Upload
lydieu
View
214
Download
0
Embed Size (px)
Citation preview
Evaluation of different pre-whitening decorrelation based adaptive feedback
cancellers in hearing aids using perceptual criteria
Kawther Essafi and Sofia Ben Jebara Research Laboratory COSIM
Ecole Superieure des Communications de Tunis, University of Carthage
Route de Raoued 3.5 Km, Cite EI Ghazala, Ariana, 2088, TUNISIA
[email protected], [email protected]
Abstract- This paper proposes a perceptual comparative evaluation of some Adaptive Feedback Cancellation (AFC) methods based on decorrelating pre filters. Our objective is to identifY suitable AFC methods which give the lowest audible degradation of the amplified speech. Hence. some perceptual criteria measuring only the audible part of the degradation are defined. We show that for hearing impaired subjects who suffer from moderate and severe hearing loss. the perceived distortions and oscillations are lower quantities. compared to the total degradation commonly measured using classical criteria. Moreover, the results of the comparative evaluation of some existing decorrelation methods show that a novel proposed approach based on the reduction of both auto-correlation and intercorrelation between signals (BRP-IVM-AFC) provide the best auditive quality for moderate hearing loss and that the method based on frequency bands reduction (bandlimited-AFC) provides an improvement of the perceived auditive quality for severe hearing loss.
Keywords- Hearing aids. acoustic feedback cancellation, decorrelation techniques, perceptual evaluation of sound quality.
I. INTRODUCTION
Hearing aids are designed to amplify sounds and to make them audible for people with hearing impairments. The input
signal is captured by a microphone, it is amplified and send to the inner ear via the loudspeaker (see Fig.l). However,
because of the acoustical coupling between the loudspeaker
and the microphone, the amplified sound is fed back into the microphone. This phenomena is called feedback and causes
distortions extremely bothering, mainly when the gain is
increased, constraining the hearing aid specialist to limit the
gain. Thus, several algorithms and solutions based on adaptive
filtering have been proposed to reduce the acoustic feedback. Despite its significant reduction, some residual feedback still
exist generating some audible distortion, altering the perceived
quality of the hearing aid output. To avoid this degradation, the Adaptive Feedback Cancellation (AFC) solutions are improved
by incorporating some decorrelation methods [1], [2], [3]. ..
Their main objective is to reduce the correlation between the desired input and the loudspeaker signals.
In the literature, the AFC decorrelation methods are mainly evaluated criteria related to the adaptive filters. Recently,
novel investigations have been interested in the quantitative
assessment of the sound quality of the hearing aid output. We
relate mainly the works of A. Spriet et at [4] who introduce
objective criteria to quantify the oscillations and distortions in
978-1-4673-5604-6/12/$31.00 ©20 12 IEEE
x(n) .-----1 J{�n)
Input �I amplifier � Output
Microphone Loudspeaker
Fig. I. A microphone-amplifier-loudspeaker system.
the hearing aid output.
In previous works, we evaluated and compared the perfor
mance of different solutions of AFC based on the use of decorrelation fi Iters in different locations of the feedback
canceller. The quality was evaluated in terms of the adaptive
filter misadjustment, the maximum stable gain (MSG) and the sound quality using the yet mentioned criteria [5].
In [6], we improved the previous criterion while making them perceptive in order to quantify only audible degradation in the
hearing aid output. In fact, it is useless to consider an inaudible
degradation, it only increases the quantitative criterion without being significant for the listener.
In this paper, we aim evaluating and comparing the perfor
mance of some AFC based on decorrelation methods using the quantitative perceptual criteria. The purpose of this paper
is to identify better structures and schemes, which give lower best audible residual feedback for different levels of hearing
loss.
This paper is organized as follows. The next section gives an overview of main feedback cancellers using signals transfor
mations in the context of hearing aids. In section 3, we recall
the classical and perceptual criteria for quantifying the amount of oscillations in the hearing aids output and we present the
results of the comparative evaluation between the methods presented in section 2. In section 4, we evaluate in the same
way the perceptible signal distortion. Finally, the conclusions
are given in section 5.
II. DECORRELATION METHODS FOR AFC
In hearing aids, the conventional solution for acoustic feedback cancellation is depicted in FigA with solid line. The
feedback path is modelled by a finite impulse response filter
F which is adaptively estimated using F(n). The estimated
feedback signal d( n) is subtracted from the microphone signal
y(n), yielding to the desired signal estimation e(n) = x(n) + d( n) - d( n). This estimation is processed by the amplification
000218
Fig. 2. Decorrelation methods of acoustic feedback cancellation in hearing aids.
gain G and its delayed version is sent to the ear through a
loudspeaker.
When the coefficients of the feedback path estimate are determined by minimizing the power of the error signal E{e(n)2}, the optimal filter F converges to the optimum Wiener solution:
F = F + E{U(n)U(nf} -1 E{x(n)U(n)}. (1)
Because of the presence of the closed loop in the hearing
aid equipment, the source signal x( n) and the loudspeaker
signal u( n) are correlated. It leads to the biased feedback path estimate expressed in Eq.l. For this reason, different
decorrelation methods have been proposed in the literature to reduce the bias of the feedback path estimate.
Throughout this paper, we will compare the performances
of different techniques based on the use of decorrelating prefilters. In these techniques, the feedback path estimate is
updated using the transformed signals. In Fig.4, we represent
a generic scheme where decorrelation filters are placed in different locations of the feedback canceller. From this scheme,
several versions can be extracted. In table 1, we present the type of each used filter and the way to obtain each transformed
signal (uj(n), ej(n), up(n), Yj(n)): • FXLMS-AFC [7]: the driving signal uj(n) and the error
signal ej(n) for adaptation of filter coefficients are processed
through a short term predictor H2. The estimated feedback
path is composed of a fixed filter HI in series with the adaptive filter F(n). The filter HI prevents the divergence
of the adaptive filter in the oscillation frequencies and it has the advantage to manage the feedback path estimate which
is HI * F(n) (* is the convolution operation). To provide
satisfactory performance for reducing acoustic feedback path, the choice of the filter HI is based on prior knowledge of
frequency range. Hence, HI is a very rough approximation of
T bl I Al .
h a e 19ont ms an d structures d escnptlOns. Decorrelation methods Algorithms adaptation
FXLMS-AFC F(n + 1) = F(n) + J.Lef(n)Uf(n) A(n) = Identity Uf(n) = H!U'(n), u'(n) = H'[U(n) H, = high-pass filter ef(n) = HJ'E(n) H2 = pre-whitening filter e(n) = y(n) - F(n)TU'(n) BL-AFC F(n + 1) = F(n) + J.Lef(n)Uf(n) A(n) = Identity ef(n) = H!E(n) H, = H2 = band-pass filters e(n) = y(n) - F(n)TU'(n) PEM-AFC F(n + 1) = F(n) + J.Lef(n)Uf(n) A(n) = H, = Identity Uf(n) = H!U(n), ef(n) = H!E(n) H 2 = adaptive pre-whitening filter e(n) = y(n) - F(n)TU(n)
hf2 = E{ef(n)2} IVM-AFC F(n + 1) = F(n) + J.Lef(n)Up(n) H, = H2 = Identity ef(n) = Yf(n) - F(n)TUp(n) A (n) = adaptive pre-whitening filter JA = E{ep(n)
2} BRP-IVM-AFC F(n + 1) = F(n) + J.Lef(n)Up(n) H, = H2 = Identity Up(n) = A(n)TU(n) A (n) = adaptive pre-whitening filter ef(n) = Yf(n) - F(n)TUp(n)
JA = L�:�k AkE {[ep(n)ep(n - k)2}
+ L::�' 'fIE {[Yf(n)up(n - l )f}
the feedback path F. • BL-AFC (Bandlimited-AFC) [8]: it is a variant of the
FXLMS-AFC algorithm used to reduce the feedback signal
only in the frequency range where instability occurs. Hence, the filters HI and H2 are band-pass filters chosen so that
the allowed band covers only the region where oscillation
may happen. Therefore, the feedback canceller may be more efficient for reducing the residual oscillation components es
pecially when the desired signal x( n) has significant energy
in the bands where oscillation frequencies are not located.
• PEM-AFC [9]: Its principle looks like the one of the
FXLMS-AFC algorithm except that the filter HI is removed. For unknown and highly time-varying source signals x(n) modeled using an auto-regressive model, the signals driving
the adaptive algorithm are pre-whitened using an adaptive linear filter H2. However, for fast time-varying feedback path,
the adaptive algorithm could diverge if the adaptive prewhitening filter H2 has a large group delay.
• IVM-AFC [2]: the loudspeaker signal is pre-whitened using
linear prediction filter A(n) and the microphone signal is filtered using the same pre-whitener A( n). In this way, the
adaptive filter F( n) does not suffer from stability problems
when the feedback path change too quickly in time.
• BRP-IVM-AFC [10]: in order to reduce the bias during
feedback path estimation of the IVM-AFC algorithm, a new
pre-whitener filter A( n) is constructed. It is adapted thanks
to the minimization of a criterion considering fourth order
statistics of the pre-whitened error signal ep(n) = AT E(n) and inter-correlation between pre-whitened loudspeaker signal
up(n) and filtered microphone signal Yj(n).
000219
III. A COMPARATIVE S TUDY OF AUDIBLE OS CILLATIONS
A. Criteria definition
To detect the audible oscillations in the hearing aid output,
two criteria have been proposed in [6].
1) Audible Transfer Variation Criterion (TVCaud): In [4], the Transfer Variation Criterion TVC is introduced
to assess the total amount of oscillations in the hearing aid
output. It is defined as the largest peak in the ratio of the Power Spectral Densities (PSD) of the signals u ( n) and r ( n) .
We recall that u ( n) is the amplified signal sent to the inner ear
and r (n) is the ideal one obtained when there is no feedback. The TVC for each frame m, is written as:
Pu(J,m) I) TVC(m) = maxj(IIOloglO
Pr(J, m) , (2)
where f is the frequency, Pu(J, m) (resp. Pr(J, m)) is the
PSD of u (n) (resp. r (n)) .
However, this criteria assesses the overall oscillations without
verifying if the detected oscillations are perceived by the ear or not. In order to measure only audible oscillations, a quan
titative perceptual criteria based on the TV C was proposed
in [6]. This criterion was computed as follows: Pu(J, m) and Pr(J, m) used in the expression of the TVC(m) are replaced
by other quantities which quantify only the audible parts of the spectrum. For such purpose, some auditory properties of
human ear are considered. More precisely, the concepts of
Masking Threshold (MT), the Audible Spectrum (AS) and the Class of Perceptual Equivalence (C P E) are used.
According to the concept of the AS, the amplified error
signal (v (n) = u (n) - r (n)) spectral components below the masking threshold of the desired output signal r ( n) will be
inaudible, while other spectral components are audible. In order to quantifiy only the audible residual feedback, the
power spectrum Pu(J, m) and Pr(J, m) can be replaced by
the audible spectrum ASu(J, m) and ASr(J, m). Moreover, according to the concept of the C P E, it is pos
sible to find other quantities perceptually equivalent to the
spectrum Pu(J, m) and Pr(J, m) in such a way that the difference between them either smaller than the one between
the audible spectrum ASu(J, m) and ASr(J, m). These new spectral shapes are limited by two curves: the Upper Bound of
Perceptual Equivalence (U B P E) equal to the original signal
r (n) over its own masking curve and the Lower Bound of Perceptual Equivalence (LBP E) equal to the original signal
r (n) when it is audible and 0 dB otherwise.
These concepts exploit the notion of masking threshold. So, this difference can be further reduced if the absolute hearing
threshold is replaced by the threshold of hearing impaired in the calculation procedure of the MTr. In this way, the
assessment of the auditive quality will be adapted to the
hearing-impaired. The new spectral quantities called Equivalent Audible Spec
trum (EAS) are used to define the audible Transfer Variation
Criterion:
Frequency in Hz
Fig. 3. Hearing threshold of two hearing-impaired.
EASu(J, m) I) TVCaud(m) = maxj(llOloglO
EASr(J, m) , (3)
where
EASu(f,m) EASr(f,m)
PvU,m) UBPErU,m) if Pu(J,m) > UBPEr(J,m)
and V'Pr(J, m) if MTr(J, m) ::; Pu(J, m)
and Pu(J, m) < Pr(J, m) if Pu(J, m) < MTr(J, m)
and Pr(J, m) ?: MTr(J, m) 1 otherwise
2) Audible Power Concentration Ratio (PCRaud): In [4], the Power Concentration Ratio PCR is defined as
the degree to which a large amount of power is concentrated at a small number of frequencies in the hearing aid output. In
order to reduce its dependency on the power spectrum of the
desired input signal x(n) and the feedback path F, the PCR is computed as the difference between the PC R of the output
signal u (n) and the PCR of the desired output signal r (n) :
PCR(m) = PCRu(m) - PCRr(m), (4)
where PCRr(m) is defined as follows:
(5)
000220
AFC FXLMS BL PEM IVM BRP
• TVCaud severe hearing loss • TVCaud moderate hearing loss • TVC
Fig. 4. Comparaison of the mean oscillation measures in terms of TVC and TVCaud for different AFC methods.
where � = {f, TV F(f) ?: 6dB}. PC Ru (m) has the same
kind of expression than PC Rr (m) where the set � is replaced
by A = {f E C five strongest TV F(f, mH. To quantify only the audible oscillations, we adopt the same methodology as the one used to define the TVCaud [6]. Then,
the audible power concentration ratio PC Raud is defined as:
LjEA' SAEu(m, j) LjEf,' SAEr(m, j) PCRaud(m) =
LjSAEu(m,j) -
LjSAEr(m,j) ,
(6)
where e = {j, TV Faud(f) ?: 6dB} and A' = {f E e, five
strongest TV Faud(f, mH.
B. Comparison results
In the following experiments, we use a realistic acoustic
feedback path for behind-the-ear hearing aid system with
LF = 64 taps (for a sampling frequency of 16 kHz). The desired input signal x ( n) is a real speech sequence (18
seconds) at a level of 60 dB SPL. The total delay from the microphone signal to the loudspeaker signal equals dG = 6ms. The adaptive filter F(n) is governed by the partitioned-block
frequency domain LMS algorithm (64 point FFT, block size of 32) and the step-size is normalized per frequency-bin by the
sum of the input and error powers. The objective measures
are computed using half-overlapping frames of 0.5 second. The validation of perceptual criteria was carried for moderate
hearing loss with an average hearing level of 43 dB and severe hearing loss with an average hearing level of 80 dB (see Fig.3).
The validation of classical criteria is the same for all degrees
of hearing loss, because the hearing threshold is not used in the decorrelation methods.
20 ,-----------------------------------
18 L...r--___ ----"
16
14
12
10
6
4
AFC FXLMS BL PEM IVM BRP
• TVCaud severe hearing loss • TVCaud moderate hearing loss • TVC
Fig. 5. Comparaison of the maximum oscillation measures in terms ofTVC and TVCaud for different AFC methods.
1) Evaluation in Terms of TVC and TVCaud: Fig.4 and Fig.5 show respectively the performances of the acoustic feedback cancellation techniques previously men
tioned in terms ofTVC and TVCaud for moderate and severe hearing loss, by considering the mean value over frames and
the maximum value. The maximum measures allow to assess
the segment of the speech output that exhibit most distortions and oscillations. The mean performance measure allows to
assess the overall sound quality of the hearing aid output. We
present a comparative assessment of the average performances for four gain settings: 18 dB is the Maximum Stable Gain
(MSG) without feedback cancellation, 24 dB, 28 dB are two intermediate gain values used during feedback cancellation and
a gain limit over which instability occurs. It is equal to 32 dB
for all solution. - When comparing the TVCaud for the two different hearing
characteristics with TVC, we remark a significant reduction
of the amount of audible oscillation in the hearing aid output. It is valid for all tested techniques using mean and maximum
values of objective measures. - We can also see that all algorithms achieve less audible
oscillations in the case of severe hearing loss compared to the
case of moderate hearing loss. It is explained by the fact that the frequency band of inaudibility is larger. So, the amount of
degradation is lesser.
- We should also remark that the PEM method gives better TVC performance than the Bandlimited method. But using
the perceptual criteria TVCaud, we obtain the opposite. It mean that the PEM provide more audible oscillations than the
Bandlimited-AFC, which renders it less powerful.
- In general, the BRP-IVM-AFC provide less oscillations than other algorithms using classical criteria and perceptual criteria
for moderate hearing loss.
000221
14 ------�==�--------------------------
12 -------1
10 ----. -..---------1
AFC FXlMS Bl PEM IVM BRP
• PCRalid smre hearing loss • PCRalid moderate hearing loss • PCR
Fig. 6. Comparaison of the mean oscillation measures in terms of PCR(1O-3) and PCRaud(10-3) for different AFC methods.
- For hearing impaired subjects who suffer from a severe
hearing loss, we also notice that the Bandlimited-AFC provides the best amount of the residual audible oscillations.
2) Evaluation in Terms of PC R and PC Raud: Fig.6 and Fig.7 show respectively the performances of the feedback cancellation solutions on the basis of mean and
maximwn values of objective measures PC R and PC Raud for moderate and severe hearing loss.
- When dealing with mean values of objective measures
PC R and PC Raud (see Fig.6), we remark that the oscillations introduced by the FXLMS-AFC method in terms of PCR and
PC Raud in the case of moderate hearing loss are the larger.
However, in term of PC Raud for severe hearing loss, it is almost the same than the standard AFC.
- When comparing the IVM-AFC method with the algorithms BL-AFC and PEM-AFC, we can clearly see that the amount of
oscillations is more elevated in terms of PC R and PC Raud. However, the BRP-IVM-AFC, which corresponds to a modified version of the IVM-AFC, achieves the best reduction
of the amount of oscillations for the two hearing-impaired
subjects. - When dealing with maximwn values of objective measures
PCR and PCRaud (see Fig.9), we notice that the IVMAFC exhibit more fluctuations than other algorithms using
all selected criteria. However, the BRP-IVM-AFC exhibit a
significant improvement of the auditive quality in term of oscillati on.
IV. A COMPARATIVE S TUDY OF AUDIBLE DIS TORTIONS
A. Criteria definition
An objective measure to quantify the distortion in the
loudspeaker signal u (n) was proposed in [4]. It is called the
20 -------------��-----
18 ------------�
16 -------------------------
14 -------1
12
10
AFC FXlMS Bl PEM IVM
• PCRalid smre hearing loss • PCRalid moderate hearing loss • PCR
BRP
Fig. 7. Comparaison of the maximum oscillation in terms of PCR(1O-3) and PCRaud(10-3) for different AFC methods.
frequency weighted log-spectral Signal Distortion (SD) and is defined as follows:
SD(m) =
6500Hz " ( ) ( Pu(f, m)
)2 � JERE f lOloglO P (f m)
, f=300Hz r ,
(7)
where IeRE(f) is a weight affected to each auditory critical
band Bi of Equivalent Rectangular Bandwidth (ERB) scale. To quantify only the audible distortion, we adopt the same
methodology as the one used to define the TVCaud [6]. Then,
the audible signal distortion SDaud is defined as:
6500Hz EASu(f, m) 2 L JERE (f) (lOloglO EAS (f m)) . f=300Hz r ,
B. Comparison results
(S)
Fig.S and Fig.9 show respectively the performances of the different algorithms of mean and maximwn values of objective
measures SD and SDaud for moderate and severe hearing
loss. -When dealing with mean values of objective measures S D and SDaud (see Fig.S), it appears that the BRP-IVM-AFC pro
vides less distortion than other algorithms using classical criteria SD. While in term of perceptual criteria SDaud, it almost
presents the same rate of degradations as the BandlimitedAFC method. We also remark that the PEM method provides
less distortion than the standard AFC and FXLMS algorithm
using classical criteria and perceptual criteria in the case of moderate hearing loss. On the other hand, for hearing
impaired subjects who suffer from a severe hearing loss, these
000222
4 ---------------------------------------
3,5
2,5
1,5
0,5
AFC FXLMS BL PEM IVM BRP
I SOaud serere hearing loss I SOaud moderate hearing loss I SO
Fig. 8. Comparaison of the mean distortion measures for different AFC methods.
three algorithms have the same performances. Therefore, this justifies the interest of the perceptual criteria in the validation
of the acoustic feedback cancellation techniques in term of the
sound quality as perceived by hearing impaired subjects. - When dealing with maximum values of objective measures SD and SDaud (see Fig.9), we notice a significant reduction in
the level of audible oscillations for different acoustic feedback reduction techniques. Oddly, the PEM method exhibit more
distortion than the standard AFC algorithm using the classical
criteria and perceptual criteria. In general, the BRP-IVM-AFC provides an improvement of the auditive quality for different
degrees of hearing loss.
V. CONCLUSION
In this paper, we have presented a comparative assess
ment of the performances of some existing pre-whitening decorrelation methods for feedback cancellation in hearing
aids. The used criteria are based on properties of the human
auditory system in such a way that they measure only audible degradation in the hearing aid output. From the comparative
simulation results, we may conclude that, among many decorrelation methods, the BRP-IVM-AFC approach improves the
performances of the hearing aid output for moderate hearing
loss. It is comparable to the Bandlimited-AFC for severe hearing loss. The calculation of the degree of correlation of
the perceptual criteria with sUbjective criteria constitutes the
perspectives of our work.
REFERENCES
[I] M. G. Siqueira and A. Alwan, "Steady-state analysis of continuous adaptation in acoustic feedback reduction systems for hearing-aids," IEEE Trans. Speech and Audio Processing, vol. 8(4), pp. 443-453, July 2000.
[2] A. Spriet, G. Rombouts, M. Moonen and J. Wouters, "Adaptive feedback cancellation in hearing aids," Journal of the Franklin Institute, 343(6), Sept. 2006.
AFC FXLMS BL PEM IVM BRP
ISOaud smre hearing loss ISOaud moderate hearing loss ISO
Fig. 9. Comparaison of the maximum distortion measures for different AFC methods.
[3] 1. Hellgren and T. B. Elmedyb, "Generation of probe noise in a feedback cancellation system," EUROPEAN PATENT APPLICATION 07112147.9, Jan. 2009.
[4] A. Spriet, M. Moonen and 1. Wouters, "Objective evaluation of feedback reduction techniques in hearing aids," 17th European Signal Processing Con! EUSiPCO, Glasgow, Scotland, Aug. 2009.
[5] K. Essafi and S. B. Jebara, "A comparative study between different prewhitening decorrelation based acoustic feedback cancellers," IEEE 12th into workshop on multimedia signal processing MMSP, Saint-Malo, France, Oct. 2010.
[6] K. Essafi and S. B. Jebara, "Criteres perceptifs d'evaluation de la qualite des signaux dans les protheses auditives," XXflie colloque GRETSi, Traitement du signal et des images, Bordeaux, France, Sept. 20 I I.
[7] 1. Hellgren, "Analysis of feedback cancellation in hearing aids with filtered-X LMS and the direct method of closed loop identification," IEEE Trans. Speech, Audio Processing, vol. 10, no. 2, Feb. 2002.
[8] H. F. Chi, S. X. Goa, S. D. Soli and A. Alwan, "Band-limited feedback cancellation with a modified filtered-X LMS algorithm for hearing aids," Speech Communication, 39(1-2), Jan. 2003.
[9] A. Spriet, 1. Proudler, M. Moonen and J. Wouters, "Adaptive feedback cancellation in hearing aids with linear prediction of the desired signal," IEEE Trans. Signal Processing, 53 (10 (Part 1)),2005.
[10] K. Essafi and S. B. Jebara, "A decorrelation based adaptive prediction filter for acoustic feedback cancellation in hearing aids," IEEE 10th into conference on information Science, Signal Processing and their Applications ISSPA, Malaysia, May 2010.
000223