CA 02291826 1999-11-29
NOISE REDUCTION APPARATUS AND NOISE REDUCTION METHOD
BACKGROUND OF THE INVENTION
Field of the Invention
s The present invention relates to a noise reduction method
of a speech signal adopted in a speech communication system
or a speech recognition system used under an environment with
background noises. The noise reduction method suppresses
a noise of the speech signal input to these systems by
1o eliminating the noise. The present invention also relates
to a noise reduction method of a noise frame of the speech
signal by suppressing its amplitude.
Background Arts
15 A conventional method to reduce the background noises is
illustrated in the drawing of Fig.lS. The method will be
described below. This method converts an input signal of
speech having background noises from analogue to digital.
An A/D converted input signal is divided into fixed periods
20 (frames). Then, the followings are performed in order to
obtain an output signal with reduced noises for every one
of the divided frames. A discrete Fourier transformation
is applied to a series of input signals including a fixed
period signal (hereinafter referred to as a frame signal) .
25 The frame signal is transformed to a frequency spectrum by
1
CA 02291826 1999-11-29
a
applying the discrete Fourier transformation to the series
of input signals. The frequency spectrum is divided into
an amplitude spectrum and a phase spectrum. An estimated
noise amplitude spectrum has been estimated from a non speech
frame. The estimated noise amplitude spectrumis subtracted
from the amplitude spectrum of the input signal using a
subtraction filter. By doing so, a speech amplitude
spectrum with reduced background noises (output amplitude
spectrum) isestimated. Then, thespeech amplitudespectrum
to is added to the phase spectrum to get a frequency spectrum
of the speech with reduced background noises. By applying
an inverse discrete Fourier transformation to the frequency
spectrum of this speech, the output signal with reduced
noises is obtained. This method is being proposed in a paper
15 by Steven F. Ball, entitled "Supprssion of acustic noise in
speech using spectral subtraction", IEEE Trans. Acoust.,
Speesh and Signal Proc., vol. ASSP-29, pp.113-120, Apr.
1979.
20 Fig.lS illustrates a block chart of the conventional noise
reduction method.
First of all, an overall operation of the conventional noise
reduction method is described with reference to the drawing
25 of Fig.lS.
2
CA 02291826 1999-11-29
a.
An input signal 107 cut into a fixed frame length is
transformed to a frequency domain at a Fourier
transformation unit 101. An input phase spectrum 108 and
an input amplitude spectrum 109 are output from the Fourier
transformation unit 101. A noise period deciding unit 102
decides that the input signal is in a speech frame (period)
if greater than a threshold value TH, and the input signal
is in a noise frame (period) if less than the threshold value
to TH.
When a current frame is a noise frame, then an estimated noise
amplitude spectrum calculating unit 103 performs a weighted
addition of the input amplitude spectrum 109 of that time
1s and an estimated noise amplitude spectrum up to that time
and outputs an updated estimated. noise amplitude spectrum
110.
A typical transmission function of the subtraction filter
2o unit 104 is expressed by an equation (1).
F(w)= B(cu)2-r ' E(N(~))2 (1)
S (~)2
whereas
F (~): subtraction filter
3
CA 02291826 1999-11-29
S (cu) : input amplitude spectrum
E (N (cu)) : estimated noise amplitude spectrum
r : reduction rate
For the equation ( 1 ) , S ( c~ ) denotes to an amplitude spectrum
of the input signal, E (N (cu)) denotes to an estimated noise
amplitude spectrum, and r is a constant expressing the
reduction rate of the estimated noise amplitude spectrum.
An amount of noise suppression increases as the reduction
to rate r increases. On the other hand, the amount of noise
suppression decreases as the reduction rate r decreases.
The reduction rate r is calculated based on a pre-determined
equation at the subtraction filter unit 104.
Further, based on the reduction rate r, an output amplitude
spectrum 111 is output from the subtraction filter unit 104
by subtracting the estimated noise amplitude spectrum from
the input amplitude spectrum, in a same manner as the equation
(1) .
In other words, at the subtraction filter unit 104, the noise
suppression is weakened by decreasing the reduction rate for
a frame with small power such as noise frame or consonant
sounds .
4
CA 02291826 1999-11-29
An inverse Fourier transformation unit 105 outputs an output
signal in a time domain.
s The first conventional spectrum subtraction method has a
disadvantage of deteriorated speech quality due to cutoff
of a speech such as onset of speech, ending part of speech,
or frame with small power such as consonant sounds. Such
cutoff of the speech will further increase by increasing the
to reduction rate r.
In order to resolve this problem, a method to vary the
reduction rates of the subtraction filter for every frames
is disclosed in Japanese unexamined patent publication
15 HEIB-221092. In this method, suppose that the F (a>) as in
the equation (1) is used as the subtraction filter, for
example, then if an input is large, r is set large. And if
an input is small, the speech cutoff is reduced by making
the r small.
As a second conventional spectrum subtraction method in
improving a noise reduction efficiency, a function called
window function is determined for' a period in time axis, and
the signals for applying the Fourier transformation is
weighted by multiplying of the window function. For example,
s
CA 02291826 1999-11-29
r
a frame signal and a fixed length signal which is continuous
to the frame signal (overlap signal) are multiplied by the
window function and then the Fourier transformation is
applied. For a frame signal output from the inverse Fourier
s transformation, an overlap signal output from the inverse
Fourier transformation of the previous frame is added to a
beginning of the sample weighted by less than 1 in the frame.
This conventional method is being utilized in North America
CDMA Automobile Telephone System (TIA/EIA, IS-127,
to "Enhanced Variable Rate Codec Service Option 3 for Wideband
Spread Spectrum DigitalSystems"). The window function used
in this example is illustrated in the drawing of Fig. l6.
In the drawing, the window function is used so that addition
15 of weights of the samples corresponding to an addition of
the frame signal and the overlap signal becomes 1. As well,
the window function is used to get smooth edges of the signals
for applying the Fourier transformation. By weighting using
such window function, an estimation accuracy of the noise
2o amplitude spectrum is improved, and eventually, the noise
reduction efficiency is improved.
As a third conventional spectrum subtraction method, there
is a method of applying Fourier transformation to a frame
25 signal and a fixed length signal (overlap signal) continuous
6
CA 02291826 1999-11-29
to the frame signal without weighting by the window function.
This method is being used in the noise reduction apparatus
disclosed in Japanese unexamined patent publication HEI
9-34497. This method implements a waveform reforming
process which overlaps an overlap signal output from the
inverse Fourier transformation of the previous frame with
a frame signal output from the inverse Fourier
transformation at a triangle window and outputs the result .
The waveform reforming process is implemented to achieve
Zo smooth outputs between frames. The overlap signal output
from the inverse Fourier transformation is stored for
overlapping to a next frame. The wave form reforming process
is expressed by an equation (2).
Oj = ( j ~ D; -I- (L- j ) ~ Z; ) /L
( j =O~-L- 1 )
Oj =Dj ( j =L~-L+M- 1 )
Z; =OM+i (j =O~~L-1)
(2)
2o whereas
O~ . output signal
D~ . frame signal
Zj , overlap signal
M : frame length
L :overlap signal length
CA 02291826 1999-11-29
a
An advantage of the third conventional method is described
below.
In order to implement a spectrum subtraction method,
theoretically, it is common to use a 16-bits fixed decimal
point calculation digital signal processor (fixed decimal
point DSP) having a limited precision calculation. The
precision calculation of the fixed decimal point DSP is
1o increased by a double precision calculation or by shifting
a decimal point up by certain places. However, these can
cause a problem of increased amount of calculation.
Basically, the Fourier transformation and inverse Fourier
transformation involve a large amount of calculation.
However, there are many cases of wanting to implement the
spectrum subtraction method by involving a small amount of
calculation, although this may sacrifice the precision
calculation. When the third conventional spectrum
2o subtraction method is adopted, since the frame signal for
applying Fourier transformation is not weighted by the
window function, a dynamic range. of the frame signal does
not spread. And an accuracy deterioration of the frame
signal is small if being applied the Fourier transformation
and inverse Fourier transformation with the limited
a
CA 02291826 1999-11-29
precision calculation. This is the reason for using the
fixed decimal point DSP involving the small amount of
calculation.
The fourth conventional noise reduction method using a
spectrum subtraction, at first, estimates a noise power for
calculating the estimated noise amplitude spectrum, and a
calculated noise power is taken as a noise frame deciding
threshold value. Then, if an input power is smaller than
io this threshold value, the input frame is decided as a noise
frame. It is common that the estimated noise amplitude
spectrum is calculated as an average of noise signals, that
is, an average of input signals of a plurality of frames which
are decided as noises . This spectrum subtraction method is
disclosed in Japanese unexamined patent publication
HEIB-221092 as an example.
As a fifth conventional method to reduce the background noise,
there is a method to decline a noise perception by suppressing
2o an amplitude of the noise period. A method which is
disclosed in Japanese unexamined patent publication
HEI7-38454 decides a state of current frame from a pre-
determined finite state which expresses a signal mode, and
suppresses the amplitude by a certain intensity if the state
of current frame is indicating a noise frame state.
9
CA 02291826 1999-11-29
The first conventional spectrum subtraction method has the
disadvantage where a noise period is distorted and become
an unpleasant residual noise called musical noise, which is
s a major practical problem.
This problem occurs in the noise period when an output
amplitude spectrum shows a form in which a power o~f the output
amplitude spectrum is sparsely concentrated to some of
to frequencies by subtracting estimated noise amplitude
spectrum. The problem occurs when the power concentrated
frequencies varies irregularly among frames. To solve the
problem of musical noise, one can adopt a method to make the
reduction rates of estimated noise amplitude spectrum
is variable, by setting an amount of reduction small for the
noise frame. But the setting of the reduction rates small
leads to a problem which is a lack in the amount of noise
suppression for the noise frame.
2o For the second conventional spectrum subtraction method, the
noise reduction efficiency will improve by applying the
Fourier transformation to a signal weighted by the window
function so that both edges of the signal approaches 0.
However, an accuracy of the signal at both edges declines
2s when this method is performed by the fixed decimal point DSP,
CA 02291826 1999-11-29
therefore, there is a problem of discontinuous feeling sound
occurring at the frame boundaries when both edges of the
signal output from the inverse Fourier transformation
deteriorates.
In the third conventional example, for the case of applying
the Fourier transformation to a signal not being weighted,
an accuracy is easily attained even if being implemented by
the fixed decimal point DSP. However, variation among
1o frames of noise amplitude spectrum component included in the
input amplitude spectrum is more enlarged for this case than
a case of applying Fourier transformation to the signal being
weighted. Therefore, as a result, the noise reduction
efficiency is reduced for the non.-weighting case comparing
1s to the weighting case.
In the fourth conventional spectrum subtraction method, for
deciding the noise period, the noise period deciding
threshold value is set higher for the cases of a large power
2o variation of noise so that the noise frame can be decided
correctly. In doing so, the speech frame is also wrongly
decided as a noise frame to a certain extent, and a speech
spectrum will be involved to the estimated noise amplitude
spectrum, resulting in the speech cutoff. On contrary, to
25 prevent this from happening, the noise period deciding
11
CA 02291826 1999-11-29
threshold value can be set lower so as not to decide the speech
frame as the noise frame. But by doing so the noise frame
is misjudged as a speech frame, therefore, the updating of
estimated noise amplitude spectrum is not performed
correctly, and as a result, the noise reduction efficiency
is reduced.
The fifth conventional method declines the noise perception
by suppressing the amplitude of noise period, and a current
to frame state is decided from a pre-determined finite states
expressing the signal mode. A problem with the fifth
conventional method is that if the frame state indicates a
noise frame state, if this method adopts a method suppressing
the amplitude of noise frame by a certain intensity, there
is a case that the frame state transition happens frequently
between the speech frame state and the noise frame state in
a short period of time depending on a type of input noise.
In addition to that, the intensity of suppressing amplitude
changes frequently. These result in problems of unstable
output power and deteriorated hearing.
The present invention attempts to solve the problems
mentioned above by aiming for a noise reduction method which
can lessen unpleasant residual noise, even if the spectrum
of each frame is sparsely concentrated to some of the
12
CA 02291826 1999-11-29
frequencies or even if the frame state transition are
happening frequently from or to the noise frame.
Disclosure of the invention
According to one aspect of the present invention, a noise
reduction apparatus for obtaining an output from a spectrum
subtraction filter by subtracting an estimated noise
amplitude spectrum from an amplitude spectrum which is
obtained by orthogonally transforming an input signal cut
to into a fixed frame length comprises an amplitude adjusting
filter unit which obtains a desired output for each frame
by multiplying a subtraction output from the spectrum
subtraction filter and an amplitude adjusting coefficient
which is determined from a power ~of the amplitude spectrum
and a power of the estimated noise amplitude spectrum.
According to another aspect of the present invention, the
noise reduction apparatus includes the spectrum subtraction
filter which varies a reduction rate of a subtraction of
2o estimated noise amplitude spectrum based on the estimated
noise amplitude spectrums and includes the amplitude
adjusting filter unit which varies the amplitude adjusting
coefficient based on the reduction rate.
According to another aspect of the present invention. the
13
CA 02291826 1999-11-29
noise reduction apparatus includes the amplitude adjusting
filter unit which increases the amplitude adjusting
coefficient if the reduction rate is large, so as to intensify
a noise suppression of speech period and to increase an output
value of output signals and includes the amplitude adjusting
filter unit which decreases the amplitude adjusting
coefficient if the reduction rate is small, so as to weaken
a noise suppression of noise period and to decrease an output
value of output signal.
According to another aspect of the present invention, the
noise reduction apparatus includes the amplitude adjusting
filter unit which multiplies the amplitude adjusting
coefficient to an amplitude spectrum is a time domain, which
has applied an inverse orthogonal transformation to the
subtraction output by the spectrum subtraction filter.
According to another aspect of the present invention, the
noise reduction apparatus includes the amplitude adjusting
2o filter unit which multiplies the amplitude adjusting
coefficient to the amplitude spectrum in a frequency domain
for each frame, and obtains an output by applying an inverse
orthogonal transformation.
According to another aspect of the present invention, the
14
CA 02291826 1999-11-29
noise reduction apparatus inlcudes the amplitude adjusting
coefficient which is a value obtained by an addition of a
weighted amplitude adjusting coefficient obtained at a
previous frame and an amplitude adjusting coefficient of a
current frame which is obtained based on a difference of the
power of amplitude spectrum of input signal of the current
frame and the power of estimated noise amplitude spectrum
of the current frame.
to According to another aspect of the present invention, a noise
reduction apparatus for obtaining an output from a spectrum
subtraction filter by subtracting an estimated noise
amplitude spectrum from an amplitude spectrum which is
obtained by orthogonally transforming an input signal cut
i5 into a fixed frame length comprises an input signal creating
unit for getting signals in periods before and after a current
frame, multiplying a weighted function of less than 1 to the
signals in the periods before and after the current frame
so that edges of the periods will be close to 0, and attaching
2o the periods before and after the current frame to the current
frame, so that an output of the input signal creating unit
is treated as an input signal for .calculating the amplitude
spectrum.
25 According to another aspect of the present invention, the
CA 02291826 1999-11-29
noise reduction apparatus further inlcudes a waveform
reforming unit for reforming the current frame by using a
period after the previous frame, which is multiplied by the
weighted function of 1.
s
According to another aspect of the.present invention, a noise
reduction apparatus for obtaining an output from a spectrum
subtraction filter by subtracting an estimated noise
amplitude spectrum from an amplitude spectrum which is
obtained by orthogonally transforming an input signal cut
into a fixed frame length comprises a reduction rate
calculating unit for obtaining a reduction rate by
calculating an average noise power from a plurality of frames
of the estimated noise amplitude spectrum, and by comparing
is the calculated average noise power with a noise period
deciding threshold valued and comprises the calculated
reduction rate which is used as the reduction rate of the
spectrum subtraction filter.
2o According to another aspect of the present invention, the
noise reduction apparatusfurther comprises an average noise
power calculating unit for calculating an average noise
power from a plurality of frames of the estimated noise
amplitude spectrum: and includes the amplitude adjusting
2s filter unit which multiplies the subtraction output which
16
CA 02291826 1999-11-29
is output from the spectrum subtraction filter for each frame
to an amplitude adjusting coefficient determined from the
average noise amplitude power and the amplitude spectrum
power, and obtains a desired output.
s
According to another aspect of the.present invention, a noise
reduction method for obtaining an output from a spectrum
subtraction filter by subtracting an estimated noise
amplitude spectrum from an amplitude spectrum which is
so obtained by orthogonally transforming an input signal cut
into a fixed frame length comprises the steps of: varying
a reduction rate of the spectrum subtraction filter based
on the estimated noise amplitude spectrums and multiplying
a subtraction output which is output from the spectrum
15 subtraction filter for each frame to the amplitude adjusting
coefficient determined from a power of amplitude spectrum
and a power of estimated noise amplitude spectrum, and
obtaining a desired output.
2o According to another aspect of the present invention, a noise
reduction method for obtaining an output from a spectrum
subtraction filter by subtracting an estimated noise
amplitude spectrum from an amplitude spectrum which is
obtained by orthogonally transforming an input signal cut
25 into a fixed frame length, comprises the steps of calculating
1~
CA 02291826 1999-11-29
a reduction rate including a step of calculating an average
noise power from a plurality of frames of the estimated noise
amplitude spectrum and a step of comparing the calculated
average noise power with a noise period deciding threshold
value; and setting the calculated reduction rate as a
reduction rate of the spectrum subtraction filter.
Brief Description of the drawings
Fig.l illustrates a configuration of spectrum subtraction
1o method of embodiment 1 of the present invention;
Fig.2 illustrates a flow chart of the noise reduction
method in an amplitude adjusting filter in the noise
reduction method of embodiment 1; ,
Fig.3 explains an operation performed by a noise period
is deciding unit of embodiment 1;
Fig.4 illustrates an operation performed by an amplitude
adjusting filter of embodiment 1;
Fig.5 explains output signals of spectrum subtraction
method of embodiment 1;
2o Fig.6illustrates a configuration of spectrum subtraction
method of embodiment 2 of the present invention;
Fig.7 illustrates a configuration of subtraction method
of embodiment 3 of the present invention;
Fig.8 illustrates an example of input signal for noise
25 reduction method of embodiment 3;
18
CA 02291826 1999-11-29
Fig.9 illustrates an example of weighting function to be
multiplied by the input signal creating unit for the noise
reduction method of embodiment 3;
Fig.lO illustrates an example of output signal of inverse
Fourier Transformation unit for the noise reduction method
of embodiment 3;
Fig.ll illustrates a configuration of spectrum
subtraction method of embodiment 4 of the present invention;
Fig. l2 illustrates an operation of the noise reduction
to method of embodiment 4;
Fig.l3 illustrates a spectrum subtraction method of
embodiment 5 of the present invention;
Fig.l4 explains an operation of a noise reduction method
of embodiment 5;
Fig. l5 illustrates a configuration of the conventional
noise reduction method;
Fig.l6 illustrates a window function used for the
conventional noise reduction method;
Fig.l7 illustrates a comparison of the operations of
2o embodiment 1; and
Fig.l8 illustrates a waveform reforming process of the
embodiment 3.
Best mode for carrying out the invention
Embodiment 1.
19
CA 02291826 1999-11-29
A noise reduction method of the embodiment 1 is described.
According to the noise reduction method of the present
embodiment, a reduction rate of the spectrum subtraction
filter is variable. The noise reduction method of the
present invention comprises an amplitude adjusting filter
unit which subtracts an amplitude from a normal output. The
amplitude subtraction relies on the reduction rate of the
spectrum subtraction filter.
1o The drawing of Fig.1 illustrates a block chart of the noise
reduction method of the embodiment 1. The drawing of Fig.2
illustrates a flow chart-of the amplitude adjusting filter
unit of Fig. 1.
The drawing of Fig.l7 illustrates a comparison of operations
for the speech period and the noise period .
To begin with, an overall operation of the noise reduction
method of embodiment 1 is described with reference to the
2o drawing of Fig. 1.
An input signal 7 having a fixed frame length is orthogonally
transformed at the Fourier transformation unit l,
transformed to a frequency domain, and an input phase
spectrum 8 and an input amplitude spectrum 9 are output. A
CA 02291826 1999-11-29
noise period deciding unit 2 decides that the input signal
is in a speech frame (period) if it exceeded a threshold value
TH, and decides that the input signal is in a noise frame
(period) if below the threshold value. If an equation (3) is
s satisfied, the threshold TH is replaced by a new threshold
THnew
Pin< 2. 0 TH (3 )
TH"eW =0. 9 TH+0. 1 Pi" (4)
to
Whereas Pi" is an input signal power, TH is a noise period
deciding threshold value before updating, and TH"eW is a noise
period deciding threshold value after updating.
15 Accordingly, a relation illustrated in the drawing of Fig.3
is obtained. If a current frame is decided as a noise frame,
an estimated noise amplitude spectrum calculation unit 3
performs a weighted addition of an input amplitude spectrum
9 of the time and an estimated noise amplitude spectrum up
2o to that time and outputs an estimated noise amplitude
spectrum 10. For example, the estimated noise amplitude
spectrum 10 is updated by an equation (5).
E (N (~)) =Eold (N (w)) a
25 -f-S (W) ~ (1-a) (5)
21
CA 02291826 1999-11-29
Whereas
E (N (cu)) . estimated noise amplitude spectrum after
updating
Eold (N (c~)) . estimated noise amplitude spectrum
before updating
S (w) . amplitude spectrum of input signal
0<a<1
At the subtraction filter unit 4, the reduction rate r is
to calculated based on a next equation (6), and an obtained
reduction rate r is output as a noise reduction intensity
13.
r =m i n { 1.0, rTH ~ POWs /POWN } (6)
POWS: power of amplitude spectrum of input signal of all
frequencies of a current frame.
POWN: power of estimated noise amplitude spectrum of all
frequencies of a current frame
2 0 0 < r ,~,,~ < 1
Further, based on the reduction rate r, an input amplitude
spectrum is subtracted by an estimated noise amplitude
spectrum in a same manner as the equation (1), to give a
subtracted output amplitude spectrum 11.
22
CA 02291826 1999-11-29
The subtraction filter unit 4 (spectrum subtraction filter)
weakens a noise suppression by making the reduction rate
small for a noise frame or a frame having small power such
as consonant sounds.
The amplitude suppression of an output amplitude 12 which
is returned to a time domain at an inverse Fourier
transformation unit 5 is intensified since an amplitude
to adjusting coefficient gets small at a next amplitude
adjusting filter unit 6 for those frames with small powers.
That is, the amplitude adjusting filter unit 6 performs the
following three operations from steps S1 to S3 of Fig.2. The
noise reduction intensity 13 (the reduction rate r) is input
at step S1. Then a power of estimated noise amplitude
spectrum of a current frame is subtracted from a power of
amplitude spectrum of input signal of the current frame. And
then, an amplitude suppressing coefficient G is calculated
2o using an equation (7) for each frame.
23
CA 02291826 1999-11-29
G= P OWs-P OWN
P OWs- r ' P OWN
(POW S-POW NCO, and, POW S-r~ POW N>O)
G=O
(POW s-POW N<O, or, POW s-r~ POW N50) (7)
Whereas POWS denotes a power of amplitude spectrum of all
frequencies of input signal of a current frame, POWN denotes
a power of estimated noise amplitude spectrum of all
frequencies of the current frame, G denotes an amplitude
suppressing coefficient of the current frame, and r denotes
the reduction rate. As illustrated in the drawing of Fig.17,
the amplitude suppressing coefficient G gets large if the
reduction rate r is large. If the reduction rate r gets small,
to the amplitude suppressing coefficient G is also small.
Further, the amplitude adjusting coefficients g (n) for each
of samples are calculated in step S2 using an equation (8) .
g (n) =g (n-1) AR-~G (1-AR) (8)
Provided that
0<AR<1
2o Whereas g (n) denotes an amplitude adjusting coefficient of
24
CA 02291826 1999-11-29
nth sample of a current frame, and n-1 expresses a previous
sample. As indicated in the drawing of Fig.l7, when the
reduction rate r is large, the amplitude suppressing
coefficient G becomes large, and the amplitude adjusting
coefficient g(n) gets large. When the reduction rate r is
small, the amplitudesuppressing coefficient G becomessmall,
and the amplitude adjusting coefficient g(n) gets small.
Further in step S3, the output amplitude 12 from the inverse
1o Fourier transformation unit 5 is expressed in Sin, and an
output signal 14 is calculated by multiplying the amplitude
adjusting coefficient g(n) to the output amplitude 12 by
using an equation (9).
lout ~n~ -'Sin ~n~ X g ~rl
The drawing of Fig.4 illustrates input/output attributes of
the amplitude adjusting filter unit 6. It is apparent from
the drawing that, because the reduction rate r is small for
2o the noise period, the amplitude adjusting coefficient g(n)
is small . An output value for the noise period becomes small
since a suppression gets large. On contrary, because the
reduction rate r is larger for the speech period, the
amplitude adjusting coefficient g(n) is large. An output
value for the speech frame comparatively increases since a
CA 02291826 1999-11-29
suppression gets small.
According to the present embodiment, the amount of reduction
rate of the noise in the frequency domain is reduced for the
noise frame or the speech frame having small power such as
consonant sounds, therefore, the occurrences of musical
noise are suppressed at the noise frame, and the cutoff of
speech periodisprevented. Fig.5illustrates the amplitude
spectrums of noise frame before reduction and after
to reduction.
In cases of intensifying the reduction rate r=1.0, only
highly powered frequency components of amplitude spectrum
of noise sparsely remains after reduction and turns into the
musical noise. In cases of weakening the reduction rate
r=0.5, sparse concentration to some frequencies of the
amplitude spectrum of noise after reduction is prevented,
and the musical noise does not occur. In the frame where
the amount of reduction is decreased, the amplitude is
2o suppressed responding to. it, therefore, an insufficient
amount of the noisesuppressionisprevented. Additionally,
the amplitude adjusting filter unit 6 of the present
embodiment gradually changes the amplitude adjusting
coefficients in a direction of time axis for every samples
sequentially inside a frame by using the equation (8),
26
CA 02291826 1999-11-29
therefore, a natural output is obtained in spite of a sudden
adjusting of the amplitude such as onset of the speech period.
Embodiment 2.
The amplitude adjustment can also be implemented to the
frequency domain. That is, the amplitude adjusting filter
unit is provided after the Fourier transformation unit so
that the inverse Fourier transformation is applied to a
signal after the amplitude is adjusted.
A method of the embodiment 2 of the present invention is
described with reference to the drawings.
The drawing of Fig.6 illustrates a block chart showing the
configuration of spectrum subtraction method of the present
embodiment.
An operation of this method is described with reference to
the drawing of Fig.6.
Fig. 6 illustrates a configuration for adjusting an amplitude
in the frequency domain for which the amplitude adjusting
filter of Fig.l is relocated immediately after the
subtraction filter 4. Operation of all other parts is same
as the operation of embodiment 1.
27
CA 02291826 1999-11-29
In the drawing, the subtraction filter unit 4 calculates the
reduction rate by using the equation (6), and based on a
calculated reduction rate, the estimated noise amplitude
spectrum 10 is subtracted from the input amplitude spectrum
9, and the subtraction filter unit 4 outputs an output
amplitude spectrum. The amplitude adjusting filter unit 15
outputs a final output amplitude spectrum by multiplying the
amplitude adjusting coefficient calculated based on the
to ~ reduction rate r to the output amplitude spectrum.
According to the present embodiment, since the intensity of
reduction is small for the noise frame or the speech frame
having small power such as consonant sounds, therefore, the
1s musical noise occurrence at the noise frame is suppressed,
which prevents diminishing or distorting of consonant sounds
period. In addition to that, in the frame with reduced
amount of reduction, the amplitude is suppressed depending
on the amount of reduction, such that the insufficient amount
20 of noise suppression is prevented.
Also, the amplitude adjusting coefficient needs be
calculated only once for each frame. In this embodiment,
there is no need to calculate the amplitude adjusting
25 coefficient for each of signal samples as in the step S3 of
28
CA 02291826 1999-11-29
Fig.2 of the embodiment 1. Therefore, the present
embodiment is implemented by small calculation amount.
Embodiment 3.
In the previous embodiment, the input signal is divided into
fixed frames and speech is extracted for every frames.
Inevitably, the input signal becomes non-continuous at
divided points of the frames, and this may result in the
occurrence of discontinuous feeling sounds. This
1o embodiment attempts to improve from this by smoothening a
change in the input signal at the divided points between the
frames.
Hereinbelow, the method of the present embodiment is
described with reference to the drawings.
The drawing of Fig.7 illustrates a block chart of the
configuration of spectrum subtraction method for the
embodiment 3 . Most of the elements are same as those of Fig. l .
2o New elements are: an input signal creating unit 19; and a
waveform reforming unit 20.
The drawing of Fig.8 illustrates the example of input signal
7 for the noise reduction method of the embodiment 3. The
drawing of Fig.9 illustrates a weighting function for
29
CA 02291826 1999-11-29
multiplying to the input signal at the input signal creating
unit for the noise reduction method of the embodiment 3.
The drawing of Fig.lO illustrates the example of output
s signal 5a of the inverse Fourier transformation unit 5 for
the noise reduction method of the embodiment 3.
The operation of noise reduction method of the present
embodiment is described with reference to the drawing of
1o Fig.7.
Fig. 8 illustrates a time series amplitude of the input signal
7 after the A/D conversion, which is inputted to the input
signal creating unit 19. Suppose that a frame signal for
15 suppressing noise of a current frame is s(i)(i=n,
n+1, ..., n+fr) , then the input signal creating unit 19 cuts out
s (i) (i=n-a, n-a+1,...,n+fr+sm+a) including signals before and
after the frame signal. A of Fig.8 illustrates a period
before the frame. B and C of Fig.8 illustrate periods after
2o the frame. Next, multiply the s(i)(i=n-a, n-
a+1,...,n+fr+sm+a) to the weighting function of Fig.9, to get
a signal s' (i) (i=n-a, n-a+1,...,n+fr+sm+a) as expressed by an
equation ( 10 ) . The period of A and C of Fig. 8 are weighted
so their edges approach 0. The period of B of Fig.8 i~
25 weighted by 1 which is same weighting as the frame signal.
CA 02291826 1999-11-29
s' (i) =S (i) ~W (i-n-I-a)
(i=n-a, n-a-f-1,..., n~-f r-I-sm-f-a)
(10)
The Fourier transformation unit 1 applies Fourier
transformation to the signal s'(i)(i=n-a, n-
a+1,...,n+fr+sm+a) . Suppose that an output signal of the
inverse Fourier transformation unit 5 is u(i)(i=0,
l0 1,...,2a+fr+sm), then the Fig.lO illustrates u(i), and
timewise, the a (i) (i=0, 1,..., 2a+fr+sm) corresponds to
s(n-a+i) (i=0, 1,...,2a+fr+sm) .
As illustrated in the drawing of Fig. l8, the waveform
reforming unit 20 processes the waveform reforming for a
frame signal u(i)(i=a, a+1,..., a+fr) which has applied the
inverse Fourier transformation so that this frame signal
will be continuous between the frames . A next equation ( 11 )
expresses an output signal after the waveform reforming
2o provided that output u(i) (i=a+fr, a+fr+1,..., a+fr+sm) from
the inverse Fourier transformation of a period B after the
previous frame is up ( i ) ( i=0, 1, ..., sm) .
u' (i) - (u (i) ~ i-hup (i-a)
2s ~ (sm-i)) /sm
31
CA 02291826 1999-11-29
(i=a, a-~1,..., a~-sm) (1 1)
u' (i):signal after waveform reforming process
s According to the present embodiment, a frame signal is
weighted so that its edges approach 0 and the weighted frame
signal is applied the Fourier transformation, therefore, the
noise reduction efficiency will improve. Also, upon
implementing the present system at the fixed decimal point
to calculation digital signal processor (fixed decimal point
DSP, digital signal processor), edges (A and C of Fig. l8)
of the signal where the accuracy is lost due to weighting
are not used in the waveform reforming, therefore, an output
accuracy is secured, at the same time, the discontinuous
i5 feeling sounds occurring at the frame boundaries is
prevented.
When the DSP is used, a fixed decimal point calculation of
short wordlengthsof aboutl6-bitsisperformed. Therefore,
2o accuracy is lost when the DSP processes a signal having a
large dynamic range. By weighting the signal so that its
edges approaches 0, a dynamic range of the signal expands,
such that the accuracy is lost by using the DSP. In this
embodiment, the edges of signal where accuracy is lost are
25 not used in the waveform reforming, therefore, the accuracy
32
CA 02291826 1999-11-29
of waveform reforming improves.
Embodiment 4.
As a method of increasing the noise reduction efficiency for
the spectrum subtraction, the present embodiment considers
a method of reducing the speech cutoff caused by an excessive
reduction and facilitates a method of distinguishing the
noise period.
1o A method of the embodiment 4 of the present invention is
described with reference to the drawings.
The drawing of Fig.ll illustrates a block chart of
configuration of the spectrum subtraction method of the
1s embodiment 4 for the present invention. A new element is
a reduction rate calculating unit 16. Other elements are
identical to those of Fig. 1.
The drawing of Fig. l2 describes the operation of the noise
2o reduction method of the present embodiment. In the drawing,
a bold line indicates an input power of noise frame POWS,
a dotted line indicates an average noise power POWA"E, and
a chained line indicates a noise period deciding threshold
value POWTH. The POWTH is a threshold value of previous noise
2s frame as described in the equation (3).
33
CA 02291826 1999-11-29
An operation of the noise reduction method of the present
embodiment is described with reference to an overall
configuration of Fig.ll.
The present configuration has added the reduction rate
calculating unit 16 to the overall configuration of the
conventionalspectrumsubtraction. In the drawing of Fig.ll,
a filter of the subtraction filter unit 4 performs
1o calculation of the equation ( 1 ) . The noise period deciding
unit 2 calculates the noise period deciding threshold value
POWTH. If an input power underlies this threshold value, it
is decided as the noise frame.
The reduction rate calculating unit 16 calculates an average
power from a plurality of noise frames close to a current
frame based on this decision, and takes it as the average
noise power. The reduction rate calculating unit 16
calculates a ratio rl of the noise period deciding threshold
2o value 17 and the average noise power based on an equation
(12), provided that the noise period deciding threshold
value is POWTH and the average noise power is POWA~E. A
calculated ratio rl is output as a noise variation level 18.
34
CA 02291826 1999-11-29
r 1 = POWAVE (1 2)
P OWTH
The noise variation level r1 is set as the reduction rate
r of noise spectrum of the equation (1) . A higher value of
the noise period deciding threshold value POWTH is used when
the power variation of noise is large so that an input signal
is correctly decided in the noise period. Fig. l2
illustrates the changes in the average noise power and the
noise period deciding threshold value against the change in
the input power of the noise period.
to
According to the present embodiment, when the change in the
noise power is large, a value of the average noise power POWA~
of the noise frame is smaller than the noise period deciding
threshold value POWT", as illustrated in the drawing of Fig.12 .
Therefore, the reduction rate rl must be smaller than 1. As
a result, the reduction by estimated noise amplitude
spectrum is effectively suppressed, and the speech cutoff
is reduced, and an estimated noise amplitude spectrum is
updated correctly.
Embodiment 5.
The present embodiment describes a method of reducing an
unpleasant remaining noise, by suppressing the amplitude of
CA 02291826 1999-11-29
input signal which was decided as a noise of the noise period.
A method of the embodiment 5 of the present invention is
described with reference to the drawings.
The drawing of Fig.l3 illustrates a block chart of the main
elements of the noise reduction method of the embodiment 5
of the present invention.
to The drawing of Fig. l4 illustrates the operation of the noise
reduction method of the present embodiment.
The configuration of the present embodiment reduces the
noise by suppressing an amplitude of the noise period, and
i5 decides the amplitude adjusting coefficient for suppressing
the amplitude depending on the current input power and the
average noise power.
The operation is described with reference to Fig. l3.
A noise period deciding unit 301 calculates an input power
POWS 305 shown by the bold line of Fig. 12 from the input signal
304, and decides a noise period by using the noise period
deciding threshold value POWTH just as in the embodiment 1.
2s Based on a decided noise period, an average noise power
36
CA 02291826 1999-11-29
calculating unit 302 calculates an average power of
plurality of past noise frames that are close to the current
frame, and obtains an average noise power POWA"E 306 which
is shown by the dotted line of Fig. l2.
An amplitude adjusting filter unit 303 calculates the
amplitude suppressing coefficient G from the input power POWS
305 and the average noise power POWAVE 306 of the input signal
by using an equation (13). The amplitude suppression
1o coefficient G is used as an amplitude adjusting coefficient
as it is. The amplitude adjusting filter unit 303 multiplies
the amplitude adjusting coefficient to the input signal 304
and obtains an output signal 307.
G= P OWs-P OWAVE
P OWs
POW s-POW Ave~O)
G=O
(POW S-POW Ave~O~ (13)
So that the amplitude adjusting coefficient will change
smoothly among the frames, by using the amplitude
suppressing coefficient G calculated~by the equation (13),
a final amplitude adjusting coefficient g(n) can be
calculated. The final amplitude adjusting coefficient g(n)
is calculated by an operation of smoothening the change in
the amplitude suppressing coefficient G for each of samples
37
CA 02291826 1999-11-29
as in the equation (8).
The drawing of Fig. l4 illustrates a relation of output power
of the output signal 307 after adjusting the amplitude and
s the input power POWS 305 of the input signal 304.
The dotted line of Fig. l4 illustrates a case of outputting
the input signal as it is, without using the amplitude
adjusting filter. That is, the dotted line illustrates a
1o case of "the input power of the input signal = output power
of the output signal". The bold line illustrates a relation
of the output power and the input power POWs in cases of using
the amplitude adjusting filter. The bold line shows a
smaller output value than the output value shown by the dotted
i5 line due to the amplitude suppression. The bold line also
shows that the output signal power is zero when the input
power is smaller than the average noise power POWAyE-
According to the present embodiment, the reduction rate is
2o made variable, and with the amplitude adjusting filter unit
for changing a degree of the amplitude suppression based on
the varying reduction rate is provided, therefore, the
speech output is stabilized without having to change the
amplitude suppression extensively and frequently as in the
2s case of directly deciding the amplitude adjusting
38
CA 02291826 1999-11-29
coefficient from a decided state of the speech.
Industrial Applicability
As described previously, according to the present invention,
not only the reduction rate is variable, it is provided with
the amplitude adjusting filter unit for changing the degree
of suppressing the amplitude based on the variable reduction
rate, therefore, even if the intensity of reduction at noise
frame is reduced for a purpose of preventing the cutoff of
Zo noise frame, the noise is suppressed by suppressing the
amplitude corresponding to the reduction rate. As a whole,
the present invention is effective in obtaining a well-
balanced and easy-to-hear output.
1s According to the present invention, the amplitude adjustment
is performed in the frequency domain, therefore, there is
no need to calculate the amplitude adjusting coefficients
for each of signal samples. The present invention is
effective in reducing the amount of calculation.
According to the present invention, the amplitude adjusting
coefficient gradually changes in the direction of time axis,
therefore, a natural output is obtained even though there
may be a sudden amplitude adjusting such as onset of the
2s speech period.
39
CA 02291826 1999-11-29
According to the present invention, since the present
invention is provided with the input signal creating unit
which multiplies a current frame signal and a signal in the
periods before and after the current frame by the weighting
function and performs addition of the weighted current frame
signal, therefore, the noise reduction efficiency will be
higher by improving an estimated accuracy of the noise
spectrum by weighting the signals for applying to the Fourier
1o transformation. Also, in case that non-weighted signal of
the inverse Fourier transformation is output as the frame
signal, this is effective in obtaining a highly accurate
signal and is effective in performing a small calculation
amount even by using the fixed decimal point calculation
i5 digital signal processor.
According to the present invention, since the intensity of
noise removal is adjusted depending on a variation
characteristic of period which was decided as the noise
2o period, the present invention is effective in pursuing and
setting a correct noise deciding threshold value even for
the cases of noise period having a large variation. The
invention is also effective in preventing the speech cutoff
due to the added speech components getting mixed.
40
CA 02291826 1999-11-29
According to the present invention, since a gradually
changing average power of noise period is used in determining
the amount of amplitude suppression instead of determining
an amount of amplitude suppression directly from the noise
period, the present invention is effective in avoiding a bad
influence caused by frequent changes in the noise period
deciding output.
41
