See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3181200

Adaptive feedback active noise control headset: Implementation, evaluation and

its extensions

Article in IEEE Transactions on Consumer Electronics · September 2005

DOI: 10.1109/TCE.2005.1510511 · Source: IEEE Xplore

CITATIONS READS
100 4,540

3 authors, including:

Woon-Seng Gan Sen M. Kuo

Nanyang Technological University Singapore Northern Illinois University
556 PUBLICATIONS 8,422 CITATIONS 217 PUBLICATIONS 8,587 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Sen M. Kuo on 07 January 2014.

The user has requested enhancement of the downloaded file.

W. S. Gan et al.: Adaptive Feedback Active Noise Control Headset: Implementation, Evaluation and Its Extensions 975

Adaptive Feedback Active Noise Control Headset:

Implementation, Evaluation and Its Extensions

Woon S. Gan, Senior Member, IEEE, Sohini Mitra and Sen M. Kuo Senior Member, IEEE

Abstract — In this paper, we present design and real-time and performance deficiencies caused by non-stationary
implementation of a single-channel adaptive feedback active reference inputs, measurement noise, acoustic feedback [1],
noise control (AFANC) headset for audio and communication and higher cost of using additional reference microphones.
applications. Several important design and implementation More recently, technique has been developed to combine the
considerations, such as the ideal position of error analog feedback with digital feedforward [4] to achieve better
microphone, training signal used, selection of adaptive noise canceling performance. However, the limited flexibility
algorithms and structures will be addressed in this paper. in using the analog filter can be a restriction for further
Real-time measurements and comparisons are also carried improvement, such as the on-line modeling of the secondary
out with the latest commercial headset to evaluate its path.
performance. In addition, several new extensions to the
In response to the preceding problems, an adaptive feedback
AFANC headset are described and evaluated.
active noise control (AFANC) communication headset is
designed. This paper is divided into three main sections. The
Index Terms — Digital signal processing, active noise first section gives an overview of the AFANC headset and
control headset, and real-time implementation. answers some of the implementation questions. Several
experiments have been conducted to model the secondary path
transfer function for different error microphone positions
I. INTRODUCTION inside the shell of the headset. In particular, we propose a
As portable audio and communication devices such as MP3 broadband musical signal as the training signal for a better
players, cellular phones, PDAs, and wireless communication secondary path modeling of the headsets. The second section
headsets become more widely used, designers will need to describes the real-time implementation of the AFANC headset
address the usage of these devices in noisy environments. using a programmable DSP processor, the evaluation of our
Acoustic noise problems become more serious as increased AFANC headset with a commercial high-end active noise
numbers of industrial equipment such as engines, blowers, control headset. In the third section, an integrated approach in
fans, transformers, and compressors are in use in many designing a noise reduction headset for the audio and
outdoor installations, planes, and automobiles. The traditional communication applications is discussed. This integrated
passive earmuffs are valued for their high attenuation over a
system not only cancels the noise for near-end listening, but
broad frequency range; however, they are relatively large,
also cleaned up the noise before transmitting to the far end.
costly, and ineffective at low frequencies. Active noise control
The performance of this integrated system is also being
(ANC) [1]-[3] systems cancel the unwanted noise based on the
principle of superposition. Specifically, an anti-noise of equal evaluated in term of the reduced noise level obtained.
amplitude and opposite phase is generated and combined with
the primary noise, thus resulting in the cancellation of both II. ADAPTIVE FEEDBACK ACTIVE NOISE CONTROL
noises. The ANC system efficiently attenuates low frequency This section covers the structure of the AFANC, its
noise, where passive methods are ineffective, bulky in size, algorithms, secondary path updates using musical signal and
and tend to be very expensive. ANC is developing rapidly the ideal position of error microphone inside the headset.
because it permits improvement in noise reduction, which
results in potential benefits in weight, volume, and cost. A A. Feedback ANC System
better approach is to use a combination of passive and ANC In an ANC application, the primary near-end noise d(n) is not
technique. available during the operation of ANC because it was canceled
by the secondary noise. Therefore, the basic idea of an
Previously, analog controller using feedback configuration
AFANC is to estimate the primary noise, and use it as a
commonly referred to as active noise reduction [4] has been
reference signal x(n) for the ANC filter, W(z). Unlike an
employed to cancel noise in the headsets. Since this is a non-
adaptive feedforward ANC system, where a separate sensor is
adaptive approach, no on-line modeling of the ear-cup transfer
available to pick up the reference signal, the AFANC
function can be carried out in real time. Current research in
regenerates its own reference signal. An AFANC system is
ANC for communication headset focuses on using adaptive
required for applications include spatially incoherent noise
feedforward technology [2]-[3]. In practice, however, the
generated from turbulence; noise generated from many
feedforward ANC systems for headset have to handle causality
Contributed Paper
Manuscript received May 18, 2005 0098 3063/05/$20.00 © 2005 IEEE
976 IEEE Transactions on Consumer Electronics, Vol. 51, No. 3, AUGUST 2005

sources, and propagation paths, and induced resonance where select a training signal that is persistently excited so as to
no coherent reference signal is available. model the secondary path over the entire frequency range of
interest. White noise, which has a flat spectral density for all
The complete AFANC system using the filtered-x least- frequencies, is commonly used as an ideal broadband training
mean-square (FXLMS) algorithm [1] is illustrated in Fig. 1, signal for system identification. In some applications, one
where Sˆ ( z ) is required to compensate for the secondary path. might use bandlimited or colored noise, where the power of
The reference signal x(n) is synthesized as an estimate of d(n), the signal is constant over the range of frequencies of interest,
which is expressed as and zero elsewhere. A chirp signal can also be used for this
M −1 purpose. However, for ANC headset application, where the
x ( n ) ≡ dˆ ( n ) = e( n ) + ¦ sˆm y ( n − m) off-line modeling signal is heard every time the AFANC
m=0 , (1) headset is turned on, it is desirable to use a more soothing
where sˆm , m = 0 , 1, ..., M − 1 are the coefficients of the Mth signal for training.
order FIR filter Sˆ ( z ) used to estimate the secondary path. d(n) + e(n)
Σ
The secondary signal y(n) is generated as: _
x(n) y(n)
L −1 W(z) S(z)
y ( n ) = ¦ wl ( n ) x ( n − l ) , (2)
l =0
Sˆ ( z) Sˆ ( z)
where wl ( n ), l = 0 , 1, ... L − 1 are the coefficients of W(z) at time
n, and L-1 is the order of the FIR filter W(z). The filter’s x′(n) y′(n)
LMS
coefficients are updated by the FXLMS algorithm expressed as +
+
follows: Σ
dˆ (n)

wl ( n + 1) = wl ( n ) + μx′( n − l ) e( n), l = 0, 1, ..., L − 1 (3) Figure 1. Block diagram of adaptive feedback ANC system

where μ is the step size, and A possibility is to use music with wide bandwidth excitation
M −1 to model the secondary path over a period of time. Since the
x′( n ) ≡ ¦ sˆm x ( n − m ) (4) standard sampling frequency for communication applications
m=0
is 8 kHz, any ideal training signal specific to such applications
is the filtered reference signal. should have a flat spectrum in the 0 to 4 kHz frequency range.
The AFANC algorithm summarized in (1) to (4) is very Thus, the first step is to search for a musical piece with rich
effective and compact since no reference sensor is required at frequency content in the 0 to 4 kHz frequency range. However,
the external of the headset. However, in some cases, large the challenge is to find a musical piece that has constant signal
noise level difference can make the algorithm unstable. An envelope over the period of training. We found that an opera
effective solution is to employ the normalized FXLMS piece from “O Fortuna from Carmina Burana,” [5] consists of
algorithm [1] as follows: a number of consecutive frames that are rich in frequency
content as shown in Fig. 2. A power spectrum analysis on
μ these cumulative frames showed it to be almost flat in the
wl ( n + 1) = wl ( n ) + x′( n − l ) e ( n ), (5)
Pˆx ( n ) + c frequency range from 0-1700 Hz. After this range, the
spectrum slowly sloped down up to 4 kHz. Thus, it was
for l = 0, 1, ..., L − 1 , where the power estimate decided that equalization be applied to the upper frequency
(using highpass filter with cutoff frequency at 1.7 kHz,
2 followed by 10 dB of amplification) of this musical piece to
Pˆx ( n ) = (1 − α ) Pˆx ( n − 1) + αx ( n ) (6)
give a flat frequency response for the 0-4 kHz frequency
is based on the first-order recursive filter with α ≈ 0.99 , and c range.
is a small constant to prevent using a large step size.
In order to test the effectiveness of the equalized music
B. Modeling of Secondary Paths signal as a training signal, white noise was replaced by the
equalized music piece in the off-line secondary-path modeling
As mentioned in the previous section, the secondary path S(z) algorithm. Figure 3 shows the estimated filter coefficients of
needs to be estimated and used in the updating process of the the secondary-path filter, S(z), obtained using musical signal
FXLMS algorithm. Estimation of S(z) is usually performed and white noise for off-line modeling. The system modeling
off-line, followed by on-line modeling [1]. Off-line capability of the chosen musical signal is as good as the white
initialization is carried out initially when a training signal is noise.
injected into both the adaptive filter Sˆ ( z ) , and the actual
secondary path S(z). This configuration becomes the
traditional adaptive system identification. It is important to
W. S. Gan et al.: Adaptive Feedback Active Noise Control Headset: Implementation, Evaluation and Its Extensions 977

20
system used is a circumaural headsets which was mounted on a
KEMAR (Knowles Electronics Mannequin for Acoustics
0 Research) mannequin. A stereo amplifier was used to amplify
the source signal being fed from the control system analyzer to
-20
the headphone loudspeaker. The error signal was picked up by
Magnitude (dB)

-40
an omnidirectional Realistic electret tie pin microphone
embedded inside the ear-pads of the headphone. The
-60 microphone signal was pre-amplified by a dual microphone
preamplifier before it reached the control system analyzer
-80 input. The sampling frequency chosen was 8 kHz, the
frequency range of interest being between 0 to 3 kHz.
-100
0 500 1000 1500 2000 2500 3000 3500 4000
Frequency (Hz)
The pinna-related measurement is taken at eight different
Figure 2. Power spectrum of the music piece from “O Fortuna from error microphone locations [5] on the median plane around the
Carmina Burana.” ear-pad of the headphone mounted on the KEMAR as shown
in Fig.4. In particular, we noted that the worst transfer function
occurs at position #2, with many peaks and nulls. The best
transfer function with the fewest peaks and nulls occurs at
position #8. The problem with transfer function with many
peaks and nulls are its difficulty in adapting to such a system
during offline modeling. Their frequency responses at position
#2 and #8 are plotted in Fig. 5.

Headphone’s
5 speaker
Headphone’s 4
shell
6
Figure 3. Estimated secondary-path filter coefficients for Sˆ ( z ) using front
music signal and white noise for off-line modeling. 3
7
After the off-line modeling, the AFANC system can be
operated without the need to continuously update the Sˆ ( z ) . 2
8
However, on-line modeling of S(z) can still be performed to Error
microphone’s
keep track on the changes of S(z) due to slight movement of
position 1
the headset. In Section 4, we will describe how on-line
modeling can be carried out using speech signal.
Figure 4. Eight different microphone locations on the right
ear-pad.
C. Ideal Position of the Error Microphone
50
The position of the error microphone used in the AFANC
headset can have a great impact on the secondary path
modeling. This is due to the fact that sound generated by the
secondary emitter (speaker) in the headset directly hits the
Magnitude (dB)

ear’s pinna, which is the external portion of the ear. The sound
from the emitter undergoes changes in its spectral content, due 0
to the filtering effects of sound hitting the pinna.

Therefore, the primary objective behind this study is to find

an optimum error microphone location inside the ear-cup for
the ANC application. In this section, we will carry out a series
-50
of measurements to identify secondary path transfer functions 0 400 800 1200 1600 2000 2400 2800 3200 3600 4000
Frequency (Hz)
at different positions. A control systems analyzer was used to
generate the random noise for estimating the poles and zeros
Figure 5. Frequency responses of S(z) at microphone locations #8 (solid
of the secondary path transfer function from the headphone line) and #2 (dotted line).
emitter (loudspeaker) to the error microphone. The headphone
978 IEEE Transactions on Consumer Electronics, Vol. 51, No. 3, AUGUST 2005

In addition, we also measured the frequency response of The DSP algorithms were implemented on a programmable
S(z) using the microphone located at the right ear cavity of the floating-point digital signal processor. This processor can
KEMAR. This measurement produces a very flat frequency handle 32-/40-bit floating-point operations. A daughter module
response, which is ideal for the development of the AFANC was used as a peripheral device for interfacing the DSP
headset. However, it is not practical and uncomfortable to processor with real-world analog signals. The sampling
insert an error microphone inside the ear canal of the listener. frequency for the 16-bit A/D converters was software
Coincidently, the measurement results [5] also shown that programmed to 8 kHz and the cut off frequency of the analog
microphone placed at the frontal positions (#6, #7, #8) anti-aliasing lowpass filter on the daughter module is 3.2 kHz.
produces the flattest response among the 8 positions. Since the The AFANC algorithm was coded in assembly language and
position #8 is very near the external auditory meatus, the debugged using the debugger.
microphone could fit in more easily inside the hollow in that
location unlike the position #6 where cartilaginous framework With the error microphone placed at the optimum
could pose a problem. Moreover, an added advantage is that at microphone location, discussed previously in Section II-C,
this location (#8), the microphone is at one of the nearest noise cancellation tests were conducted for different primary
possible positions to the ear canal. Therefore, the ideal noises [5]. The length of the filter L for off-line modeling of
position at found at position #8, which has the flattest the secondary path was fixed at 65, the step size parameter is
frequency response and at position closed to the ear canal set at 0.05; the length of the adaptive filter W(z) used for the
among the 8 positions. on-line cancellation was set at 110. The measurements are
carried out over a period of 4 seconds (or 32,000 iterations at
sampling frequency of 8 kHz).
III. REAL-TIME IMPLEMENTATION AND EVALUATION OF
PERFORMANCE B. Performance Evaluation

A. Real-Time Implementation In this section, we investigate the performance of the AFANC

using white training noise and equalized musical signal for off-
In the real-time implementation of the AFANC headset, the line secondary-path modeling. Two engine noises recorded at
error signal e(n) was picked up by an omni-directional 2,200 rpm and 3,700 rpm are used as the primary noise to test
electret microphone embedded inside the ear-pads of the the performance of the AFANC. At 2,200 rpm, there are four
headphone system. The headphone was mounted on a dominant narrowband noise components at harmonics 76 Hz,
KEMAR mannequin. The error signal was pre-amplified by a 116 Hz, 156 Hz and 196 Hz; whereas at 3,700 rpm,
dual preamplifier and sent to the DSP system. The anti-noise narrowband components are at 61 Hz, 122 Hz and 183 Hz.
y(n) generated by the DSP was amplified by a stereo amplifier, The noise reduction capability of the AFANC (using white
before it was sent to the secondary loudspeaker mounted on noise as training signal) is measured and plotted in Fig. 7 for
the headphone system. The engine noise was played from a engine noise recorded at 2,200 rpm. The average noise
Sony digital audio tape (DAT) deck, pre-amplified by an reduction over the four noise harmonics is around 15.424 dB.
amplifier and played from a speaker, which faced the In the case of 3,700 rpm, the average noise reduction over the
KEMAR, 27 cm in front and 24 cm below as shown in Fig. 6. three harmonics is around 23.27 dB.

-20

-30

-40
Magnitude (dB)

-50

-60

-70

-80
0 100 200 300 400 500 600 700 800
Frequency (Hz)

Figure 7. Power spectrum for engine noise (dotted line) and residual
error (solid line) at 2,200rpm

Instead of using white noise as the training signal, we next

examine the noise reduction performance using the equalized
Figure 6. Primary loudspeaker positioned in front of KEMAR musical source in Section II-B. The average noise reduction
for the noise harmonics in 2,200 rpm and 3,700 rpm are
W. S. Gan et al.: Adaptive Feedback Active Noise Control Headset: Implementation, Evaluation and Its Extensions 979

around 10 dB and 15.2 dB, respectively. Therefore, a slight IV. EXTENSION OF AFANC
drop of 5-7 dB is observed using the equalized musical
training signal over the white noise training signal. A. Integrated Feedback Active Noise Control

In the final performance evaluation, we compare the We can further integrate the above AFANC with the receiving
AFANC headset against a high-end commercially available audio input to form an integrated adaptive feedback active
noise canceling headset. For consistent comparison, the noise noise control (IAFANC) system [5]. In IAFANC, the residual
cancellation performances for both headsets were monitored noise picked up by the error microphone is used to synthesize
by the KEMAR microphone embedded in the KEMAR’s ear. the primary noise x(n) for updating the adaptive filter
The net noise attenuation of the proposed AFANC and the coefficients using the FXLMS algorithm. Since the IAFANC is
commercial available noise headset are shown in the bar integrated with the existing audio playback systems (such as
diagrams of Fig. 8 for 2,200 rpm and 3,700 rpm. It is shown walkman or MP3 players) or communication headset, the
that the proposed AFANC headset has very good low- microphone placed inside the ear cup picks up both the
frequency cancellation capability compared to the commercial residual noise and the desired audio signal. However, the
available headset. In addition, we also measured the passive audio components will also become the interference to the
noise cancellation (by turning off the active noise cancellation IAFANC algorithm and a method is devised to neatly combine
feature) performance of both headsets and found that both the audio and the ANC system.
headsets give a comparable and good level of high frequency
(above 2 kHz) attenuation. In this section, we introduce the detailed integration of
AFANC [6] with the audio system as shown in Fig. 9. The
error sensor output signal, e(n), contains the residual noise
plus the desired audio signal. The audio interference
(a) cancellation filter, Sˆ ( z ) , uses the audio signal, a(n), as the
30
reference signal to estimate and then remove the audio
25 components in e(n). The difference error signal, e′(n), consists
only of the residual noise, is used to update the adaptive noise
20
AFANC control filter, W(z). Note that the updating of Sˆ ( z ) using the
15
Commercial LMS algorithm can be conducted off-line using a wideband
10 musical signal as explained in Section II-B and followed by
using the audio signal as the reference input to adjust the
5
weights on-line. However, the step size used in on-line
0 modeling must be kept small so as to adapt effectively to the
76 116 156 196 small changes of S(z).
Frequency (Hz)

Acoustic d(n)
(b) domain
u(n) +
e(n)
25 x(n) y(n) _
W(z) Σ S(z) Σ
20 +
Sˆ ( z) Sˆ ( z) +
Sˆ ( z )
15 Σ
AFANC y′(n)
x′(n) copy
LMS
10 Commercial LMS -1
+
+ a(n)
Σ
5 dˆ (n) e′(n)

0
61 122 183
Frequency (Hz)
Figure 9. An integrated audio and ANC system

Assuming that the audio signal a(n) is of persistent

Figure 8. Net noise cancellation comparison for the proposed AFANC excitation and uncorrelated with the primary noise d(n), the
and a commercial noise canceling headset for engine noise at (a) 2,200
rpm and (b) 3,700 rpm secondary path S(z) can be modeled on-line by the audio
interference cancellation filter Sˆ ( z ) . After Sˆ ( z ) has
converged, we obtained the difference error and the error
respectively as
980 IEEE Transactions on Consumer Electronics, Vol. 51, No. 3, AUGUST 2005

E ' ( z ) = E ( z ) + Sˆ ( z ) A( z ) , (7) assigned as the error microphones to pick up the noise entering
the ear-cup. A communication microphone, which is located
E ( z) = D( z) − U ( z)S ( z ) . (8) external of the ear-cup, is used to pick up the near-end speech,
Substituting (8) into (7), we obtain but corrupted by the near-end noise.
E ' ( z ) = D( z ) − U ( z ) S ( z ) + Sˆ ( z ) A( z ) . (9)
Since U ( z ) = A( z ) + Y ( z ) , therefore (9) becomes In the proposed integrated ANC-communication system [7],
the residual noise picked up by the error microphone is used to
E ' ( z ) = D ( z ) − A( z ) S ( z ) − Y ( z ) S ( z ) + Sˆ ( z ) A( z ) . (10) synthesize the primary noise x(n) for updating the adaptive
filter coefficients using the FXLMS algorithm. The far-end
The optimal solution for the audio interference cancellation speech component will also become the interference to this
integrated system, and a method is developed to combine the
filter is Sˆ ( z ) = S ( z ) , therefore, (10) becomes
communication and the ANC systems.
E ' ( z ) = D( z ) − Y ( z ) S ( z ) . (11)
Equation (11) shows that the E′(z) is reduced to the residual The detailed integration of adaptive feedback ANC with
error of ANC system, where the primary noise D(z) is communication system is shown in Fig. 11. The error sensor
cancelled by the anti-noise Y(z)S(z). Note that output signal e(n) contains both the residual noise and the
desired speech signal. The speech interference cancellation
X ( z ) = E ' ( z ) + Y ( z ) Sˆ ( z ) . (12) filter Sˆ ( z ) uses the far-end speech signal a(n) as the reference
signal to estimate and then remove the speech components in
Comparing (11) and (12), resulted in D(z) = X(z) if e(n). The difference error signal, e′(n), consists of the residual
Sˆ ( z ) = S ( z ) . That is, we obtain an accurate reference signal noise only, is used to update the adaptive noise control filter
for the noise control filter W(z). W(z). Note that the update of Sˆ ( z ) using the LMS algorithm is
located at the output of the feedback ANC, and can be used for
B. IAFANC with Adaptive Noise Cancellation both off-line and on-line modeling as explained in the previous
section.
In speech communication, there is also a need to remove the
near-end noise before sending it to the far-end. As shown in It has been shown analytically in the previous section that
Fig. 10, a simple adaptive noise canceling filter H(z) with the the regenerated reference signal x(n) derived from the adaptive
LMS algorithm can be used to remove the near-end noise. The filter W(z) and the error signal, is a close estimate of the
microphone used to pick up the desired near-end signal also primary noise d(n). That is, we are able to obtain an accurate
sensed the undesired near-end noise. P(z) is the primary path reference signal for the noise control filter W(z). The noise
from the noise source to the microphones. However, a cancellation filter H(z) can be neatly integrated into the system
correlated noise input must be used to train the noise canceling by channeling the anti-noise signal y(n) into its input. The
filter H(z) using the LMS algorithm. The adaptive LMS signal y(n) is highly correlated to the near-end noise d(n), and
algorithm is given as: therefore it is a good candidate for the input of H(z). This filter
provides good cancellation of the near-end noise to produce a
hl ( n + 1) = hl ( n ) + μy ( n − l ) e ( n ), l = 0, 1, ..., L − 1 (13) cleaner near-end speech for transmission. Furthermore, no
where y(n) is correlated with the reference near-end noise. additional microphone is required to pick up the reference
noise, thus solving the crosstalk interference between
microphones.
Near-end signal Reference mic. input,
b(n) signal&noise_near We can summarize several advantages of the integrated
ANC-communication system as follows: (1) good estimation
Near-
P(z) End + ef(n) Transmit of the true residual noise e′(n) without interfering with the
Unit
Noise
noise
H(z)
- (To far end) speech signal a(n); (2) large step size can be used in adapting
source
the cancellation filter W(z) since the difference error signal
LMS e′(n) used by the FXLMS algorithm is not corrupted by the
high volume speech signal; (3) the adaptive feedback ANC
Correlated noise input, y(n) technique provides a more accurate noise cancellation since
the microphone is placed inside the ear-cup of the headset; (4)
Figure 10. Adaptive noise cancellation filter for far-end transmission the system uses single microphone per ear cup, thus produces a
compact, lower power consumption, and a cheaper solution;
This section describes the integration of the feedback ANC (5) the audio signal can be neatly used to drive both on-line
filter W(z), the secondary path modeling filter Sˆ ( z ) , and the and off-line modeling of the secondary path transfer function;
adaptive noise canceling filter H(z). Altogether, three and (6) the use of adaptive noise cancellation filter enhances
microphones are used in this integrated ANC-communication the near-end speech before sending to the far-end. The next
headset. One microphone is placed close to the emitter inside section will examine some of the performance results obtained
each ear-cup as shown in Section II-C. These microphones are using this integrated system.
W. S. Gan et al.: Adaptive Feedback Active Noise Control Headset: Implementation, Evaluation and Its Extensions 981

Near-end signal Reference mic. input,

b(n) signal&noise_near

Near- ef(n)
P(z) End + Transmit
noise Unit
Noise H(z)
- (To far end)
source

LMS

Ear-cup

d(n)
primary
noise
u(n)
x(n) y(n) - e(n)
W(z) + S(z) +
error mic.

Sˆ ( z)

Sˆ ( z) +
LMS copy
LMS -1
Sˆ ( z)
a(n), received signal Figure 12. Noise spectral for the error signals without (dotted line) and
Sˆ ( z)
from far-end with (solid line) using ANC filter under an engine disturbance.

e′(n)
+

Figure 11. An integrated communication and ANC system.

C. Performance of IAFANC with Adaptive Noise

Cancellation
Computer simulation was conducted to examine the combined
performance of the integrated ANC-communication headset.
The impulse response of the headset was measured using a
dummy head, and recorded speech and noise signals were used
in these computer simulations.
Experiment was carried out in order to evaluate the Figure 13. The top and bottom plots show the original signal corrupted
capabilities of the integrated system. An engine noise with by engine noise and enhanced near-end speech respectively
prominent harmonics at 61 Hz, 122 Hz and 183 Hz was used
as the noise source. The adaptive filters W(z), H(z), and
V. CONCLUSIONS
Sˆ ( z ) used in this simulation are 128-tap, 64-tap, and 80-tap
FIR filters, respectively. Step sizes of 0.01, and 0.05 were used In this paper, we presented several topics on the AFANC
to adapt W(z) and H(z) respectively. Off-line modeling of S(z) system. These topics include the structure of the AFANC, its
was first performed to obtain the secondary path estimation adaptive algorithm, using musical signal to model the
Sˆ ( z ) by using a wide spectrum musical signal with a step size secondary path, finding the ideal position of the error
microphone, its real-time implementation, its performance
of 0.01, and followed by on-line modeling using the speech
signal with a small step size of 0.0005. evaluation with a commercial active noise control headset, and
its extension for a clearer speech communication.
Figure 12 shows the case when the speech signal was Several important findings have been obtained which
corrupted by the engine noise. Note that the engine noise was includes: (a) optimum position of the error microphone for
recorded from the welding power generator running at 3,700 AFANC headset, (b) ability to use a wide spectrum musical
rpm. The narrowband harmonics (61 Hz, 122 Hz, and 183 Hz) signal for off-line modeling of the secondary path, (c)
were canceled by more than 30 dB. Figure 13 shows the integration of AFANC into existing audio and communication
original signal corrupted by engine noise, and the enhanced
system, (d) good performance of the AFANC headset and its
near-end signal after passing through the adaptive filter. The
performance has also been evaluate against existing
near-end speech signal was enhanced by more than 25 dB.
commercial noise canceling headset.
Overall, the strength of the AFANC headsets is its good noise
estimation using a single microphone inside each ear-cup of
the headset. Its also has a nice structure to integrate neatly to
existing speech-communication devices, like cell phones.
982 IEEE Transactions on Consumer Electronics, Vol. 51, No. 3, AUGUST 2005

References
[1] Sen M. Kuo and Dennis R. Morgan, Active noise control
BIOGRAPHY
systems: Algorithms and DSP implementations, John Wiley &
Sons, Inc., New York, 1996. Woon-Seng Gan (M ’93, SM ’00) received his B.Eng
[2] C. H. Hansen and S. D. Snyder, Active control of noise and (1st Class Hons) and PhD degrees, both in Electrical and
Electronic Engineering from the University of
vibration, E&FN Spon, London, 1997.
Strathclyde, UK in 1989 and 1993 respectively. He
[3] B. Widrow and E. Walach, Adaptive inverse joined the School of Electrical and Electronic
control, Prentice Hall, Upper Saddle River, NJ, 1996. Engineering, Nanyang Technological University,
[4] B. Rafaely, Active noise reducing headset, OSEE Online Singapore as a Lecturer in 1993. Currently, he is an
Symposium, 2001, pp 1-8. Associate Professor. His research interests include
[5] Sohini Mitra, Adaptive Feedback Active Noise Control Headset, adaptive signal processing, psycho-acoustic signal processing and real-time
MS Thesis, Northern Illinois University, DeKalb, IL, 2004. DSP implementation. He has recently co-authored a book on “Digital Signal
[6] W. S. Gan and Sen M. Kuo, “An integrated audio active noise Processors: Architectures, Implementations, and Applications”, (Prentice
control headset”, IEEE Transaction on Consumer Electronics, Hall, 2005).
Vol. 48, No. 2, May 2002, pp 242-247.
[7] W. S. Gan and Sen M. Kuo, “Integrated Active Noise Control Sohini Mitra received her MSEE degree from Northern Illinois University in
Communication Headset”, In Proc. of International Symposium 2004. She is currently with Motorola Inc. IL.
on Circuits and Systems, May 25 - 28, 2003, pp. IV353-356.

Sen M. Kuo (M ’85, SM ’04) received the B.S. degree

from National Taiwan Normal University, Taipei,
Taiwan, in 1976 and the M.S. and Ph.D. degrees from
the University of New Mexico in 1983 and 1985,
respectively. In 1993, he was with Texas Instruments,
Houston, TX. He is currently a Professor and Chair at
the Department of Electrical Engineering, Northern
Illinois University, DeKalb, IL. He is the author of
Active Noise Control Systems: Algorithms and DSP Implementations (New
York: Wiley, 1996), Real-Time Digital Signal Processing (Wiley, 2002),
Digital Signal Processors: Architectures, Implementations, and Applications
(Prentice Hall, 2005), and of numerous technical papers. He has been
awarded seven US patents. His research focuses on active noise and vibration
control, adaptive echo and noise cancellation, digital audio applications, and
digital communications.

View publication stats