Reduced-bandwidth and Reduced-bandwidth and
distributed MWF-based noise distributed MWF-based noise
reduction algorithms reduction algorithms
Simon Doclo, Tim Van den Bogaert, Jan Wouters, Marc Moonen
Dept. of Electrical Engineering (ESAT-SCD), KU Leuven, Belgium Laboratory for Exp. ORL, KU Leuven, Belgium
WASPAA-2007, Oct 23 2007
Outline Outline
• Hearing aids: bilateral vs. binaural processing
• Binaural multi-channel Wiener filter: transmit all microphone signals large bandwidth of wireless link
• Reduce bandwidth: transmit only one contralateral signal
o signal-independent: contralateral microphone, fixed beamformer o signal-dependent: MWF on contralateral microphones
o iterative distributed MWF procedure:
– rank-1 speech correlation matrix converges to B-MWF solution ! – can still be used in practice when assumption is not satisfied
• Performance comparison:
o SNR improvement (+ spatial directivity pattern)
o dB-MWF performance approaches quite well binaural MWF performance for all conditions
3
• Many hearing impaired are fitted with hearing aid at both ears:
o Signal processing to reduce background noise and improve speech intelligibility
o Signal processing to preserve directional hearing (ILD/ITD cues) o Multiple microphone available: spectral + spatial processing
IPD/ITD
ILD
Hearing aids: bilateral vs. binaural Hearing aids: bilateral vs. binaural
Bilateral/binaural
Binaural MWF
Bandwidth reduction
Experimental results
Conclusions
Hearing aids: bilateral vs. binaural Hearing aids: bilateral vs. binaural
Bilateral system
Independent left/right processing:
binaural cues for localisation are
Binaural system
- Larger SNR improvement (more microphones)
Bilateral/binaural
Binaural MWF
Bandwidth reduction
Experimental results
Conclusions
5
Hearing aids: bilateral vs. binaural Hearing aids: bilateral vs. binaural
• Binaural multi-microphone noise reduction techniques:
o Fixed beamforming
– Low complexity, but limited performance o Adaptive beamforming
– Mostly based on GSC structure + e.g. passing low-pass portion unaltered to preserve ITD cues
o Computationalauditorysceneanalysis – Computation of (real-valued) binaural
mask based on binaural and temporal/spectral cues o Multi-channel Wiener filtering
– MMSE-based estimate of speech component in both hearing aids
– Extensions for preserving binaural cues of speech and noise components
[Desloge 1997, Merks 1997, Lotter 2006]
[Welker 1997, Nishimura 2002, Lockwood 2004]
[Kollmeier1993,Wittkop 2003, Hamacher2002,Haykin2004]
[Doclo, Klasen, Van den Bogaert, Wouters, Moonen 2005-2007]
Bilateral/binaural
Binaural MWF
Bandwidth reduction
Experimental results
Conclusions
Configuration and notation Configuration and notation
• M microphones on each hearing aid: Y0 , Y1
• Speech and noise components:
• Single speech source: (acoustic transfer functions)
• Collaboration: 2N signals transmitted between hearing aids
Bilateral/binaural
Binaural MWF
Bandwidth reduction
Experimental results
Conclusions
7
Binaural MWF (B-MWF) Binaural MWF (B-MWF)
• SDW-MWF using all 2M microphones from both hearing aids:
o All microphone signals are transmitted:
o MMSE estimate of speech component in (front) microphone of left and right hearing aid + trade-off ()
noise reduction speech distortion
speech component in front microphone
• Binaural MWF cost function:
Estimated during speech-and-noise and noise-only periods: VAD
Bilateral/binaural
Binaural MWF
Bandwidth reduction
Experimental results
Conclusions
Binaural MWF (B-MWF) Binaural MWF (B-MWF)
• Optimal filters (general case):
• Optimal filters (single speech source):
o is complex conjugate of speech ITF o Optimal filters at left and right hearing aid are parallel
Bilateral/binaural
Binaural MWF
Bandwidth reduction
Experimental results
Conclusions
9
• To limit power/bandwidth requirements, transmit N=1 signal from contralateral hearing aid
o B-MWF can still be obtained, namely if F01 is parallel to and F10 is parallel to infeasible at first sight since full
correlation matrices can not be computed !
Reduced-bandwidth algorithms Reduced-bandwidth algorithms
Bilateral/binaural
Binaural MWF
Bandwidth reduction -fixed beamformer -contralateral MWF -distributed scheme
Experimental results
Conclusions
Fixed beamformer Fixed beamformer
• Filters F01 and F10 , which can be viewed as monaural beamformers, are signal-independent
• MWF-front: front contralateral microphone signals
• MWF-superd: monaural superdirective beamformer
limited performance
-15 -10 -5 0
30 210
240
270
300
150
330
180 0
-15 -10 -5 0
30 210
240
270
300
150
330
180 0
Bilateral/binaural
Binaural MWF
Bandwidth reduction -fixed beamformer -contralateral MWF -distributed scheme
Experimental results
Conclusions
11
Contralateral MWF Contralateral MWF
• Transmitted signals = output of monaural MWF, estimating the contralateral speech component only using the contralateral microphone signals
o Signal-dependent (better performance than signal-independent) o Increased computational complexity (two MWF solutions for each
hearing aid)
• In general suboptimal solution:
o Optimal solution is obtained in case of single speech source and when noise components between left and right hearing aid are uncorrelated (unrealistic)
Bilateral/binaural
Binaural MWF
Bandwidth reduction -fixed beamformer -contralateral MWF -distributed scheme
Experimental results
Conclusions
Distributed MWF (dB-MWF) Distributed MWF (dB-MWF)
• Iterative procedure:
o In each iteration F10 is equal to W00 from previous iteration, and F01 is equal to W11 from previous iteration
Bilateral/binaural
Binaural MWF
Bandwidth reduction -fixed beamformer -contralateral MWF -distributed scheme
Experimental results
Conclusions
13
Distributed MWF (dB-MWF) Distributed MWF (dB-MWF)
Bilateral/binaural
Binaural MWF
Bandwidth reduction -fixed beamformer -contralateral MWF -distributed scheme
Experimental results
Conclusions
Distributed MWF (dB-MWF) Distributed MWF (dB-MWF)
• Single speech source: convergence to B-MWF solution (!)
o MWF cost function decreases in each step of iteration
o Convergence to B-MWF solution, since it minimises J(W) AND satisfies with
• General case where Rx is not a rank-1 matrix:
o MWF cost function does not necessarily decrease in each iteration o usually no convergence to optimal B-MWF solution
Bilateral/binaural
Binaural MWF
Bandwidth reduction -fixed beamformer -contralateral MWF -distributed scheme
Experimental results
Conclusions
15
Experimental results Experimental results
• Setup:
o Binaural system with 2 omni microphones on each hearing aid, mounted on CORTEX MK2 artifical head in reverberant room
Bilateral/binaural
Binaural MWF
Bandwidth reduction
Experimental results -SNR improvement -directivity pattern
Conclusions
o HRTFs: T60 500 ms (and T60 140 ms), fs = 20.48kHz o Configurations:
– speech source at 0 and several noise configurations (single, two and four noise sources)
– speech source at 90 and noise source at 180
o speech material = HINT, noise material = Auditec babble noise o Input SNR defined on LF microphone = 0dB (broadband)
o Intelligibility-weighted SNR improvement between output signal and front microphone (L+R)
• MWF processing:
o Frequency-domain batch procedure o L = 128, =5
o Perfect VAD,
o dB-MWF procedure: K=10,
SNR improvement (500 ms - left HA) SNR improvement (500 ms - left HA)
4 5 6 7 8 9 10 11 12 13
Performance comparison (left, L=128, T60=500 ms)
AI weighted SNR improvement (dB)
B-MWF MWF-front dB-MWF Original signal
17
• B-MWF:
o In general largest SNR improvement of all algorithms o Up to 4 dB better than MWF-front (3 vs. 4 microphones)
• MWF-superd:
o Performance between MWF-front and B-MWF, but in general worse than (signal-dependent) MWF-contra and dB-MWF
o relatively better performance when (signal-independent) directivity pattern of superdirective beamformer approaches optimal (signal- dependent) directivity pattern of B-MWF, e.g. v=300 (left HA)
• MWF-contra:
o Performance between MWF-front and B-MWF
• dB-MWF:
o Best performance of all reduced-bandwidth algorithms
o Substantial performance benefit compared to MWF-contra, especially for multiple noise sources
o Performance of dB-MWF approaches quite well performance of B-MWF, even though speech correlation matrices are not rank-1 due to FFT overlap and estimation errors, i.e.
Experimental results Experimental results
Bilateral/binaural
Binaural MWF
Bandwidth reduction
Experimental results -SNR improvement -directivity pattern
Conclusions
Experimental results Experimental results
• Directivity pattern:
o Fullband spatial directivity pattern of F01, i.e. the pattern
generated using the right microphone signals and transmitted to the left hearing aid
o Configuration v=[-120 120], T60 = 140 ms
o B-MWF: null steered towards direction of noise sources
optimally signal with high SNR should be transmitted
o MWF-front, MWF-superd: directivity pattern not similar to B-MWF directivity pattern low SNR improvement
o MWF-contra: directivity pattern similar to B-MWF directivity pattern high SNR improvement
o dB-MWF: best performance since directivity pattern closely matches B-MWF directivity pattern
• Using these spatial directivity patterns, it is possible to explain
Bilateral/binaural
Binaural MWF
Bandwidth reduction
Experimental results -SNR improvement -directivity pattern
Conclusions
1919
-20 -15 -10 -5
30 210
60 240
270
120
300
150
330
180 0
(e)
-15 -10 -5 0
30 210
60 240
90 270
120
300
150
330
180 0
(d)
-20 -15 -10 -5
30 210
60 240
90 270
120
300
150
330
180 0
(a)
-20 -15 -10 -5 0
30 210
60 240
270
120
300
150
330
180 0
(c)
-10 -5 0 5
30 210
60 240
90 270
120
300
150
330
180 0
(b)
Contralateral directivity patterns (140 ms) Contralateral directivity patterns (140 ms)
B-MWF MWF-front MWF-superd
MWF-contra dB-MWF
v=[-120 120]
Conclusions Conclusions
• Binaural MWF: large bandwidth/power requirement
• Reduced-bandwidth algorithms:
o MWF-front, MWF-superd: signal-independent
o MWF-contra: monaural MWF using contralateral microphones
– Signal-dependent, but suboptimal
o dB-MWF: iterative procedure
– Converges to B-MWF solution for rank-1 speech correlation matrix – Also useful in practice when this assumption is not satisfied
• Experimental results:
o dB-MWF > MWF-contra > MWF-superd > MWF-front
– Signal-dependent better than signal-independent – 2 or 3 iterations sufficient for dB-MWF procedure
– dB-MWF performance approaches quite well B-MWF performance
• Extension: distributed processing in acoustic sensor
Bilateral/binaural
Binaural MWF
Bandwidth reduction
Experimental results
Conclusions