Adaptive beamforming (GSC) : not robust against signal model errors Robust generalised multi-microphone noise reduction scheme
Efficient implementation using stochastic gradient algorithms
• Robustness: small deviations from assumed microphone characteristics (gain, phase, position) large deviations from desired spatial directivity pattern
- measurement or calibration procedure: expensive, not effective against drift
- incorporate random deviations into design: consider all feasible microphone characteristics and optimise average performance using probability as weight
• Simulations :
N=3, [-0.01 0 0.015] m, L=20, end-fire beamformer (passband: 0o-60o)• Spatial directivity patterns for non-robust and robust beamformer in case of no position errors and small position errors:
[0.002 –0.002 0.002] mDesign, implementation, and evaluation of a robust
multi-microphone noise reduction algorithm
Simon Doclo
1), Ann Spriet
1-2), Jan Wouters
2)and Marc Moonen
1)1)
ESAT-SCD, KULeuven, Kasteelpark Arenberg 10, 3001 Leuven, Belgium
2)
Lab. Experimental ORL, KULeuven, Kapucijnenvoer 33, 3000 Leuven, Belgium
1. Multi-microphone noise reduction techniques
• reduction of noise wrt useful speech signal in different acoustic environments
• exploit spatial + spectral information of speech and noise sources
• small-size microphone arrays increased sensitivity to signal model errors (e.g. microphone mismatch)
2. Spatially pre-processed SDW-MWF
Spatial pre-processing
u0
u1
1
uN
Fixed
beamformer Speech
reference
0 0 0
y x v
Blocking matrix
Noise references
1 1 1
y x v
2 2 2
y x v
1 1 1
N N N
y x v
Multi-channel Wiener Filter (SDW-MWF)
]
0[k w
] [k
z]
1[k w
]
2[k w
]
1[k wN
( ) A z
( ) B z
• Structure of SP-SDW-MWF resembles Generalised Sidelobe Canceller (GSC):
- spatial pre-processor speech reference and noise references
- adaptive stage : adaptive estimation of noise component in speech reference
v
0[k- ]
• Standard GSC minimises output noise power :
• Fixed + adaptive stage rely on assumptions (e.g. no mismatch, no reverberation), but in practice these assumptions are not satisfied speech distortion
- distortion of speech component in speech reference - speech leakage into noise references
• Design of robust noise reduction algorithm :
- robust fixed beamformer
limit distortion in x
0[k] and limit speech leakage
- robust adaptive stage limit effect of (remaining) speech leakage
] [ ] [ ]
[ ]
[ k x
0k k k
z
x w
Tx
0 2
]
[
[ ] [ ] [ ]
min
E v k T k kk w v
w
3. Robust spatial pre-processor
4. Robust adaptive stage: SDW-MWF
4.1. Cost function
4.2. Implementation: stochastic gradient algorithms
• Robustness: limit effect of speech leakage w
T[k]x[k] by controlling filter w[k]
- Quadratic inequality constraint (QIC-GSC): conservative approach - Take speech distortion into account in optimisation criterion (SDW-MWF)
o 1/ trades off noise reduction and speech distortion (1/=0 GSC) o regularisation term ~ amount of speech leakage
• Wiener solution (using )
• Generalised scheme different algorithms, depending on
1/ andw
0- Without w0 : Speech Distortion Regularised GSC (SDR-GSC), i.e. standard ANC criterion is supplemented with regularisation term
- With w0 : Spatially pre-processed SDW-MWF (SP-SDW-MWF)
[k] [k]
E [k] [k]
E [k] [k]
E x xT y yT v vT
]
[k w
Algorithm Complexity (MAC) MIPS
QIC-GSC (3N-1)FFT + 16N - 9 2.16
SDW-MWF (3M+2)FFT + 10M
2+ 15M + 4 2.71
(a), 4.31
(b)• Set-up:
3-microphone BTE (d=1cm,1.5cm) mounted on dummy head - speech (0o) + 5 speech-like noise sources (75o,120o,180o,240o,285o) - microphone gain mismatch 2=4 dB at second microphone• Performance measures: Intelligibility-weighted signal to noise ratio SNR
intelligand spectral distortion SD
intellig• Performance of SP-SDW-MWF:
- GSC (1/ = 0, no w0): degraded performance if significant leakage
- SDR-GSC: 1/ > 0 increases robustness (speech distortion noise reduction) - SP-SD-MWF (w0 ) : performance not degraded by mismatch
• Comparison with QIC-GSC: QIC increases robustness of GSC, but QIC f (amount of speech leakage)
4.3. Experimental validation
SP-SDW-MWF achieves better noise reduction than
QIC-GSC, for a given maximum speech distortion level
noise reduction speech distortion
0 2
2
[ ]
min [ ]
T[ ] [ ] 1
T[ ] [ ]
k
E v k k k E k k
w
w v w x
• Stochastic gradient algorithm in time-domain: LMS-based updating formula
- allows transition to classical LMS-based GSC by tuning parameters (1/, w0) - approximation of regularisation term in time-domain using data buffers
• Complexity reduction in frequency-domain:
block-based implementation (FFT) - approximation of regularisation term replace buffers by correlation matrices• Computational complexity
(N = 3 (mics), M = 2 (a), M = 3 (b), L = 32, fs = 16kHz)Complexity comparable to FD implementation of QIC-GSC
regularisation term Classical GSC
0
[ 1] [ ] [ ] 1
[ ] [ ]
[ ]
T[ ] [ ]
T[ ]
k k k v k k k k k k
w w v v w x x w
speech-and-noise periods noise-only periods
[ [ ] [ ]
1 [ ] [
0
1 1
[ k ] E k ]
T[ k ] 1 E k
Tk E k v k ]
v v v
w y y
0 1
0 1 0 1 0 1
( , , ) ( ) ( )
N
mean N N N
A A
J J A A f A f A dA dA
Non-robust design Robust design