Note: Descriptions are shown in the official language in which they were submitted.
CA 02397080 2002-08-07
1
SUB-BAND ADAPTIVE SIGNAL PROCESSING IN AN OVERSAMPLED
FILTERBANK
FIELD OF THE INVENTION
The present invention relates to signal processing and more specifically to a
method and a system for adaptive signal processing.
BACKGROUND OF THE INVENTION
A conventional approach in the signal processing applications listed above is
to
use a time domain approach, where a filterbank is not used, and a single
adaptive filter
acts on the entire frequency band of interest. This single time domain filter
is typically
required to be very long, especially when applied to acoustic echo
cancellation.
Computational requirements are a concern because longer filters require
exponentially
increasingly more processing power (i.e., doubling the filter length increases
the
processing requirements by more than two). A longer filter typically requires
more
iterations by its adaptive controlling algorithm to converge to its desired
state. In the
case of an adaptive noise cancellation algorithm, slow convergehce hampers the
ability
of the system to quickly reduce noise upon activation and to track changes in
the noise
environment.
In summary, the problems with time domain adaptive signal processing are: 1)
Long filters are required - cannot interleave the update of multiple filters.
2) Slower filter
convergence due to longer filter length, 3) Performance problems in the
presence of
coloured noise, and 4) Inability to set varying algorithm parameters for
individual
frequency bands.
Solutions to problems in time domain adaptive signal prooessing arising from
coloured noise and a long filter are limited. A long filter is often a
requirement that is
dictated by the particular application, and shortening it would degrade
performance. In
cases when it is allowable, white noise can be inserted into the signal path
to allow the
filter to adapt quicker.
CA 02397080 2002-08-07
2
Slow convergence is usually dealt with by choosing algorithm parameters that
result in fast convergence while still guaranteeing filter stability. In the
Least Mean
Squares (LMS) algorithm this is done by increasing the step-size parameter
(mu).
However, this approach causes considerable distortion in the processed output
signal due
to the larger fluctuations of the adaptive filter resulting from a high mu
value.
A method used to increase computational speed in time domain signal processing
is to perform operations in the Fourier transform domain (see J. J. Shynk,
"Frequency
Domain and Multirate Adaptive Filtering", IEEE Signal Processing Magazine,
vol. 9, no.
1, pp.15-37, Jan 1992). A section of the signal is transfon.ned, operated on,
then
undergoes an inverse transformation. Methods are well known for performing
specific
operations in the transform domain that directly correspond to linear
convolution (a
common operation) in the time domain, but require less processing time. The
added
requirement of having to calculate the Fourier transform and inverse Fourier
transform is
offset when the signal can be transformed in blocks that are sufficiently
large.
SUMMARY OF THE INVENTION
The invention seeks, through the use of WOLA filterbanks and other components,
to alleviate these and other problems found in prior art implementations. In
doing so,
cost-effective solutions are achieved. Each of the shortcomings of the earlier
technologies is addressed in turn.
In accordance with an aspect of the present invention, there is provided an
adaptive signal processing system for improving a quality of a signal, which
includes: a
first analysis filterbank for receiving a primary information signal in the
time domain and
transforming the primary information signal into a plurality of oversampled
sub-band
primary signals in the frequency domain; a second analysis filterbank for
receiving a
reference signal in the time domain and transforming the reference signal into
a plurality
of oversampled sub-band reference signals numerically equal to the number of
primary
signal sub-bands; a plurality of sub-band processing circuits for processing
these signals
CA 02397080 2002-08-07
3
to improve a quality of an output signal; and a synthesis filterbank for
combining the
outputs of the sub-band processing circuits to generate the output signal.
The oversampled WOLA filterbanks also address the problems with traditional
FFT-based sub-band adaptive filtering schemes. WOLA filterbank processing is
described in US Patent No. 6,236,731 for hearing aid applications. These
problems
include highly overlapped bands that provide poor isolation, and lengthy group
delay.
In addition, oversampled WOLA filterbank processing also provides the
following advantages for sub-band adaptive signal processing: 1) Programmable
power
versus group delay trade-off, adjustable oversampling, 2) Stereo analysis in a
single
WOLA, 3) Much greater range of gain adjustment in the bands, and 4) The use of
complex gains.
An oversampled WOLA filterbank sub-band adaptive system can also be
implemented on ultra low-power, miniature hardware using the system described
in US
Patent No. 6,240,192 (Schneider and Brennan).
Through the use of the oversampled WOLA filterbank, the single time domain
filter can be replaced by a plurality of shorter filters, each acting in its
own frequency
sub-band. The oversampled WOLA filterbank and sub-band filters provide equal
or
greater signal processing capability compared to the time domain filter they
replace - at a
fraction of the processing power.
Utilising the oversampled WOLA filterbank results in faster convergence and
improved overall effectiveness of the signal processing application.
Yet another benefit of sub-band adaptive signal processing in an oversampled
filterbank is referred to as the "whitening" effect [see W. Kellermann.
"Analysis and
design of multirate systems for cancellation of acoustical echoes."
Proceedings IEEE
International Conference on Acoustics, Speech, and Signal Processing, pp. 2570-
2573,
New York, NY, USA, April 1988. A white signal has a flat spectrum; a coloured
signal
has a spectrum that significantly varies with frequency. The WOLA filterbank
CA 02397080 2002-08-07
4
decomposes coloured input signals into sub-band signals with spectra that are
"whiter"
than the wide-band signal. Due to oversampling, the whitening effect occurs in
only part
of the spectrum; however, this behaviour is predictable and uniform across all
bands and
can therefore be compensated for by emphasis filters (described hereafter).
The
commonly used least-mean-square (LMS) algorithm for adaptive signal processing
performs best with white signals [Haykin, Simon. Adaptive Filter Theory
Prentice Hall.
1996]. Thus, the whitening effect provides a more ideally conditioned signal,
improving
system perfonnance.
Yet another benefit of sub-band adaptive signal processing in an oversampled
filterbank is the ability to set varying algorithm parameters for individual
frequency
bands. For example, a noise cancellation algorithm can have filters that are
set up to
converge at different rates for different sub-bands. In addition, the adaptive
filters can
have different lengths. The increased number of possible parameters allows the
system
to be more effectively tuned according to the requirements of the application.
In situations in which processing power is limited or must be conserved, the
update of the adaptive filter groups can be interleaved. Thus, although an
adaptive filter
may be occasionally skipped in the update process but it will still be updated
at periodic
intervals. This is in contrast to the situation of a single time domain filter
where the
processing cannot be split across time periods in this way.
Although some solutions have utilised some degree of oversampling- less than
two times - (see M. Sandrock, S. Schmitt. "Realization of an Adaptive
Algorithm with
Sub-band Filtering Approach for Acoustic Echo Cancellation in
Telecommunication
Applications". Proceedings of ICSPAT 2000), they do not provide the low group
delay,
flexibility in power versus group delay trade-off and excellent band isolation
of
oversampled WOLA based adaptive signal processing.
The following are some of the combined advantages of adaptive signal
processing
using oversampled WOLA filterbank compared to earlier techniques: 1) Very low
group
delay, 2) A flexible power versus group delay trade-off, 3) Highly isolated
frequency
CA 02397080 2002-08-07
bands, 4) Wide-ranging band gain adjustments, 5) Variable algorithm parameters
in
different sub-bands (filter length, convergence rate, etc; algorithm
parameters can be
optimally adjusted to meet computation as well as other performance
constraints), 6)
Faster convergence of adaptive filters, 7) Reduced computation time, 8)
Improved
5 performance in coloured noise, and 9) Ability to split computational load
associated with
updating adaptive filters across multiple time periods.
A further understanding of the other features, aspects, and advantages of the
present invention will be realized by reference to the following description,
appended
claims, and accompanying drawings.
Brief Description of the Drawings
Embodiments of the invention will now be described with reference to the
accompanying drawings, in which:
Figure 1 shows a signal path through the oversampled WOLA filterbank
operating in mono mode;
Figure 2 shows a signal path through the oversampled WOLA filterbank
operating in stereo mode;
Figure 3 shows a block diagram of a time-domain adaptive noise cancellation
system;
Figure 4 shows a block diagram of a frequency-domain adaptive noise
cancellation system;
Figure 5 is a schematic diagram showing a spectral emphasis operation;
Figure 6 shows the signal flow of the LMS,,, block when a spectral emphasis
filter
is used;
Figure 7 shows a block diagram of a two-microphone Wiener noise cancellation
system;
CA 02397080 2002-08-07
6
Figure 8 shows a block diagram of a sub-band adaptive acoustic echo
cancellation
system with the oversampled WOLA filterbank;
Figure 9 shows a processing block for the sub-band adaptive acoustic echo
cancellation system with the oversampled WOLA filterbank using LMS;
Figure 10 shows a block diagram of an oversampled WOLA filterbank processing
system using a microphone array for the primary signal;
Figure 11 shows a block diagram of a WOLA filterbank processing system with
multiple reference inputs using LMS; and
Figure 12 shows a sub-band processing block for WOLA filterbank processing
system with multiple reference inputs using LMS.
Detailed Description of the Preferred Embodiment(s)
Figure 1 shows the signal path through a basic oversampled WOLA filterbank
system operating in mono mode. The signal from the Microphone 100 passes
through a
preamplifier 102 to an analog to digital converter 104. The resultant digital
signal output
by the converter is passed into the analysis filterbank 106 that is programmed
to divide
the signal into sub-bands. Each sub-band is then passed to one of the
Processing Blocks
108 whose outputs are combined by the Synthesis Filterbank 110 into a single
digital
signal that is passed in turn to a digital to analog converter 112 to produce
an analog
output 114. Similarly Figure 2 shows the signal path through a basic
oversampled
WOLA filterbank system operating in'stereo mode', although in this case the
term is
somewhat misleading, since although there are two inputs to the system, there
is only one
output. The signals from the two microphones 200, 202 each pass through
respective
preamplifiers 201, 203 to respective analog to digital converters 204, 206.
The resultant
digital signal outputs from the converters are passed into the analysis
filterbank 208 that
is programmed to divide each signal into a number of sub-bands. Each sub-band
is then
passed to one of the Processing Blocks 210 whose inputs are the equivalent sub-
bands of
both inputs, and whose outputs are combined by the Synthesis Filterbank 212
into a
CA 02397080 2002-08-07
7
single digital signal that is passed in turn to a digital to analog converter
214 to produce
an analog output 216. In both cases, the logic contained in the processing
blocks is
dependent on the particular application. For sub-band adaptive signal
processing, these
blocks contain adaptive filters and their associated control logic.
The type of filters (recursive or non-recursive), method of controlling the
adaptive filters, and number of inputs (one or many) ran vary. The LMS
algorithm and
its variants are widely used in adaptive signal processing for their relative
simplicity and
effectiveness. Many applications use the two-input stereo configuration, but
sub-band
adaptive signal processing with one or many inputs is also within the scope of
this
invention. Furthermore, this invention is not limited to any particular
configuration of
the oversampled WOLA filterbank (i.e., number of sub-bands, sampling rate,
window
length, etc).
The WOLA filterbank provides an input to each sub-band adaptive processing
block that is highly isolated in frequency. The sub-band adaptive processing
blocks may
have independent adaptive parameters, or they may be grouped into larger
frequency
bands and share properties.
After adaptive processing, the modified sub-band signals are sent to the
synthesis
filterbank, where they are recombined into a single output signal. The net
effect of the
sub-band adaptive filters on this output signal is equal to a single time
domain filter that
is much longer than any one of the sub-band filters.
United States Patent 6,236,731 "Filterbank Structure and Method for Filtering
and Separating an Information Signal into Different Bands, Particularly for
Audio Signal
in Hearing Aids" by R. Brennan and T. Schneider, incorporated herein by
reference,
discloses the WOLA filterbank signal processing. A brief summary of that
patent is
included in an Appendix A attached hereto for convenience.
A description of two prefened embodiments of the present invention follows.
Both described embodiments are for noise cancellation applications. This is a
typical
application of adaptive oversampled WOLA processing, but the present invention
is not
CA 02397080 2002-08-07
8
limited thereto. The first preferred embodiment is a sub-band noise
cancellation
algorithm that uses a variant of the LMS algorithm, and the oversampled WOLA
filterbank in stereo mode. The second preferred embodiment also performs noise
reduction with a two-microphone configuration and an alternative method for
deriving
the adaptive coefficients.
In a first preferred embodiment a sub-band noise cancellation system is
described
that uses a variant of the LMS algorithm together with an oversampled WOLA
filterbank
operating in stereo mode. Although least-mean squares signal processing is
described
here, other techniques well known in the art are also applicable. For example,
recursive
least squares can also be used.
The LMS algorithm is typically used to cancel the noise in transmitted speech
when the speaker is located in a noisy environment. The listener, not the
speaker,
experiences the improvement in signal quality. Examples of where is algorithm
can be
used include telephone handsets, and boom-microphone headsets. This algorithm
is
useful for all headset styles that use two microphones for speech
transmission. The
algorithm can be applied to other applications as well. For example, one
skilled in the art
may modify this algorithm for acoustic echo cancellation or acoustic feedback
cancellation.
Two-microphone adaptive noise cancellation works on the premise that one
signal contains noise alone, and the other signal contains the desired signal
(in this case
speech) plus noise that is correlated with the noise in the first signal. The
adaptive
processing acts to remove the correlated elements of the two signals. Since
the noise
signals are (assumed to be) correlated and the speech is not, the noise is
removed.
Figure 3 shows a block diagram of a time-domain, two-microphone adaptive
noise cancellation system. A first microphone 301, which is arranged to pickup
the
wanted signal, passes its signal, which includes a noise component from the
acoustical
environment, to a Voice activity detector (VAD) 306 and a summer 310. A second
microphone 302 which is arranged to pick up mainly the noise of the acoustical
CA 02397080 2002-08-07
9
environment, passes its signal, which might include an attenuated version of
the wanted
signal, to the VAD 306, to an LMS processor 308, and to an adaptive Finite
Impulse
Response (FIR) filter 304. An LMS processor 308 uses the output from the VAD
306 as
well as the output from the second microphone 302 to control the adaptive FIR
filter 304
in order to minimize the noise appearing at the system output. The voice
activity
detector (VAD) 306 is used in some embodiments to stop or slow down adaptation
when
speech is present. This reduces obtrusive artefacts in the output signal 312
that are caused
by misadjustments of the FIR filter due to the presence of speech. The VAD 306
typically uses both microphone signals 301, 302 as inputs and may employ the
differential level as an indicator that speech is present, or it may use any
one of a variety
of more complex techniques. In a typical application, the microphone 301 faces
the
talker, and therefore will receive a higher level wanted signal that
microphone 302 which
is placed at some distance from the speakers mouth, but arrange to ensure a
similar level
of acoustical noise is received. For example, in a headset application, the
two
microphones could be located on a boom with the first microphone 301 facing in
and the
second microphone 302 facing out.
In a further embodiment, the algorithm is implemented in the frequency domain.
Figure 4a shows a block diagram of such a system. In this case, again, two
microphones
are used. A first microphone 401, which is arranged to pickup the wanted
signal, passes
its signal ('signal + noise'), which includes a noise component from the
acoustical
environment, to a Voice activity detector (VAD) 408 and to a first analysis
filterbank
404. A second microphone 402 which is arranged to pick up mainly the noise of
the
acoustical environment, passes its signal ('noisa-only'), which might include
an
attenuated version of the wanted signal, to the VAD 408, and to a second
analysis
filterbank 405. Each filterbank is arranged to provide an equal number of sub-
bands
derived from the incoming signals. These sub-band outputs are passed in turn
to a like
number of sub-band processing blocks 410, 412, 414, each of which uses a sub-
band
from the first analysis filterbank 404, as an input and the equivalent sub-
band from the
second analysis filterbank 405 as one adaptive or controlling input. Other
controlling
inputs are possible in further preferred embodiments. Thus the processing of
the signal is
CA 02397080 2002-08-07
achieved in a number of sub-bands, each with a complex output signal
(magnitude and
phase), and each requiring much less processing than would be required for the
whole
band, the total processing being less than that required if the full band were
to be
processed at once. Again, the voice activity detector (VAD) 408 is used in
some
5 embodiments to stop or slow down adaptation when speech is present. Such a
frequency
domain implementation offers better performance than a time-domain
implementation
because it converges faster and, because of its sub-band operation, implements
longer
adaptive filters in an efficient manner. Interleaved or decimated updates are
used in
some embodiments to further reduce the computational load. Also, noise
rejection for
10 frequency-localized noise is likely to be improved.
Each of the sub-band processing blocks 410, 412, 414 implements what is well
known in the art as the leaky normalized LMS algorithm. In a sub-band
implementation
of the leaky normalised LMS algorithm the LMS step-size can possibly vary in
each sub-
band; lower sub-bands contain high speech content and have a smaller ste"ize,
while
higher sub-bands can be more aggressively adapted with a larger step-size due
to
relatively low speech content. A typical sub-band processing block is shown in
more
detail in Figure 4b. In the figure, the sub-band outputs of the two analysis
filterbanks =
404, 405 are shown. The 'signal + noise' component from filterbank 404 is
passed
directly to a summer 440. The 'noise-only' component from filterbank 405 is
passed to
both a FIR filter 430, and to an LMS filter 435. The output of the LMS filter
435 is used
to adapt the response of the FIR filter 430. In some embodiments, a fraction
of the output
signal from the summer 440 is fed back and used as a further input to the LMS
filter 435.
In a further preferred embodiment, the leaky normalised LMS algorithm is
supplemented by a spectral emphasis filter. This additional filter is static
and serves to
whiten the LMS input signals for faster convergence. Oversamplingin
filterbanks such
as those shown in the Figure 4, 404, 405, inherently produces subband signals
that are
coloured in a predictable way. In the case of two times oversampling, the
bottom half of
the sub-band spectrum has relatively high energy and is relatively flat
compared to the
upper half of the spectrum, which contains very little energy. The spectral
emphasis
CA 02397080 2002-08-07
11
filters amplify the part of the spectrum known to have lower energy, thus the
signal is
modified towards the ideal case of being white.
Figure 5 illustrates the effect of the spectral emphasis operation. The
oversampled
input signal shown in 501 has a drop off in energy towards high frequencies,
and the
emphasis filter response 503 is designed to amplify the high frequencies. The
filtering
operation results in a signal spectrum that is flatter 505, or a process known
as
'whitening'.
Figure 6 shows the signal flow of a typical sub-band processing block of
Figure
4a, incorporating the spectral emphasis filter of Figure 5. In this block,
both the'signal +
noise' 601 and the 'noise only' 602 inputs are filtered and whitened by
emphasis filters
606, 607 before they are used by the LMS block 610 to update a secondary
Finite
Impulse Response (FIR) filter 620. It is not desirable to have a synthesis
filterbank
output signal that has been noticably emphasized in some frequency regions,
since the
perceived signal is then somewhat distorted. To avoid this distortion, the
coefficients that
define the secondary filter 620 are copied to the FIR filter 630 used on the
unemphasized
noise signa1602 to generate the signal to be synthesized. The output of this
FIR filter 630
is then summed 640 with the 'signal + noise' signa1601 to produce the output
signal 650
which is later assembled by the synthesis filterbank 420 of Figure 4 to
produce the
required audio signal.
The design of the emphasis filter is dependent on the oversampling factor used
in
the WOLA filterbank. Given the oversampled WOLA filterbank parameters, the
spectral
properties of the sub-band signals can be determined, and an appropriate
emphasis filter
can be designed. It can be implemented as a FIR filter or an infinite impulse
response
(IIR) filter.
A further preferred embodiment of the invention is describe in the context of
a
transmit algorithm based on Wiener noise reduction technique which is well
known in
the art. Again this algorithm is useful for all headset styles that use two
microphones for
speech transmission. This embodiment uses the stereo processing mode of the
WOLA
CA 02397080 2002-08-07
12
filterbank. Two signals are simultaneously transformed to n sub-bands in the
frequency
domain: one is 'signal + noise', the other is 'noise only'. The processing
acts to remove
the noise that is correlated between the two signals. Figure 7 shows a block
diagram of
this processing. Again, for convenience the action of the various components
between the
analysis filterbanks704, 705 and synthesis filterbank 730 is described in
terms of a single
sub-band, although there will typically be a number of sub-bands. The outputs
of the
'signal + noise' microphone 701 and the noise only microphone 702 are passed
to two
analysis filterbanks 704, 705 respectively. The sub-band outputs of the signal
+ noise
filterbank 704 are each modified by a summer 720 before being assembled by the
synthesis filterbank 730 to produce the required output 735. For each of the
sub-bands of
the 'signal + noise' signal, an equivalent sub-band of the 'noise only' signal
is processed
by filter W 710, controlled by a Least Squares block 712 whose inputs are the
sub-bands
from the 'signal + noise' filterbank 704 and a fraction of the appropriate
summer result
720. The overall aim of the algorithm is to minimize E2 in the expression:
E2 = (YW-X)(YW-X)*
Where E is the level of the sub-band input to the synthesis filterbank 730, X
is the
level of the signal + noise sub-band output by the analysis filterbank 704, Y
is the level
of the noise in the same sub-band output by the analysis filterbank 705, and W
is the
function of the filter block 710. The * operator denotes complex conjugation.
The algorithm is next discussed in some detail. The solution that minimizes
EZis
the equation:
W=rxy/RX, (1)
where R, is the auto-correlation matrix of X and rXY is the cross-correlation
matrix of X
and Y (see M. H. Hayes. Statistical Digital Signal Processing and Modeling.
John
Wiley & Sons, Inc. 1996, pages 337-339).
If k and rxY are estimated using only the most recent sample of X and Y, the
value of adaptive weight Wk at time index n is
CA 02397080 2002-08-07
13
Wh(n) = Yk(n) /Xk(n),
where k is the sub-band index.
Thus, update of an adaptive weight only requires division of the complex
values
Yk(n) and Xk(n). Taking one-sample estimates of the auto-correlation and cross-
correlation matrices eliminates the need to perform the matrix inversion of
kin equation
(1).
A novel addition to this algorithm is the use of frequency constraints. If
left
unconstrained, adjacent bands may have very different gains. While this will
result in the
lowest noise level (since Ez will be minimized), it may also result in some
undesirable
processing artifacts giving rise to a lessening in perceived quality of the
signal.
Constraining the adjustment of the gain vector (W) results in less noise
reduction, but
fewer artifacts. Equation (2) defines a scheme where the gain in a given band
is
constrained by the two adjacent bands. Note that this case uses only a single
(complex)
weight per band. It is possible to extend this scheme to allow for multiple
weights per
band. For the single gain case, the matrix is block-diagonal; thus, there are
efficient
solution methods.
Y, YZ 0... W, X,
Y, Y2 Y, 0 Wz _ XZ
(2)
0 Yz Y, Y,
0 0 ... Wk Xk
Multi-microphone Wiener algorithms like this have been successfully used for
noise reduction in other applications; for example, seeMulti-Channel Spectral
Enhancement In a Car Environment Using Wiener Filtering and Spectral
Subtraction,
Meyer and Simmer, Proc. ICASSP-97, Vol. 2, pp. 1167-1170.
A yet further preferred embodiment of the invention is use in an echo
cancellation
system. The goal of acoustic echo cancellation is to remove the far end
speaker's voice
from the signal that enters the near end microphone and eventually reaches the
CA 02397080 2002-08-07
14
loudspeaker at the far end. This allows the near end speaker's voice to be
transmitted
without echoes of the far end speaker's voice (caused by room reverberation),
for better
intelligibility and less listening effort. An adaptive signal processing
system must deal
with a significantly long room response. A single time domain filter
implementation
would typically contain thousands of coefficients to adequately model this
response, with
consequently high processing power requirements. The use of the present
invention to
implement an LMS algorithm is used to control the adaptive filters, as
illustrated in
Figure 8, allows for shorter filters and therefore a savings in processing
power over the
traditional time domain approach. A shown in Figure 8, the far end speaker
(person)
makes use of a microphone 801 and receiver (loudspeaker) 802 which are
connected to
the near end speaker (person) who also has a receiver (loudspeaker) 803 and a
microphone 804. Because of the typical acoustic properties of rooms in which
such
systems are used, some fraction of the sound emitted by the receiver 803
inevitably
enters the microphone 804 at the near end. Of course the same is true if the
far end uses a
similar receiver and microphone system, but this discussion will be restricted
to a simpler
configuration where only one end is so arranged. In a generalised system, to
mitigate the
problems caused by this inadvertent signal path, a system comprising analysis
filterbanks
806, 812 and a synthesis filterbank 808 with sub-band processing blocks 810
interposed
is used. The behaviour of the sub-band processing blocks conforms
substantially to the
algorithm described above.
Figure 9 shows in more detail a sub-band processing block 810 of Figure 8. In
each of these blocks, the input 901 passed from an analysis filterbank is
summed with a
signal derived from a further input 902 passed from another analysis
filterbank to provide
an output 907 from which the unwanted signal is substantially removed. The
processing
of the input 902 is performed using an LMS filter 901, whose inputs are the
modified
output 907 and the input 902. The output of the LMS filter 901 is used to
adjust and
adapt the characteristics of a FIR filter 912 which processes the input 902
and passes the
result to the summer 905. As may be expected, the configuration is much like
the noise
cancellation system described earlier, but in this case the far end speech is
considered to
be the unwanted noise, and the desired output signal is the near end speech.
CA 02397080 2006-03-13
The previously described embodiments are examples of adaptive sub-band
adaptive signal processing with two inputs. It should be noted that they could
be
extended to make use of a multiplicity of inputs. A microphone array could be
used to
capture several input signals, all of which are summed to form the primary
(i.e. signal
5 plus noise) signal. Also, in some situations there are several noise sources
to be
cancelled, therefore a multiplicity of noise censors are required for the
reference (i.e.
noise) signals.
Time domain adaptive algorithms with more than two inputs signals are well
known in the art. The benefits of sub-band adaptive signal processing over
time domain
10 adaptive signal processing still hold for these applications (see US Patent
Application
Publication No. 2003/0063759, entitled "Directional Audio Signal Processing
Using An
Oversampled Filterbank").
In a further embodiment of the invention a microphone array is used as the
source
of the primary signal composed of the signal-of-interest and noise, and a
reference
15 microphone collects the environment noise, substantially free of the signal-
of-interest. In
other respects the system is the same as in earlier embodiments. Figure 10
illustrates the
signal flow. An array of primary microphones 1001, which pickup the signal of
interest
with noise, are connected to a first preamplifier 1002 and associated first
analog-to-
digital converter 1003 which passes its output to a first analysis filterbank
1010. A
reference microphone 1005, which picks up the noise, is connected to a second
preamplifier 1006 and its associated second analog-to-digital converter 1007
which
passes its output to a second analysis filterbank 1020. The sub-bands derived
by the
analysis filterbanks 1010, 1020 are passed to processing blocks 1030. The
action of the
processing blocks may be any of the previous noise cancelling or noise
reduction
strategies. The processing blocks 1030 then pass their outputs to the
synthesis filterbank
1040 whose output is converted to analog by a digital-to-analog converter 1050
to
produce the required output 1060.
A further embodiment of the invention uses multiple reference microphones,
each
with an analysis filterbank, together with processing making use of the LMS
algorithm.
CA 02397080 2002-08-07
16
This type of configuration is used in a noise cancellation application when
there
are more than one noise source. One microphone is used for each noise source
to
provide a reference signal, which is adaptively filtered and then subtracted
from the
primary signal. Figure 11 illustrates the signal flow and Figure 12 shows the
detail of
each processing block in Figure 11. Referring first to Figure 11, a primary
microphone
1101, arranged to pick up substantially the signal-of-interest, but which also
picks up
environmental noise from n discrete sources is connected to a first analysis
filterbank
1130 through its associated preamplifier 1102 and analog-to-digital converter
1103. Note
that, although in the figure and following description, three reference
microphone 1110,
1115, 1120, arranged to pick up one of the three substantially independent
noise sources
are shown with their associated components, this number may be fewer or larger
as
required to cover the number of discrete noise sources identified. Each
reference
microphone 1110, 1115, 1120, is connected to an associated analysis filterbank
1132,
1133, 1134 respectively through their respective preamplifiers 1111, 1116,
1121 and
analog-to-digital converters 1112, 1117, 1122. Each sub-band generated from
the signal
of interest (first) analysis filterbank is passed to one of a number of
processing blocks
1140, which will be describe below, and the outputs of the processing blocks
1140 are
combined by the synthesis filterbank 1150 whose output is passed to a digital
to analog
converter 1160 to produce the desired, substantially noise-free output 1170.
Turning now to Figure 12, the processing blocks 1140 of Figure 1 I are
described
in more detail. Each processing block accepts one sub-band derived from the
signalof-
interest filterbank 1201, and this is then mixed or summed by a first summer
1230 with
the results of processing the noise signals and then ouput 1240 to the
synthesis filterbank
1150 of Figure 11. Processing of the noise signals proceeds as follows: each
appropriate
sub-band output 1205 from the analysis filterbanks 1132, 1133, 1134 of Figure
11 is
passed to the input of a FIR filter, 1216, 1217, 1218 respectively and to an
LMS
controller 1210 which also receives the output of the first summer 1230. The
FIR filters
1216, 1217, 1218 are controlled by the outputs of the LMS controller 1210. The
outputs
of the FIR filters 1216, 1217, 1218 are summed in a second summer 1220, before
the
result is applied to the first summer 1230.
CA 02397080 2002-08-07
17
The system removes from the primary signal (or signal of interest) the
component of the
primary signal which is correlated to the reference signal (or noise).
Appendix B attached hereto includes some details of an example algorithm for
use in the present invention.
While the present invention has been described with reference to specific
embodiments, the description is illustrative of the invention and is not to be
construed as
limiting the invention. Various modifications may occur to those skilled in
the art
without departing from the true spirit and scope of the invention as defined
by the
appended claims.
CA 02397080 2002-08-07
18
Appendix A
Summary from United States Patent 6,236, 731 "Filterbank Structure and Method
for Filtering and Separating an Information Signal into Different Bands,
Particularly for
Audio Signal in Hearing Aids" by R. Brennan and T. Schneider
In accordance with the first aspect of this earlier invention, there is
provided ai
oversampled filterbank for filtering an information signal, the filterbank
having a
filterbank structure comprising a filter means defining a filter bandwidth,
said filter
means filtering said information signal and separating said information signal
inb a
plurality of frequency band signals each representing one of a plurality of
uniformly
spaced frequency bands within said filter bandwidth, said frequency bands
being stacked
in one of an even and an odd manner and said frequency bands overlapping, such
that the
summation of the unmodified frequency hand responses of the plurality of said
frequency
bands sums to a function within a predetermined passband ripple over said
filter
bandwidth, wherein the filter means includes a selection input enabling at
tmst one of the
following to be selected:
(i) the number of frequency band signals,
(ii) the bandwidth of said frequency bands,
(iii) selection of stacking of said frequency bands in one of an even and an
odd
manner,
(iv) the degree of overlap between said frequency bands,
(v) an oversampling factor by which said frequency band signals are sampled
above the theoretical minimum of critical sampling.
The filterbank can be configured to enable one or more of usual parameters of
a
digital filterbank to be adjustable, and these can include: the number of
bands; the width
of each band; whether the bands have abutting band edges, overlap or are
spaced apart;
coefficients for both analysis and synthesis windows; whether there is any
relationship
CA 02397080 2002-08-07
19
between the analysis and synthesis windows; even or a odd stacking of bands;
and the
degree of oversampling above the critical sampling rate.
Preferably, the selection input enables at least one of the number of
frequency
bands and selection of stacking of said frequency bands in one of an even and
an odd
manner to be selected, said number of frequency bands being equal to N, and
the filter
means comprises: (a) a first analysis filterbank means for separating said
signal into the
plurality of N separate frequency band signals; (b) processing means for
receiving and
processing each of said separate frequency band signals to provide N separate
processed
frequency band signals; and (c) a second synthesis filterbank means for
receiving and
recombining the N separate processed frequency band signals into a single
output signal,
wherein both of the first analysis filterbank means and the second synthesis
filterbank
means are connected to the selection input, the processing means being coupled
between
the first analysis filterbank means and the second synthesis filterbank means.
In another aspect of the earlier invention, the filterbank comprises a
dedicated
application specific integrated circuit (ASIC), said ASIC including the first
analysis and
the second synthesis filterbanks, and a programmable digital signal processor
for
controlling the number of frequency bands and the bandwidth of each frequency
band,
said digital signal processor being provided with the selection input.
The filterbank may be adapted to receive a single real monaural information
signal, wherein said transform means generates non-negative frequency band
signals and
negative frequency band signals, said negative frequency band signals being
derivable
from the non-negative frequency band signals, and said processing means
processes only
said non-negative frequency band signals. Alternatively is adapted to filter
an audio
signal comprising first and second real monaural information signals which are
combined
into a complex stereo signal and wherein said transform means generates N
combined
frequency band signals, and wherein said processing means includes: (a)
channel
separation means for separating the N combined frequency band signals into the
N
frequency band signals corresponding to said first information signal and the
N
frequency band signals corresponding to said second information signal, each
of said N
CA 02397080 2002-08-07
frequency band signals comprising non-negative and negative frequency band
signals;
(b) first independent channel processing means connected to the channel
separation
means for receiving and processing each of said separate frequency band
signals of said
first information signal to provide a first set of N separate processed
frequency band
5 signals; (c) second independent channel processing means connected to
channel
separation means for receiving and processing each of said separate frequency
band
signals of said second information signal to provide a second set of N
separate processed
frequency band signals; and (d) channel combination means connected to the
first and
second independent channel processing means for combining said first set of N
processed
10 separate frequency band signals and said second set of N processed separate
frequency
band signals.
In accordance with another aspect of the earlier invention, there is provided
a
method of processing an information signal to selectively modify different
frequency
bands, the method comprising the steps of: (1) defining a filter frequency
bandwidth to
15 be analyzed; (2) dividing the filter frequency bandwidth into a plurality
of uniformly
spaced bands, said frequency bands being stacked in an even or odd manner and
said
frequency bands abutting, overlapping, or being spaced apart from one another;
(3)
filtering the information signal to separate the signal into a plurality of
frequency band
signals, each representing one of said uniform filter bands; (4) processing
the frequency
20 band signals; (5) recombining the signals of the individual bands to form
an output
signal; and (6) providing an input for enabling at least one of the following
to be
selected: (i) the number of frequency band signals, (ii) the bandwidth of said
frequency
bands, (iii) whether said frequency bands are stacked in an even or odd
manner, (iv)
whether said frequency bands abut, overlap, or are spaced apart from one
another, and (v)
a decimation factor by which said frequency band signals are downsampled.
In another aspect the method of the earlier invention includes transfonming
the
information signal into the frequency domain, providing N separate frequency
band
signals in the frequency domain, and effecting an inverse transform of the N
separate
processed frequency band signals into the output signal in the time domain
CA 02397080 2002-08-07
21
Appendix B: Algorithm Description
This is a brief description of a sample algorithm for use with the present
invention.
An input signal x contains the desired speech and some additive noise. A
second input
signal y contains just the additive noise. Each signal will be filtered in
slightly different
ways before reaching the adaptive filtering due to spatial and physical
differences in the
transducers and equipment used to capture them. Ideally, if this filtering did
not occur
then one could simply subtract y from x to recover the speech signal. Because
of the
unknown filtering, a new filter W must somehow be determined to transformy
such that
it matches the noise in x. Applying this filter toy and then performing the
subtraction
will yield a clean speech signal.
A WOLA filterbank provides the frequency domain representation of the signals
necessary to compute the sub-band adaptive filter. The WOLA produces N=2
complex
frequency domain results for each signal from N point Fast Fourier Transforms
FFTs of
the incoming signal frames. The goal of the algorithm is to perform a least-
squares fit of
the WOLA output for both signals on a sub-band-by-sub-band basis. That is, the
fit is
computed independently for each band of the filterbank. Put mathematically,
the least-
squares fit ( Xk ) attempts to determine the complex filter weight Wk in the
kth band that
fits the data to the following equation:
Xk = W,rY,E (1)
In order to determine Wk in each band, a model for computing a least-squares
estimator
is required. Using equation 1, the model is straightforward, taking the form
of equation 2
where 8kl is the residual error for the ith frame in the kth band while Xki
andYk~ are
the WOLA outputs of the filterbank in band k for the ith frame.
-'ki = X ki - X ki (2)
For n frames, the sum of the squared error will be:
CA 02397080 2002-08-07
22
n n
eki-cki -L~(Xki - Wki) (Xki - Wki(3)
i=1 i=1
Note that we actually use complex conjugation (*), because the filterbank
outputs are all
complex values, and therefore two-dimensional vectors. The least squares
solution
requires that we minimize the magnitude of the error vector squared. To find
an estimator
for Wk , the derivative of equation 3 with respect to Wk is set to 0 and
solved for Wk ,
the estimator in a particular band. The result of this produces the estimator
described by
equation 4.
n
XkiYki
Wkn (4)
EYkil
i=1
Not surprisingly, equation 4 is an n sample elicitation of the cross-power
spectral density
over the auto-power spectral density within a particular band, matching
directly to the
classic optimal Wiener filter. Since the output of the filterbank at each band
is the output
of a bandpass filter, this spectral estimation is essentially a periodogram-
based estimate.
Because the WOLA filterbank results are all complex values, the resulting
filter weight
Wk is a complex value that should compensate for both magnitude and phase
differences between the correlated noise portions of each channel. The
resulting filter
which is composed of the N = 2 estimators (one per band), is then simply
applied to the
secondary channel results, Y, and subsequently subtracted fromXto produce a
cleaned
signal.
Computing the extended summations and complex division required by equation 4
is not
feasible for real-time. Therefore, some alternative method of computing the
estimator
over a reasonable timeframe is required. The adaptive filter described below
does this by
smoothing and averaging the instantaneous outputs from the filterbank in every
frame.
Also, there is the problem of dealing with the error in the adaptive filter.
Even the ideal
result from equation 4 is merely an estimate, and reducing its accuracy due to
CA 02397080 2002-08-07
23
computational constraints would theoretically reduce its performance. In the
best case,
the result of these errors in the filter will be simply reduced noise
suppression. In the
worst case, they will cause distortion and artifacts in the output signal. In
order to reduce
this effect, the algorithm does not subtract the entire result of the
filtering. Rather, it
subtracts an attenuated version of the adaptive filter output. This produces
fewer speech
artifacts at the cost of lower noise suppression.
Using only the instantaneous output of the filterbank analysis reduces the
calculation for
Wk to a single complex division with no summations required as shown in
equation 5.
This is exactly equivalent to using equation 4 with n = 1.
,. X Y* X
Wk = XkYr = Yk (5)
k k k
This estimate is then smoothed across frames using an exponential average with
parameter a such that the filter weight for the kth band in the nth frame is:
j,yk (n) = a X k ~n) + (1- a)yyk (n -1) (6)
Yk (n)
This first order difference equation acts as a low-pass filter, smoothing out
frame-by-
frame variations in the spectra which would cause the values of the filter
weights to
change quickly. This form of low-pass filter is extremely advantageous because
it is very
simple, and in particular requires only one multiplication operation. This is
useful for a
future real-time implementation. The preliminary results from using this
strategy were
successful in removing various kinds of artificial noise (white, pink,
high/low-pass), and
A
the decision was made to not extend the estimator for Wk beyond a single frame
in the
initial design. The value of a directly effects the rate at which the
algorithm develops a
"good" solution for each Wk by smoothing out all variations which deviate the
estimator
from it's optimal value. Smaller values of a cause the algorithm to converge
to a
CA 02397080 2002-08-07
24
solution at a slower rate, however values of a which are too large allow the
filter
weights to change with large jumps and creates significant artifacts which
seriously
hamper the quality of the noise reduced signal. It is possible to modify the
adaptive filter
so that it uses a small number of past frames to calculate Wk in each sub-
band, within
the limits of computational cycles of the technology.
The attenuated noise subtraction is subtracted on a band-by-band basis:
6k = Xk - ,8kWkYk (7)
Where 8k is a decimal between 0 and 1 indicating the portion of the filtered
noise in the
secondary channel to subtract from the primary channel. The intended usage of
the entire
vector, fl, is to weight the noise subtraction such that audible speech
artifacts are
minimized. This technique has been successfully used in single-microphone
noise
reduction techniques. This means using less noise suppression in sub-bands
with large
amounts of speech where the effect of the algorithm would produce the most
distortion.
Since in a real-time implementation, the frequency spectrum of the speech
signal is
unknown, a heuristic is necessary to determine a suitable )6 in advance. Based
on the
reasoning that the majority of human speech is confined to below 4 kHz, the
decision
was made to subtract all noise in bands higher than 4 kHz.Various attenuations
in bands
below 4 kHz can be chosen depending on the application and on the perception
of the
user.
The SNR improvements of the sub-band Wiener filtering algorithm presented
here, the
versatility in different noise environments and low computational cost of the
algorithm
make it an ideal candidate for bringing true digital signal processing into
the headset
market.
