prediction-error-method-based adaptive feedback cancellation in hearing aids using pitch estimation
Kim Ngo 1 , Toon van Waterschoot 1 , Mads Græsbøll Christensen 2 , Marc Moonen 1 , Søren Holdt Jensen 3 and Jan Wouters 4
1 Katholieke Universiteit Leuven, ESAT-SCD, Leuven, Belgium
2 Aalborg University, Dept. of Media Technology, Aalborg, Denmark
3 Aalborg University, Dept. Electronic Systems, Aalborg, Denmark
4 Katholieke Universiteit Leuven, ExpORL, O. & N2, Leuven, Belgium
Adaptive Feedback Cancellation (1)
Concept:
• Adaptively model the feedback path and estimate the feedback signal.
• Subtract estimated feedback signal from the microphone signal.
forward path
feedback cancellation
path
acoustic feedback path
+ + − u(t)
G F
ˆ
y [t|ˆf(t)]
x (t) y(t) v(t)
F ˆ
d[t , ˆf(t)]
u(t)=loudspeaker signal x(t)=feedback signal v(t)=near-end signal
ˆf(t)=Feedback path estimate
y[t|ˆf(t)]=predicted feedback signal ˆ
Microphone signal: y(t) = v(t) + x(t) = v(t) + F (q, t)u(t)
Feedback-compensated signal: d(t) = v(t) + [F (q, t) − ˆ F (q, t)]u(t).
Problem:
• Standard adaptive filtering results in a biased solution.
• Correlation between near-end signal and loudspeaker signal caused by closed signal loop.
Current solutions:
• Prediction error method (PEM)-based AFC with linear prediction (LP) model [1].
• PEM-based AFC with cascaded near-end signal model [2].
Proposed solution:
• PEM-based AFC with cascaded near-end signal model using pitch estimation [3][4].
Near-End Signal Model (3)
Harmonic sinusoidal near-end signal model:
Near-end signal: v(t) =
P
X
n=1
a n cos(nω 0 t + φ n ) + r(t)
• ω 0 is the fundamental frequency, a n the amplitude, and φ n the phase
• r(t) is the noise which is assumed to be autoregressive (AR)
• r(t) = C (q,t) 1 e(n) with C(q, t) = 1 + P n
Ci=1 c (i) (t)q −i
Near-end signal model parameter estimation:
Fundamental frequency: ω ˆ 0 = arg max ω
0E{|h H d ˜ (t)| 2 }
• where h is the optimal filter.
Amplitude: ˆ a =
Z H Z −1
Z H d
• where Z is the Vandermonde matrix containing the sinusoids Order: P ˆ = arg min
P M log ˆ σ P 2 + P log M + 3
2 log M
• The AR parameters of the noise component r(t) can be estimated using LP
Experimental Results (5)
Performance measures:
Maximum Stable Gain: MSG(t) = −20 log 10
"
max ω∈P |J(ω, t)[F (ω, t) − ˆ F (ω, t)]|
#
Misadjustment: MA F = 20 log 10 ||ˆf(t) − f|| 2
||f|| 2 .
0 5 10 15 20 25 30
10 12 14 16 18 20 22 24 26 28 30
t(s)
MSG(dB)
20 log10K(t) MSG F (q)
AFC-optfilt (variable amplitude) AFC-optfilt (variable order)
0 5 10 15 20 25 30
10 12 14 16 18 20 22 24 26 28 30
t(s)
MSG(dB)
20 log10K(t) MSG F (q)
AFC-optfilt (P =15) AFC-optfilt (P =10) AFC-optfilt (P =5)
0 5 10 15 20 25 30
10 12 14 16 18 20 22 24 26 28 30
t(s)
MSG(dB)
20 log10K(t) MSG F (q)
AFC-CPZLP (P =15) AFC-CPZLP (P =10) AFC-CPZLP (P =5)
0 5 10 15 20 25 30
−16
−14
−12
−10
−8
−6
−4
−2 0
t (s) MAF(dB)
AFC-optfilt (variable amplitude) AFC-optfilt (variable order)
0 5 10 15 20 25 30
−16
−14
−12
−10
−8
−6
−4
−2 0
t (s) MAF(dB)
AFC-optfilt (P =15) AFC-optfilt (P =10) AFC-optfilt (P =5)
0 5 10 15 20 25 30
−16
−14
−12
−10
−8
−6
−4
−2 0
t (s) MAF(dB)
AFC-CPZLP (P =15) AFC-CPZLP (P =10) AFC-CPZLP (P =5)
Prediction Error Method based AFC (2)
Concept:
• Reduce the correlation between the near-end signal and the loudspeaker signal
• Prefiltering of loudspeaker and microphone signals with inverse near-end signal model
+ +− forward
path
+−
feedback cancellation
path
acoustic feedback path
decorrelating
source signal model prefilter
G Fˆ
y[t|ˆf(t)]ˆ
x(t)
H e(t) y(t) v(t)
y[t, ˆh(t)]˜ ˆf(t)
u(t)
Hˆ−1
u[t, ˆh(t)]˜
Hˆ−1 Fˆ
ε[t, ˆh(t), ˆf(t − 1)]
F
d[t, ˆf(t)]
• Single all-pole model (Short-term predictor) fails to remove the periodicity
• A cascade of near-end signal models removes the coloring and periodicity
• PEM delivers an unbiased estimate of the feedback path coefficient vector f (t) Cascaded near-end signal model: H 1 (q, t)H 2 (q, t) = B (q, t)
A(q, t)
1 C (q, t) Minimize prediction error: min
f (t) t
X
k=1
ε 2 (k)
Prediction error: ε[t, ξ(t)] = H 1 −1 (q, t)H 2 −1 (q, t)[y(t) − F (q, t)u(t)]
Prediction Error Filter (4)
PEF for sinusoidal components:
• A sum of P sinusoids can be described using an all-pole model of order 2P
• When noise is added a pole-zero model of order 2P should be used
Cascade of second-order sections: H 1 −1 (q, t) =
P
Y
n=1
1 − 2ν n cos nω 0 z −1 + ν n 2 z −2 1 − 2ρ n cos nω 0 z −1 + ρ 2 n z −2 (1)
• The poles and zeros are on the same radial lines
• the poles are positioned between the zeros and the unit circle, i.e., 0 ≪ ρ n < ν n ≤ 1.
0 1000 2000 3000 4000 5000 6000 7000 8000
20 25 30 35 40 45 50 55
Frequency (Hz)
Magnitude (dB)
500 1000 1500 30
35 40 45 50
Frequency (Hz)
Magnitude (dB)
(a)
1000 2000 3000 4000 5000 6000 7000 8000
−100
−80
−60
−40
−20 0
Frequency (Hz)
Magnitude (dB)
200 400 600 800 1000 1200 1400
−100
−80
−60
−40
−20 0
Frequency (Hz)
Magnitude (dB)
(b)
0 200 400 600 800 1000 1200 1400
−300
−200
−100 0
Frequency (Hz)
Magnitude (dB)
200 400 600 800 1000 1200 1400
−60
−40
−20 0
Frequency (Hz)
Magnitude (dB)