Note: Descriptions are shown in the official language in which they were submitted.
CA 02406754 2002-11-07
-1-
Title: Method and Apparatus for Noise Reduction, Particularly in Hearing
Aids
FIELD OF THE INVENTION
This inventioii relates to noise reduction in audio or other
signals and more particularly relates to noise reduction in digital hearing
aids.
BACKGROUND OF THE INVENTION
Under noisy conditions, hearing impaired. persons are
severely disadvantaged compared to those with normal hearing. As a result
of reduced cochlea processing, hearing impaired persons are typically much
less able to distinguish between meaningful speech and competing sound
sources (i.e., noise). The increased attention necessary for understanding of
speech quickly leads to listener fatigue. Unfortunately, conventional
hearing aids do little to aid this problem since both speech and noise are
boosted by the same amount.
In Sheikhzadeh et al., "Comparative Performance of
Spectral Subtraction and HMM Based Speech Enhancement Strategies with
Application to Hearing Aid Design", Proceedings of the International
Conference on Acoustics, Speech, Signal Processing (ICASSP), Vol. 1, pp. I-
13 to 1-16 (April 19, 1994, IEEE), a basic spectral subtraction noise
suppression
approach and various HMM noise reduction approaches are generally
described. The basic spectral subtraction or Wiener filtering approach
theoretically minimizes noise power relative to speech. This noise
suppression method includes performing an FFT on each frame of the
input noisy signal to estimate the noisy speech spectrum, An estimate of the
noise spectrum is updated during periods of non-speech activity with the
aid of an autocorrelation-based voicing and pitch detector (i.e. when no
speech is detected the signal is assumed to be noise). A frequency domain
Wiener filter is constructed from the speech and noise spectral estimates
and is used to obtain a noise reduced or -enhanced signal (after an inverse
CA 02406754 2002-11-07
-2-
FFT transformation). However, this and the other approaches described in
the Sheikhzadeh et al. reference may, in practice, still result in
unacceptable
levels of noise in the output signal. Furthermore, these approaches also
suffer from the musical noise phenomenon as well as from degradation in
the perceptual quality of the noise reduced signal.
In addition, compression algorithms used in some hearing
aids boost low level signals to a greater extent than high level signals. This
works well with low noise signals by raising low level speech cues to
audibility. A disadvantage associated with prior art hearing aid systems that
process inputs using both noise suppression and signal compression is that,
at high noise levels, compression performs only modestly since the action
of the compressor is unduly influenced by the noise and merely boosts the
noise floor. For persons that frequently work in high ambient sound
environments, this can lead to unacceptable results.
BRIEF SUMMARY OF THE INVENTION
The present invention provides a two-fold approach to
sound quality improvement under high noise situations and its practical
implementation in a hearing aid. In one aspect the present invention
removes noise from the input signal and controls a compression stage with
a cleaner signal, compared to the use of the original noisy input signal for
controlling compression in the prior art. The signal for amplification is,
optionally, processed with a different noise reduction algorithm. Under
certain circumstances, it may be desirable to use the same noise reduced
signal for application and compression control in which case the two noise
reduction blocks merge. In another instance, it may be desirable to use
different noise reduction algorithms in each path.
Clearly, noise reduction is not suitable for all listening
situations. Any situation where a desired signal could be confused with
noise is problematic. Typically these situations involve non-speech signals
such as music. A remote control or hearing aid control will usually be
provided for enabling or disabling noise reduction.
CA 02406754 2002-11-07
-3-
The present invention is based on the realization that, what is
required, is a technique for boosting speech or other desired sound source,
while not boosting noise, or at least reducing the amount of boost given to
noise.
In accordance with a first aspect of the present invention,
there is provided a method of reducing noise in an input signal (10), said
input
signal (10) containing speech and having a signal to noise ratio, the method
comprising the steps: (1) detectirtg the presence and absence of speech; (2)
in
the absence of speech, determining a noise magnitude spectral estimate
(JI~( f) l); (3) in the presence of speech, comparing the magnitude spectrum
of
the input signal (JX(f)J) to the noise magnitude spectral estimate (J&(f)J);
characterized in that the method further comprises the steps of: (4)
calculating
an attenuation function (H(f)) from the magnitude spectrum of the input signal
(JX(f)J) and the noise magnitude spectral estimate (IN(f)l), the attenuation
function (H(f)) being dependent on the signal to noise ratio; and (5)
modifying
the input signal (10) by the attenuation function (H(f)), to generate a noise
reduced signal (12, 14) wherein there is no substantial modification to the
input
signal (10) for very low and for very high signal to noise ratios.
Preferably, the method further comprises the steps of (6)
supplying the input signal (10) to an amplification unit (22); (7) providing
the
noise reduced signal (12) to a compression circuit (20) which generates a
control input for the amplification unit (22); and (8) controlling the
amplification unit (22) with the control signal to modify the input signal
(10) to
generate an output signal (24) with compression and reduced noise.
Advantageously, step (7) comprises subjecting the input signal to an auxiliary
noise reduction algorithm (18) to generate an auxiliary noise reduced signal
(14) and providing the auxiliary noise reduced signal (14) to the compression
circuit (20).
In one embodiment the auxiliary noise reduction algorithm
(18) comprises the same noise reduction method as the main
CA 02406754 2002-11-07
-4-
noise reduction algorithm. In another embodiment, the auxiliary noise
reduction algorithm (18) is different from the noise reduction method in
the main noise reduction algorithm.
Conveniently, the square of the speech magnitude spectral
estimate (19(f)l) may be determined by subtracting the square of the of the
noise magnitude spectral estimate (J14( f)j) from the square of the magnitude
spectrum of the input signal (IX(f')I). In a preferred embodiment, the
attenuation factor is a function of frequency and is calculated in accordance
with the following equation:
H(~= Ix~)12-R1 Ntf)I2 a
I X(f) 12
where f denotes frequency, H(f) is the attenuation function, I X(f) I is the
magnitude spectrum of the input audio signal; 1l~(f)j is the noise magnitude
spectral estimate, (3 is an oversubtraction factor and a is an attenuation
rule,
wherein a and 0 are selected to give a desired attenuation function. The
oversubtraction factor 0 is, preferably, varied as a function of the signal to
noise ratio, with 0 being zero for high and low signal to noise ratios and
with 0 being increased as the signal to noise ratio increases above zero to a
maximum value at a predetermined signal to noise ratio and for higher
signal to noise ratios 0 decreases to zero at a second predetermined signal to
noise ratio greater than the first predetermined signal to noise ratio.
Advantageously, the oversubtraction factor P is divided by
a preemphasis function of frequency P(f) to give a modified oversubtraction
factor 0(f), the preemphasis function being such as to reduce j3 at high
frequencies, to reduce attenuation at high frequencies.
Preferably, the rate of the attenuation factor is controlled to
prevent abrupt and rapid changes in the attenuation factor, and it preferably
is calculated in accordance with the following equation where Gn(f) is the
smoothed attenuation function of frequency at the n'th time frame:
CA 02406754 2002-11-07
-5-
Gn(fl=(1-Y)H(fl +7Gn-,(f)
The oversubtraction factor 0 can be a function of perceptual
distortion.
The method can include remotely turning noise suppression on
and off. The method can include automatically disabling noise
reduction in the presence of' very light noise or extremely adverse
environments.
The method can include detecting speech with a modified auto-
correlation function comprising (1) taking an input sample (50) and
separating it into short blocks and storing the blocks in correlation
buffers (52); (2) correlating the blocks with one another, to form partial
correlations (56); and (3) summing the partial correlations to obtain a
final correlation (58).
The method can alternatively comprise: determining the
presence of speech by taking a block of the input signal and
performing an auto-correlation on that block to form a correlated
signal; and checking the correlated signal for the presence of a periodic
signal having a pitch corresponding to that for speech.
In a further aspect the present invention provides an apparatus
for reducing noise in an input signal (10), the apparatus including an
input for receiving the input signal (10). The apparatus comprising a
compression circuit (20) for receiving a compression control signal and
generating an amplification control signal in response, and an
amplification unit (22) for receiving the input signal (12) and the
amplification control signal and generating an output signal (24) with
compression and reduced noise. The apparatus further comprises an
auxiliary noise reduction unit (18) connected to the input for
generating an auxiliary noise reduced signal (14), the compression
control signal being the auxiliary noise reduced signal.
The apparatus may further comprise a main noise reduction
unit (16) connected to the input for generating a noise reduced signal
and supplying the noise reduced signal in place of the input signal to
the amplification unit (22).
Preferably, the input signal (10) contains speech and the
CA 02406754 2002-11-07
-6-
main noise reduction unit cornprises (1) a detector (34) connected to said
input and providing a detection signal indicative of the presence of speech;
(2) magnitude means (36) for determining the magnitude spectrum of the
input signal (IX(f) l), with both the detector (34) and the magnitude means
(36) being connected to the input of the apparatus; (3) spectral estimate
means (38) for generating a noise magnitude spectral estimate (111( f)l) and
being connected to the detector (34) and to the input of the apparatus; (4) a
noise filter calculation unit (40) connected to the spectral estimate means
(38) and the magnitude means (36), for receiving the noise magnitude
spectral estimate (111(f)1) and magnitude spectrum of the input signal
(IX(f)1)
and calculating an attenuation function (H(f)); and (5) a multiplication unit
(42) coupled to the noise filter calculation unit (40) and the input signal
(10)
for producing the noise reduced signal (12).
BRIEF DESCRIPTION OF THE DRAWING FIGURES
For a better understanding of the present invention and to
show more clearly how it may be carried into effect, reference will now be
made, by way of example, to t:he accompanying drawings in which:
Figure 1 is a conceptual blocked diagram for hearing aid
noise reduction and compression;
Figure 2 shows a detailed blocked diagram for noise
reduction in a hearing aid;
Figure 3 shows a modified auto-correlation scheme
performed in segments.
DESCRIPTION OF THE PREFERRED EMBODIMENT
Referring first to Figure 1, there is shown schematically a
basic strategy employed by the present invention. An input 10 for a noisy
signal is split into two paths 12 and 14. In the upper path 12, the noise
reduction is effected as indicated in block 16. In the lower path 14, noise
reduction is effected in unit 18. The noise reduction unit 18 provides a
cleaner signal that is supplied to compression circuitry 20, and the
CA 02406754 2002-11-07
-6A-
compression circuitry controls amplification unit 22 amplifying the
signal in the upper path to generate an output signal at 24.
Here, the position of the noise reduction unit 18 can
advantageously provide a cleaner signal for controlling the compression
stage. The noise reduction unit 18 provides a first generating means
which generates an auxiliary signal from an auxiliary noise reduction
algorithm. The auxiliary algorithm performed by unit 18 may be
identical to the one performed by unit 16, except with different
parameters. Since the auxiliary noise reduced signal is not heard, unit 18
can reduce noise with increased aggression. This auxiliary signal, in
turn, controls the compression circuitry 20, which comprises second
generating means for generating a control input for controlling the
amplification unit 22.
The noise reduction unit 16 is optional and can be
effected by using a different noise reduction algorithm from that in the
noise reduction unit 18. If the same algorithm is used for both noise
reduction processes 16 and 18, then the two paths cart be merged prior
to being split up to go to units 20 and 22.
With reference to Figure 2, this show's a block diagram
of a specific realization of the proposed noise reduction technique which
is preferably carried out by n.oise reduction unit 18 (and possibly also
noise reduction unit 16). The incoming signal at 10 is first blocked and
windowed, as detailed in applicants' simultaneously filed international
application no. PCT/CA98/00329 corresponding to international
publication no. WO 98/47313. The blocked and windowed output
provides the input to the frequency transform (all of these steps take
place, as indicated, at 32), which preferably here is a Discrete Fourier
Transform (DFT), to provide a signal X( f). The present invention is not
however restricted to a DFT and other transforms can be used. A
known, fast way of implementing a DFT with mild restrictions on the
transform size is the Fast Fourier Transform (FFT). The input 10 is also
connected to a speech detector 34 which works in parallel to isolate the
pauses in the incoming speech. For simplicity,
CA 02406754 2002-11-07
- 6B -
reference is made here to "speech", but it will be understood that this
encompasses any desired audio signal capable of being isolated or detected by
detector 34. These pauses provicle opportunities to update the noise spectral
estimate. This estimate is updated only during speech pauses as a running slow
average. When speech is detected, the noise estimate is frozen.
As indicated at 38, the outputs from both the unit 32 and the
voice detection unit 34 are connected to block 38 which detects the magnitude
spectrum of the incoming noise, J N (f ) 1. The magnitude spectrum detected by
unit 38 is an estimate. The output of unit 32 is also connected to block 36
for
detecting the magnitude spectrum of the incoming noisy signal, JX(f) 1.
A noise filter calculation 40 is made based on JX(f)J and
1~1( f) to calculate an attenuation function H( f). As indicated at 42, this
is used
to control the original noisy signal X(f) by multiplying X(f) by H(f). This
signal
is subject to an inverse transform and overlap-add resynthesis in known
manner at 44, to provide a noise reduced signal 46. The noise reduced signal
46
in Figure 2 may correspond to either of the signals at 12 or 14 in Figure 1.
During speech utterances, the magnitude spectrum is
compared with the noise spectral estimate. In general, frequency dependent
attenuation is calculated as a function of the two input spectra. Frequency
regions where the incoming signal is higher than the noise are attenuated less
than regions where the incoming signal is comparable or less than the
CA 02406754 2002-11-07
-7-
noise. The attenuation function is generally given by
_ IS(f )I2 a
H~ IS(f12 + IN(f)IZ
where H(f) is the attenuation as a function of frequency
S(f) is the clean speech spectrum
N(f) is the noise spectrum
a is the attenuation rule
The attenuation rule preferably selected is the Wiener attenuation rule
which corresponds to a equal to 1. The Wiener rule minimizes the noise
power relative to the speech. Other attenuation rules can also be used, for
example the spectral subtraction rule having a equal to 0.5.
Since neither S(f) nor N(f) are precisely known and would
require a priori knowledge of' the clean speech and noise spectra, they are
replaced by estimates S(f) and 14(f):
Ig(t) I2= IxV) 1 2-Ig (f)I2
where X(f) is the incoming speech spectrum and R(t) is the noise spectrum
as estimated during speech pauses. Given perfect estimates of the speech
and noise spectra, application of this formula yields the optimum (largest)
signal-to-noise-ratio (SNR). Although the SNR would be maximized using
this formula, the noise in the resulting speech is still judged as excessive
by
subjective assessment. An improved implementation of the formula taking
into account these perceptual aspects is given by:
H(f)- ~XV) 1 2-a1 1,1 ~f) 1 2
Ix(t) I
where: (3 is an oversubtraction factor
CA 02406754 2002-11-07
-$-
a is the attenuation rule
H(f) should be between 0.0 and 1.0 to be meaningful. When
negative results are obtained, H(f) is simply set to zero at that frequency.
In
addition, it is beneficial to increase the minimum value of H(f) somewhat
above zero to avoid complete suppression of the noise. While counter-
intuitive, this reduces the musical noise artifact (discussed later) to some
extent. The parameter a governs the attenuation rule for increasing noise
levels. Generally, the higher a is set, the more the noise is punished as X(f)
drops. It was found that the best perceptual results were obtained with a =
1Ø The special case of a = 1.0 and (3=1.0 corresponds to power spectrum
subtraction yielding the Wiener filter solution as described above.
The parameter 0 controls the amount of additional noise
suppression required; it is ideally a function of the input noise level.
Empirically it was noticed that under very light noise (SNR > 40 dB) 0
should be zero. For lower SNR signals, the noise reduction becomes less
reliable and is gradually turned off. An example of this additional noise
reduction is:
0=0 for SNR<0
0=00 SM for 0<SNR<5
0=00 [ 1- CS'~ s) for 5<SNR<40
(3=0 for SNR>40
In this example, 0a refers to the maximum attenuation, 5Ø In effect, from
SNR = 0, the attenuation P is ramped up uniformly to a maximum, 0o, at
SNR = 5, and this is then uniformly ramped down to zero at SNR = 40.
Another aspect of the present invention provides
improvements in perceptual quality making P a function of frequency. As
CA 02406754 2002-11-07
-9-
an instance of the use of this feature, it was found that to avoid excessive
attenuation of high frequency information, it was necessary to apply a
preemphasis function, P(f), to the input spectrum X(f) , where P(f) is an
increasing function of frequency. The effect of this preemphasis function is
to artificially raise the input spectrum above the noise floor at high
frequencies. The attenuation rule will then leave the higher frequencies
relatively intact. This preemphasis is conveniently accomplished by
reducing ~ at high frequencies by the preemphasis factor.
~i( fl= , where ~i is (3 after preemphasis.
Ptf)
Without further modification, the above formula can yield
noise reduced speech with an audible artifact known as musical noise. This
occurs, because in order for the noise reduction to be effective in reducing
noise, the frequency attenuation function has to be adaptive. The very act
of adapting this filter allows isolated frequency regions of low SNR to
flicker
in and out of audibility leading to this musical noise artifact. Various
methods are used to reduce this problem. Slowing down the adaptation
rate significantly reduces this problem. In this method, a forgetting factor,
y
is introduced to slow abrupt gain changes in the attenuation function:
G n (f) = (1-1')H(f) +'YGõ_, (f)
where Gn( f) and Gn_i( f) are the smoothed attenuation functions at the n'th
and (n-1)'th time frames.
Further improvements in perceptual quality are possible
by making 0 (in addition to being a function of frequency) a function of
perceptual distortion. In this method, the smoothing function (instead of a
simple exponential or forgetting factor as above) bases its decision on
adapting Gn( f) on whether such a change is masked perceptually. The
perceptual adaptation algorithm uses the ideal attenuation function H(f) as
a target because it represents the best SNR attainable. The algorithm decides
CA 02406754 2002-11-07
_10-
how much Gn(f) can be adjusted while miniinizing the perceptual
distortion. The decision is based on a number of masking criteria in the
output spectrum including:
1. Spread of masking - changes in higher frequency energy
are masked by the presence of energy in frequencies in the vicinity -
especially lower frequencies;
2. Previous energy - changes in louder frequency
components are more audible that changes in weaker frequency
components;
3. Threshold of hearing - there is no point in reducing the
noise significantly below the threshold of hearing at a particular frequency;
4. Previous attenuation - low levels should not be allowed
to jump up rapidly - high levels should not suddenly drop rapidly unless
masked by 1), 2) or 3).
For applications where the noise reduction is used to
preprocess the input signal before reaching the compression circuitry
(schematically shown in Figure 1), the perceptual characteristics of the noise
reduced signal are less important. In fact, it may prove advantageous to
perform the noise reduction with two different suppression algorithms as
mentioned above. The noise reduction 16 would be optimized for
perceptual quality while the other noise reduction 18 would be optimized
for good compression performance.
A key element to the success of the present noise
suppression or reduction system is the speech or voicing detector. It is
crucial to obtain accurate estimates of the noise spectrum. If the noise
spectral estimate is updated during periods of speech activity, the noise
spectrum will be contaminated with speech resulting in speech cancellation.
Speech detection is very difficult, especially under heavy noise situations.
Although, a three-way distinction between voiced speech, unvoiced speech
(consonants) and noise is possible under light noise conditions, it was
found that the only reliable distinction available in heavy noise was
between voiced speech and noise. Given the slow averaging of the noise
CA 02406754 2002-11-07
-11-
spectrum, the addition of low--energy consonants is irisignificant.
Thus, another aspect of the present invention uses an
auto-correlation function to detect speech, as the advantage of this function
is the relative ease with which a periodic signal is detected. As will be
appreciated by those skilled in the art, an inherent property of the auto-
correlation function of a periodic signal is that it shows a peak at the time
lag corresponding to the repetition period (see Rabiner, L.R., and Schafer,
R.W., Digital Processing of S17eech Signals, (Prentice Hall Inc., 1978)).
Since
voiced speech is nearly periodic in time at the rate of its pitch period, a
voicing detector based on the auto-correlation function was developed.
Given a sufficiently long auto-correlation, the uncorrelated noise tends to
cancel out as successive pitch periods are averaged together.
A strict short-time auto-correlation requires that the signal
first be blocked to limit the time extent (samples outside the block are set
to
zero). This operation is followed by an auto-correlation on the block. The
disadvantage of this approach is that the auto-correlation function includes
fewer samples as the time lag increases. Since the pitch lag (typically
between 40 and 240 samples (equivalent to 2.5 to 15 milliseconds) is a
significant portion of the auto-correlation frame (typically 512 samples or 32
milliseconds), a modified version of the auto-correlation function avoiding
this problem was calculated. This modified version of the auto-correlation
function is described in Rabiner, L.R., and Schafer, R.W., Digital Processing
of Speech Signals, supra. In this method, the signal is blocked and correlated
with a delayed block (of the same length) of the signal. Since the samples in
the delayed block include samples not present in the first block, this
function is not a strict auto-correlation but shows periodicities better.
It is realized that a hearing aid is a real-time system and
that all computational elements for each speech block are to be completed
before the next arrives. The calculation time of a long auto-correlation,
which is required only every few speech blocks, would certainly bring the
system to a halt every time it must be calculated. It is therefore recognized
CA 02406754 2002-11-07
-12-
that the auto-correlation should be segmented into a number of shorter
sections which can be calculated for each block and stored in a partial
correlation table. The complete auto-correlation is determined by stacking
these partial correlations on top of each other and adding as shown in
Figure 3.
Referring to Figure 3, input sample 50 is divided into
separate blocks stored in merr-ory buffers as indicated at 52. The correlation
buffers 52 are connected to a block correlation unit 54, where the auto-
correlation is performed. Partial cross-correlations 56 are summed to give
the final correlation 58.
This technique quickly yields the exact modified auto-
correlation and is the preferred embodiment when sufficient memory is
available to store the partial correlations.
When memory space considerations rule out the above
technique, a form of exponential averaging may be used to reduce the
number of correlation buffers to a single buffer. In this technique,
successive partial correlations are summed to the scaled down previous
contents of the correlation buffer. This simplification significantly reduces
the memory but implicitly applies an exponential window to the input
sequence. The windowing action, unfortunately, reduces time periodicities.
The effect is to spread the autocorrelation peak to a number of adjacent time
lags in either direction. This peak smearing reduces the accuracy of the
voicing detection somewhat.
In the implementations using an FFT transform block,
these partial correlations (for either technique given above) can be
performed quickly in the frequency domain. For each block, the correlation
operation is reduced to a sequence of complex multiplications on the
transformed time sequences. The resulting frequency domain sequences
can be added directly together and transformed back to the time domain to
provide the complete long auto-correlation. In an alternate embodiment,
the frequency domain correlation results are never inverted back to the
time domain. In this realization, the pitch frequency is determined directly
CA 02406754 2002-11-07
-13-
in the frequency domain.
Since the auto-correlation frame is long compared to the
(shorter) speech frame, the 'voicing detection is delayed compared to the
current frame. This compensation for this delay is accomplished in the
noise spectrum update block.
An inter-frame constraint was placed on frames
considered as potential candidates for speech pauses to further reduce false
detection of noise frames. The spectral distance between the proposed
frame and the previous estimates of the noise spectrum are compared.
Large values reduce the likelihood that the frame is truly a pause. The
voicing detector takes this information, the presence or absence of an auto-
correlation peak, the frame energy, and a running average of the noise as
inputs.
