustrates an efficient post - Filtering to suppress the remained musical noise, which occurs at GSC beamformer’s output due to the complex and adverse recording scenario. Consequently, the effectiveness of the author’s suggested technique is confirmed by decreasing the background noise to 11.7 dB and increasing the speech quality in the term of signal - to - noise ratio from 12.5 to 13.4 dB. The numerical results has confirmed the effectiveness of the author’s post - Filtering in reducing noise component and improving the speech quality, perceptual metric listener and speech intelligibility. The numerical simulation has confirmed the effectiveness of the author’s suggested method in improving the robust speech enhancement. The above proposed approach can be integrated into multi-channel system.
The authors have not employed any Generative AI tools.