JPH05158494

Patent Translate
Powered by EPO and Google
Notice
This translation is machine-generated. It cannot be guaranteed that it is intelligible, accurate,
complete, reliable or fit for specific purposes. Critical decisions, such as commercially relevant or
financial decisions, should not be based on machine-translation output.
DESCRIPTION JPH05158494
[0001]
BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a
noise reduction device for reducing noise mixed into a microphone from the surroundings when
a voice signal or the like is collected by the microphone.
[0002]
2. Description of the Related Art In a communication device or a recording device, etc. provided
with a microphone for picking up audio signals such as voice and music in general, noise mixed
in the microphone is suppressed to make it easy to hear received voice and reproduced voice etc.
In order to do this, a noise reduction device is used. The noise reduction device may be applied to
a voice recognition device and used for the purpose of reducing false recognition due to noise
mixing.
[0003]
As this noise reduction device (noise canceller), an adaptive noise reduction device using an
adaptive filter is known in which filter characteristics are adaptively controlled according to an
input signal. Here, an example of the adaptive noise reduction device conventionally used will be
described with reference to FIG.
08-05-2019
1
[0004]
In FIG. 5, the voice or the like from the signal source 1 is received by the microphone 2, and the
noise from the noise source 3 is received by the microphones 2 and 4, respectively. A signal
component received from the signal source 1 by the microphone 2 is s. The microphone 2 also
receives the noise component na from the noise source 3 uncorrelated with the signal
component s. An input s + na obtained by adding the signal component s and the noise
component na is called a main input. On the other hand, another microphone 4 receives from the
noise source 3 a noise component nb which is uncorrelated with the signal component s but
correlated with the noise component na. This microphone input nb is called a reference input.
[0005]
The main input s + na is sent to the adder 5. The reference input nb is sent to the adaptive filter 6
as an input x and filtered to be an output y. The filter output y is sent to the adder 5 as a
subtraction signal and subtracted from the main input, and the residual e is taken out from the
output terminal 7 and becomes the output of the adaptive noise reduction device of FIG. The
adaptive filter 6 operates to minimize the power of this residual e. That is, the adaptive filter 6
generates a pseudo output (pseudo noise) y of the noise component na by learning based on the
reference input nb, and cancels the noise component na by subtracting this from the main input s
+ na. That's why. As the adaptive filter output y approaches the noise component na, the residual
e approaches the signal component s.
[0006]
By the way, for example, an FIR (finite impulse response) filter is used as the above-mentioned
adaptive filter 6, and this FIR filter approximates the transfer characteristic of the path from the
microphone 2 to the microphone 4 as a linear characteristic. . Therefore, although the
approximation of the transfer characteristic is correctly performed, the noise will be completely
eliminated, but in practice, the approximation can often only be incompletely performed.
[0007]
08-05-2019
2
One of the reasons is that the circuit scale of the FIR filter, in particular, the number of taps can
not be made sufficiently large. The accuracy of the approximation can be improved simply by
increasing the number of taps, but in order to make the learning converge as fast as when the
number of taps is small, it is necessary to further increase the operation speed. From this point of
view, at present, it is difficult to realize the problem due to the restriction due to the size of
hardware and the restriction in terms of cost due to the requirement of an expensive high speed
device.
[0008]
Another reason is that the transfer characteristics to be approximated are not necessarily linear.
This is because, for example, distortion occurs when the quality of the microphone is poor or the
microphone is used beyond the operating range, so that the characteristics appear in a non-linear
manner.
[0009]
As described above, if the approximation of the transfer characteristic by the adaptive filter 6 is
not perfect in the adaptive noise reduction apparatus, a residual noise component correlated with
the noise component na remains in the residual e. Of course, the power of this residual noise
component is smaller than the power of the mixed noise component na itself in the abovementioned main input, but it may have the following harmful effects.
[0010]
That is, for example, when the frequency component of the residual noise is unevenly distributed
in a specific band, even a small level may be very offensive to hearing. In this case, the adaptive
noise reduction device does not achieve the original purpose of making it easy to hear voice and
the like. Furthermore, such noise localized in a specific frequency band may unbalance the
frequency distribution by being superimposed on signal components such as speech. For
example, in many speech recognition devices, the feature of the frequency distribution of speech
is extracted and used as the material of the determination, so that the change of the frequency
distribution of speech may be a cause of misrecognition.
08-05-2019
3
[0011]
The present invention has been made in view of such circumstances, and can prevent adverse
effects due to residual noise included in the noise reduction output being localized in a specific
frequency band, and adaptively regardless of the nature of the mixed noise An object of the
present invention is to provide a noise reduction device that makes it easy to hear or understand
the voice that has passed through the noise reduction device, and that prevents false recognition
even when applied to a voice recognition device.
[0012]
SUMMARY OF THE INVENTION A noise reduction apparatus according to the present invention
is provided with a first microphone provided to receive a signal from a signal source and a noise
from the noise source. A filter based on a second microphone, subtraction means supplied with a
first input obtained and received by said first microphone, and a second input obtained and
received with said second microphone An adaptive filter for controlling the filter characteristics
so as to send the processed output to the subtraction means and minimize the power of the
output from the subtraction means, and random for masking residual noise components in the
output from the subtraction means The problem described above is solved by including noise
generating means for generating noise.
[0013]
Here, it is considered that random noise from the noise generation means is superimposed on the
output from the subtraction means, and as this superposition position, a position before sending
the output of the subtraction means to the adaptive filter, and The position is considered.
The random noise may be superimposed on the first input sent from the first microphone to the
subtracting means.
The random noise may be superimposed on the filtered output sent from the adaptive filter to
the subtracting means. Although the adaptive filter has a filter unit and an adaptive algorithm
unit, the random noise may be superimposed on a second input from the second microphone
supplied to the adaptive algorithm unit. Furthermore, the random noise may be superimposed on
the filter coefficient used in the filter unit of the adaptive filter.
[0014]
08-05-2019
4
The residual noise component from the subtraction means is masked by random noise from the
noise generation means, uneven distribution to a specific frequency is prevented, voices etc.
become easy to hear, and false recognition of the speech recognition device is prevented. Be
done.
[0015]
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS First, FIG. 1 is a block circuit
diagram showing a schematic configuration of a noise reduction apparatus according to an
embodiment of the present invention.
In FIG. 1, the voice or the like from the signal source 11 is received by the microphone 12, and
the noise from the noise source 13 is received by the microphones 12 and 14, respectively. The
microphone 12 receives a signal s + na on which the signal component s from the signal source
11 and the noise component na from the noise source 13 are superimposed (added), which is
referred to as a main input. The noise component nb from the noise source 13 is obtained in the
microphone 14, and this is referred to as a reference input. The signal component s and the noise
components na and nb are uncorrelated, but the noise components na and nb are correlated.
[0016]
The main input s + na obtained by being received by the microphone 12 is sent to the adder 15
which is a subtracting means. The reference input nb received and obtained by the microphone
14 is sent to the adaptive filter 16 as the input x and filtered to be the output y. The filter output
y is sent as a subtraction signal to the adder 15 and subtracted from the main input s + na, and
the residual e is sent to the adaptive filter 16 and taken out from the output terminal 17. The
adaptive filter 16 adaptively controls the filter characteristic so as to minimize the power of the
residual e.
[0017]
Here, an adder 18 for superimposing random noise is inserted and connected between a point at
which the output (the above residual e) from the adder 15 as subtraction means is sent to the
adaptive filter 16 and the output terminal 17 ing. The random noise generator 19 outputs, for
08-05-2019
5
example, white random noise that has no correlation with the noise components na and nb, is
sent to the adder 18, and is superimposed on the residual e output from the adder 15. Ru. The
residual output on which the random noise is superimposed is taken out from the output
terminal 17 as the output of the adaptive noise reduction apparatus shown in FIG.
[0018]
As a result, residual noise included in the residual output is masked by the random noise, and for
example, noise components other than voice can be regarded as white noise. In general, white
noise is less audiblely unpleasant than noise in which a specific band is emphasized. In addition,
since whiteness noise is uniformly averaged in frequency distribution, it is difficult to affect the
speech analysis result of the speech recognition device, and adverse effects such as
misrecognition due to residual noise localized in a specific frequency band are effectively made.
It will be possible to prevent.
[0019]
By the way, the adaptive filter 16 generates a pseudo output (pseudo noise) y of the noise
component na by learning based on the reference input nb, and as shown in FIG. And a section
22. The reference input nb is supplied as the input x to the filter unit 21 and the adaptive
algorithm unit 22 through the terminal 16a. The output y from the filter unit 21 is taken out as
the output of the adaptive filter 16 through the terminal 16b, and is sent to the adder 15, which
is the subtraction means. The residual e from the adder 15 is supplied to the adaptive algorithm
unit 22 through the terminal 16c. The adaptive algorithm unit 22 changes the filter coefficient of
the filter unit 21 to change the filter characteristic, thereby subtracting the output y obtained by
filtering the input x from the main input s + na. Adaptive control to minimize the power of
[0020]
FIG. 2 shows the internal configuration of the adaptive filter 16, and shows a specific example
using a so-called FIR (finite impulse response) filter as the filter unit 21. As shown in FIG. In FIG.
2, the reference input x from the input terminal 16a is sent to a series circuit of delay elements
231, 232,..., 23L according to the number of taps. The input x0 from the input terminal 16a and
the outputs x1, x2, ..., xL from the delay elements 231, 232, ..., 23L are coefficient multipliers
240, 241, 242, ..., 24L, respectively. , And are multiplied by the filter coefficients w 0, w 1, w 2,...,
08-05-2019
6
W L and sent to the adder 25. The filter coefficients w0, w1, w2,..., WL are corrected by the
coefficient correction signal from the adaptive algorithm unit 22, and the output y from the
adder 25 is taken out from the output terminal 16b.
[0021]
As the adaptive algorithm used in the adaptive algorithm unit 22, many techniques have been
proposed. As a specific example, an LMS (least mean square, least mean square) algorithm will be
described.
[0022]
The input data at the k-th sample cycle time point (time k) of the data series of the input x and
the delayed output data from each of the delay elements 231, 232, ..., 23L are xk0, xk1, xk2, ... ..,
XkL, an input vector Xk subjected to FIR filter processing is set as Xk = [xk0 xk1 xk2 ... xkL] T (1).
T in this equation (1) represents a transpose. Assuming that the above filter coefficients
(weighting coefficients) are wk0, wk1, wk2,..., WkL for this input vector Xk, and the FIR filter
output is yk, the relationship between input and output is given by the following equation (2) It
will be. yk = wk0xk0 + wk1xk1 +... + wkLxkL (2) Further, if the filter coefficient vector (weighting
vector) Wk is defined as Wk = [wk0 wk1 wk2 wkL] T (3), input / output The relationship is
described as yk = XkTWk (4). Assuming that the desired response is dk, the error .epsilon.k from
the output is expressed as .epsilon.k = dk -yk = dk -Xk T Wk (5). Using these, the LMS algorithm
is expressed as Wk + 1 = Wk + 2μεk Xk (6). (6) is a gain factor that determines the speed and
stability of adaptation.
[0023]
The FIR filter approximates the transfer characteristic of the spatial path from the second
microphone 14 to the first microphone 12 in FIG. 1 as a linear characteristic. Therefore, if the
accuracy of the approximation is increased, the adaptive filter output y approaches the noise
component na, and the noise component na can be canceled by subtracting this from the main
input s + na, and the residual e is It approaches the signal component s.
08-05-2019
7
[0024]
However, in many cases, the above approximation can only be incompletely performed due to the
fact that the number of taps of the FIR filter (filter unit 21) can not be made sufficiently large or
the transfer characteristic to be approximated is not necessarily linear. As described above, if the
approximation of the transfer characteristic by the adaptive filter 16 can not be completely
performed, a residual noise component correlated with the noise component na remains in the
residual e. When the frequency component of this residual noise is unevenly distributed in a
specific band, it is not preferable because it may be aurally disturbing even if it is a small level, or
it may cause misrecognition when applied to a speech recognition device. That is as described
above.
[0025]
On the other hand, as in the embodiment of the present invention, the white noise random noise
from the random noise generator 19 is superimposed only on the output of the output (the
residual e) from the adder 15 which is the subtraction means. The residual noise is prevented
from being unevenly distributed in the specific frequency band by taking it out from the output
terminal 17. The random noise is provided in an amount sufficient to mask the residual noise so
that components other than the output audio signal s can be regarded as white random noise.
White noise has almost uniform frequency distribution, no offensive feeling like noise localized in
a specific band, less unpleasant hearing, and less adverse effect on speech recognition device .
[0026]
By the way, the position to which the random noise from the random noise generator 19 is added
is not limited to the insertion position of the adder 18, but may be, for example, any position of
each point P2 to P5 in FIG. That is, in FIG. 3, by inserting and connecting an adder at the position
of each point P2 to P5 and adding random noise, the second to fifth embodiments of the present
invention can be obtained respectively. The other configuration is the same as that of the first
embodiment of the present invention shown in FIG. 1 and, therefore, the corresponding portions
are denoted by the same reference numerals and the description thereof will be omitted.
[0027]
First, as a second embodiment of the present invention, an adder (the adder 18 in FIG. 1) is
08-05-2019
8
inserted and connected to a point P2 between the first microphone 12 and the adder 15 as the
subtracting means. There is a configuration in which random noise from the noise generator
(random noise generator 19 in FIG. 1) is added to the adder.
[0028]
Next, as a third embodiment of the present invention, an adder is inserted and connected to a
point P3 immediately after the adder 15 as the subtraction means, and a random noise is added
to the adder.
At this time, the adaptive algorithm unit 22 of the adaptive filter 16 is supplied with the residual
e from the adder 15 in which the random noise is superimposed.
[0029]
Next, as a fourth embodiment of the present invention, random noise is superimposed on the
output from the filter section 21 of the adaptive filter 16 at the position of point P4, and this is
supplied to the adder 15 as a subtraction signal. Configuration is mentioned.
[0030]
In the case of the second to fourth embodiments, unlike the first embodiment, the superimposed
random noise affects the filter coefficient of the adaptive filter 16.
That is, since the filter coefficient (weighting coefficient) of the filter unit 21 is updated based on
the residual to which random noise is added, it is slightly shaken. As a result, the frequency
component of the filter output of the adaptive filter 16 has a minute component dispersed in a
wide band in addition to the main component of the approximate noise of the noise component
na. The amount of random noise to be added to the positions of the points P2 to P4 is
appropriately adjusted, and residual noise included in the residual after subtraction of the filter
output from the main input is white noise and random noise. Make it look like. As a result, as in
the first embodiment, the voice after the noise reduction processing can be easily heard, and
erroneous recognition can be prevented when applied to the voice recognition device.
08-05-2019
9
[0031]
Furthermore, as a fifth embodiment of the present invention, a configuration in which random
noise is superimposed on the reference input at the position of the point P5 immediately before
the adaptive algorithm section 22 of the adaptive filter 16 can be mentioned. In the case of the
fifth embodiment, the reference input supplied to the adaptive algorithm unit 22 is slightly
shaken, and the same effect as that of the first to fourth embodiments can be obtained.
[0032]
Next, FIG. 4 shows a specific configuration of the adaptive filter used in the sixth embodiment of
the present invention, and using the adaptive filter of the configuration of FIG. A reduction device
(but excluding the adder 18 and the random noise generator 19) can be configured. The same
reference numerals as in FIG. 2 denote the same parts in FIG. 4 and a description thereof will be
omitted.
[0033]
In FIG. 4, in the transmission path of each filter coefficient from the adaptive algorithm 22 to
each coefficient multiplier 240, 241, 242,..., 24L of the filter unit 21, an adder 280, 281, 282,.
28L are inserted and connected, and random noises from the random noise generator 29 are
sent to and superimposed on these adders 280, 281, 282,..., 28L. Thereby, each filter coefficient
w0 ', w1', w2 ', ..., wL' on which random noise is superimposed is sent to each coefficient
multiplier 240, 241, 242, ..., 24L. become.
[0034]
Therefore, random noise is superimposed on the coefficient after being corrected by the abovedescribed adaptive algorithm, so that each coefficient that determines the filter characteristic of
the filter unit 21 is slightly shaken, and the frequency of residual noise of the filter output The
components have minute components dispersed in a wide band. Therefore, the same effect as the
first to fifth embodiments can be obtained.
08-05-2019
10
[0035]
The present invention is not limited to the above embodiment. For example, the specific
configuration of the filter unit 21 and the algorithm used for the adaptive algorithm unit 22 are
not limited to the FIR filter and LMS algorithm of the above embodiment. . Also, the noise from
the noise generator is not limited to white noise.
[0036]
As is apparent from the above description, according to the noise reduction device according to
the present invention, the input from the microphone for receiving the signal from the signal
source is sent to the subtraction means, and the noise reduction device from the noise source An
input from a microphone for receiving noise is sent to the adaptive filter, an output from the
adaptive filter is sent to the subtracting means, and the filter characteristic of the adaptive filter
is adjusted to minimize the power of the output from the subtracting means. Since the residual
noise component in the output from the subtraction means is masked by random noise from the
noise generation means while being controlled, the residual noise component in the output from
the subtraction means is masked and unevenly distributed to a specific frequency Is made easy to
hear voice and the like, and false recognition of the voice recognition device is prevented.
08-05-2019
11