Superdirective beamforming Superdirective beamforming
robust against microphone robust against microphone
mismatch mismatch
Simon Doclo, Marc Moonen
Dept. of Electrical Engineering (ESAT-SCD), KU Leuven, Belgium
ICASSP-2006, May 17 2006
Overview Overview
• Fixed superdirective beamforming:
o Optimal suppression of diffuse noise field
o Sensitive to uncorrelated noise and microphone mismatch
• Robust design procedures:
o Limit white noise gain
o Take into account statistics of microphone characteristics:
– Mean and worst-case directivity factor – Mean noise and distortion energy
– Mean deviation from desired directivity pattern
• Simulation results:
o Mean/worst-case directivity factor is preferred procedure
o Suitable parameter range for other design procedures
3
Fixed beamforming Fixed beamforming
• Speech and noise sources with overlapping spectrum at different positions
Exploit spatial diversity by using multiple microphones Spatial focus on speech source + suppress noise and reverberation from certain directions
• Fixed beamformers:
o Direction of speech source and microphone configuration assumed to be known
o Applications: hearing aids, teleconferencing, pre-processing stage in adaptive beamformers (GSC)
o Different types: delay-and-sum beamformer, differential microphone array, frequency-invariant beamformer, superdirective beamformer
Fixed beamforming
Superdirective beamforming
Robust
superdirective beamforming
Simulation results
Conclusions
Design of fixed beamformer (1) Design of fixed beamformer (1)
• Configuration:
o Linear microphone array (N microphones, distance dn) o Far-field assumption
o Speech source (s,s) + noise field
• Steering vector:
2( , , )
v
s S
Y g V g
n , , A
n , , e
j n , microphone characteristics
(gain, phase)
delay (position)
Fixed beamforming
Superdirective beamforming
Robust
superdirective beamforming
Simulation results
Conclusions
5
Design of fixed beamformer (2) Design of fixed beamformer (2)
• Output signal:
• Directivity pattern: transfer function between source and output
H
Z W Y
, ,
H , ,
H
W
g
• Array Gain: SNR improvement between input and output signal
with normalised noise correlation matrix (i.e. noise coherence matrix for homogeneous noise field)
• Directivity factor (DF): ability of array to suppress diffuse noise
H 2
s H
VV
G
W g WΦ
WVV
Φ
2
VV
H
s
H diff
DF
W g WΦ
W
H 2
s
WNG H
W g
W W
• White noise gain (WNG): ability of array to suppress spatially uncorrelated noise (e.g. sensor noise) measure for robustness
Fixed beamforming
Superdirective beamforming
Robust
superdirective beamforming
Simulation results
Conclusions
Depend on
mic char
Superdirective beamforming (1) Superdirective beamforming (1)
• Optimization criteria: maximize array gain for diffuse noise
• Solution:
min , s.t. 1
VV
H diff H
s
W WΦ W
W g
1
1
VV
VV
diff
s
sd H diff
s s
W
Φ ggΦ g
Sensitive to uncorrelated noise, i.e. small WNG, especially for small-size microphone arrays at low frequencies
• WNG constraint: limit amplification of uncorrelated noise
needs to be chosen in function of the amount of sensor noise
• Solution:
min , s.t. 1,
VV
diff
s
H
H H
W WΦ W
W g W W
1
, 1
VV
VV
diff
s
sd H diff
s s
W
Φ
gΦ I g g
I
2
max , s.t. 1
VV
H
s H
H diff s
W
W g W g WΦ W
Unity constraint in speech direction
Fixed beamforming
Superdirective beamforming
Robust
superdirective beamforming
Simulation results
Conclusions
7
Superdirective beamforming (2) Superdirective beamforming (2)
• Sensitivity to microphone mismatch:
o N=3, [0 0.01 0.025]m, s=0o, fs=16 kHz
o Deviation: [0 2 0]dB, [-5 10 5]o, [0.001 –0.001 0.001]m
o Determine such that requirements are met for this mismatch
0 1000 2000 3000 4000 5000 6000 7000 8000
0 1 2 3 4 5 6 7 8 9 10
Frequency [Hz]
Decrease of Directivity factor [dB]
=0
=0.0001
=0.001
=0.01
=0.1
=1
=10
Fixed beamforming
Superdirective beamforming
Robust
superdirective beamforming
Simulation results
Conclusions
Robust superdirective beamforming Robust superdirective beamforming
• In practice microphone characteristics are never exactly known
• Instead of measuring/calibrating or limiting WNG, take all feasible microphone characteristics into account and optimise a mean performance criterion using probability as weight:
1. Mean (or worst-case) directivity factor
2. Weighted sum of mean noise and distortion energy
3. Mean deviation from desired superdirective directivity pattern
• Related to earlier proposed design procedures for robust beamformers with an arbitrary directivity pattern
[S. Doclo, M. Moonen, “Design of broadband beamformers robust against gain and phase errors in the microphone array characteristics,” IEEE Trans. Signal Processing, vol. 51, no. 10, pp. 2511-2526, Oct. 2003]
Take into account stochastic deviations in design
, , , ,
j n , , j ncosc fsn n
A a e e
Knowledge of probability density function required f
A
Fixed beamforming
Superdirective beamforming
Robust
superdirective beamforming
Simulation results
Conclusions
9
Robust superdirective beamforming Robust superdirective beamforming
1. Mean directivity factor
o filter W cannot be extracted from integrals discrete sum
o Iterative optimization techniques (e.g. quasi-Newton method)
0 1
, ( )
0(
1)
0 1m A AN N N
DF DF f A f A dA dA
W
W A
0 1
0 1 0 1
, ( ) ( )
N
m N N
A A
DF DF f A f A A A
W
W A
2. Worst-case directivity factor
o finite grid of microphone characteristics
o Minimax optimization problem (e.g. sequential quadratic program)
1
2
tot
T
DF DF DFK
F W W W
Wmin
max min
k
DF
k DFW W
Fixed beamforming
Superdirective beamforming
Robust
superdirective beamforming
Simulation results
Conclusions
Preferred design procedures
Robust superdirective beamforming Robust superdirective beamforming
3. Weighted sum of mean noise and distortion energy
,
tm vm dm
J W
J W
J W
0 1
( )
0(
1)
0 1N VV
H diff
vm A A N N
J f A f A dA dA
W
WΦ
A W
0 1
2
0 1 0 1
1 ( ) ( )
N
H
dm A A s N N
J f A f A dA dA
W W g A
4. Mean deviation from desired superdirective directivity pattern
0 1
,
, ( )
0(
1)
0 1LS m A AN LS N N
J J f A f A dA dA
W
W A
,
02 0 , , , , , ,
2JLS W A
F
H
D
d d
directivity pattern of superdirective BF when no microphone
mismatch occurs directivity pattern for
specific microphone characteristic A
Fixed beamforming
Superdirective beamforming
Robust
superdirective beamforming
Simulation results
Conclusions
Quadratic cost functions closed-from expression
Not optimising directivity factors
11
Simulation results (1) Simulation results (1)
• Set-up and microphone characteristics:
o N=3, [0 0.01 0.025]m, s=0o, fs=16 kHz, design frequency 1 kHz o Nominal microphone characteristics: An()=1, same pdf
o Only gain deviations: uniform gain pdf (a=1, sa=0.3)
o Grid spacing for mean/worst-case directivity factor: a=0.02 o Measures: DF without deviation + mean/worst-case DF
f
A
Fixed beamforming
Superdirective beamforming
Robust
superdirective beamforming
Simulation results
Conclusions
Simulation results (2) Simulation results (2)
10-4 10-3 10-2 10-1 100 101 102
0 2 4 6 8
DF [dB]
Directivity factor no deviation
10-4 10-3 10-2 10-1 100 101 102
0 2 4
DFm [dB]
Mean directivity factor - max = 4.88dB
10-4 10-3 10-2 10-1 100 101 102
-40 -20 0
DFmin [dB]
Worst-case directivity factor - max = 2.43dB
=0.01
=0.07
• Limit WNG: parameter needs to be tuned
13
Simulation results (3) Simulation results (3)
• Directivity patterns: with/without gain deviation [0.7 1.3 1.2]
Fixed beamforming
Superdirective beamforming
Robust
superdirective beamforming
Simulation results
Conclusions
-20 -10 0
30
210
60
240
90
270 120
300
330
180 0
Mean directivity factor Wm
-20 -10 0
30
210
60
240
90
270 120
300
330
180 0
Worst-case directivity factor Wmin
-20 -10 0 10 20 30
30
210
60
240
90
270 120
300 150
330
180 0
Superdirective (=0), Jvm=17dB, Jdm=17dB
-20 -10 0
30
210
60
240
90
270 120
300 150
330
180 0
=0.03, Jvm=-4.1dB, Jdm=-9.5dB
-20 -10 0
30
210
60 90
120 150
330
180 0
=0.1, Jvm=-3.6dB, Jdm=-12dB
-20 -10 0
30
210
60 90
120 150
330
180 0
Delay-and-sum (=), Jvm=-0.2dB, Jdm=-20dB
Simulation results (4) Simulation results (4)
Fixed beamforming
Superdirective beamforming
Robust
superdirective beamforming
Simulation results
Conclusions
scale: 30dB
15
Conclusions Conclusions
• Superdirective beamforming:
o Commonly used fixed beamforming technique (hearing aids) o Maximises array gain for diffuse noise field
o Sensitive to uncorrelated noise and microphone mismatch, especially for small-size arrays at low frequencies
Fixed beamforming
Superdirective beamforming
Robust
superdirective beamforming
Simulation results
Conclusions
• Robustness improvement:
o Limit white noise gain parameter needs to be tuned
o Take into account statistics of microphone characteristics and optimize mean performance criterion:
– Mean/worst-case directivity factor: preferred designed procedure – Weighted sum of mean noise and distortion energy parameter
needs to be tuned
– Mean deviation from desired directivity pattern lowest performance