JP3128870B2 - Noise reduction device - Google Patents

Abstract

PURPOSE:To make it easier to listen to signals such as voice sound even if remaining noise in remainder is maldistributed to a specific frequency range so as to prevent erroneous recognition when used in a speech recognizing device. CONSTITUTION:Related to the main input received by a microphone 12, signal content s from a signal source 11 is mingled with noise content na from a noise source 13. A microphone 14 receives noise content nb from the noise source 13. The output y similar to the noise content na is obtained by using an adaptive filter 16. This is subtracted from the main input to obtain remainder e similar to the signal content s. The remainder e is superimposed with white color-like random noise from a random noise generator 19 to mask the remaining noise in the remainder e.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise reduction device for reducing noise (noise) mixed into a microphone from the surroundings when a sound signal or the like is picked up by the microphone.

[0002]

2. Description of the Related Art Generally, in a communication device or a recording device equipped with a microphone for picking up an audio signal such as voice or music, noise received in the microphone is suppressed to make it easy to hear a received voice or a reproduced voice. To do
A noise reduction device is used. This noise reduction device
It is sometimes applied to a speech recognition apparatus for the purpose of reducing erroneous recognition due to noise contamination.

[0003] This noise reduction device (noise canceller)
There is known an adaptive noise reduction device using an adaptive filter whose filter characteristics are adaptively controlled according to an input signal. Here, an example of a conventionally used adaptive noise reduction device will be described with reference to FIG.

In FIG. 5, sound from a signal source 1 is received by a microphone 2, and noise from a noise source 3 is received by microphones 2 and 4, respectively. Signal source 1
Is a signal component received by the microphone 2 from s.
The microphone 2 receives _a noise component na from the noise source 3 that is uncorrelated with the signal component s. The input s + n _a plus and these signal components s and a noise component n _a is referred to as a primary input. Meanwhile, another microphone 4, but no correlation with the signal component s will be receiving the noise component n _b correlated to the noise component n _a from the noise source 3. It is referred to as a reference input to the microphone input n _b.

[0005] The primary input s + n _a is fed to the adder 5. The reference input _nb is sent to the adaptive filter 6 as an input x, is filtered, and becomes an output y. The filter output y is sent to the adder 5 as a subtraction signal, and is subtracted from the main input. 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 so as to minimize the power of the residual e. In other words, the adaptive filter 6 based on the reference input n _b, learned by generating a pseudo output (pseudonoise) y of the noise component n _a, the noise component n by subtracting it from the main input s + n _{a It} tries to offset a. Greater the adaptive filter output y approaches the noise component n _a, the residual e becomes closer to the signal component s.

[0006]

Incidentally, for example, an FIR (finite impulse response) filter is used as the adaptive filter 6, and this FIR filter is
The transfer characteristic of the path from to the microphone 4 is approximated as a linear characteristic. Therefore, if the transfer characteristic is accurately approximated, the noise will be completely removed. However, in practice, this approximation can often be performed only incompletely.

One of the reasons is that the circuit scale of the FIR filter, especially the so-called tap number, cannot be made sufficiently large. If the number of taps is simply increased, the accuracy of the approximation will be increased, but in order to converge the learning at the same speed as when the number of taps is small, it is necessary to further increase the calculation speed. Under such circumstances, it is currently difficult to achieve the above-mentioned problems due to restrictions on the size of the hardware and restrictions on the cost due to the need for expensive high-speed devices.

Another reason is that the transfer characteristic to be approximated is not necessarily linear. this is,
For example, when the quality of the microphone is poor, or when the microphone is used out of the operation range, distortion occurs, which is caused by non-linear characteristics.

[0009] Approximation of the transfer characteristic by the adaptive filter 6 in this way adaptive noise reduction system is not complete, so that the residual noise component having a correlation with the noise component n _a in the residuals e remains. Of course, the power of the residual noise component, although smaller than the power of the mixed noise component n _a per se in the main input, sometimes exerts the following harm.

That is, for example, when the frequency components of the residual noise are unevenly distributed in a specific band, even a small level may be very unpleasant in audibility. In this case, the adaptive noise reduction device does not achieve the original purpose of making it easy to hear voices and the like. further,
Noise that is unevenly distributed in a specific frequency band as described above may cause the frequency distribution to be unbalanced by being superimposed on a signal component such as voice. For example, in many voice recognition devices, since the characteristics of the frequency distribution of the voice are extracted and used as a material for determination, a change in the frequency distribution of the voice may be a cause of erroneous recognition.

The present invention has been made in view of such circumstances, and can prevent an adverse effect due to uneven distribution of residual noise included in noise reduction output in a specific frequency band. In addition, an object of the present invention is to provide a noise reduction device that makes it easy to hear or understand voices passed through an adaptive noise reduction device, and that does not easily cause erroneous recognition when applied to a voice recognition device.

[0012]

SUMMARY OF THE INVENTION A noise reduction device according to the present invention is provided with a first microphone provided for receiving a signal from a signal source, and for receiving noise from a noise source. A second microphone, and an FIR to which a time-series output signal from the second microphone is supplied.
(Finite impulse response) filter section, subtraction means for subtracting the time series output signal of the FIR filter section from the time series output signal from the first microphone, and the time series output signal from the second microphone. Adaptive control means for controlling the characteristics of the FIR filter unit based on the time series output signal from the subtraction means so as to minimize the power of the time series output signal from the subtraction means; and noise for generating random noise. Generating means, and adding means for adding the random noise from the noise generating means only to the output of the time series output signal from the subtracting means, by the random noise output from the noise generating means, By masking the residual noise component unevenly distributed in a specific band in the time-series output signal from the subtraction means, To resolve.

Here, the random noise from the noise generating means is superimposed on the output from the subtracting means. In addition, the superimposed position of the random noise is obtained by using the output of the subtracting means in the adaptive mode. The position before sending to the control means and the position after it are considered. Further, the random noise may be superimposed on a filter processing output sent from the filter unit to the subtraction means. Further, the random noise may be superimposed on the first input sent from the first microphone to the subtraction means. The random noise may be superimposed on a second input from the second microphone supplied to the adaptive control means. further,
The random noise may be superimposed on a filter coefficient used in the filter unit.

[0014]

The residual noise component from the subtracting means is masked by random noise from the noise generating means, thereby preventing uneven distribution to a specific frequency, making it easier to hear voices and the like, and preventing erroneous recognition by a voice recognition device. Is done.

[0015]

FIG. 1 is a block circuit diagram showing a schematic configuration of a noise reduction device according to an embodiment of the present invention. In FIG. 1, sound and the like from a signal source 11 are transmitted to a microphone 12, and noise from a noise source 13 is transmitted to a microphone 12.
And 14 respectively. The microphone 12, the signal and the signal component s from the signal source 11, and a noise component n _a of the noise source 13 is superimposed (added) s + n _a
Are received and are referred to as primary inputs. Microphone 14
, A noise component n _b from the noise source 13 is obtained, which is referred to as a reference input. A signal component s and a noise component n _a ,
Although a n _b are uncorrelated, there is a correlation between the noise component n _a and n _b.

The main input s + n _a received by the microphone 12 is sent to an adder 15 which is a subtraction means. The reference input n _b obtained by being received by the microphone 14 is sent as an input x to the adaptive filter 16 is filtered, the output y. The filter output y is sent as a subtraction signal to the adder 15 is subtracted from the primary input s + n _a, the residuals e is taken out from an output terminal 17 with are sent to the adaptive filter 16. The adaptive filter 16 adaptively controls the filter characteristics so as to minimize the power of the residual e.

Here, an adder 18 for superimposing random noise is provided between an output terminal 17 and a point at which the output (residual e) from the adder 15 serving as a subtracting means is sent to the adaptive filter 16. Inserted and connected. The random noise generator 19 outputs, for example, white random noise having no correlation with each of the noise components n _a and n _b , is sent to the adder 18, and outputs the residual e output from the adder 15. Superimposed. 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 device shown in FIG.

As a result, the residual noise included in the residual output is masked by the random noise. For example, noise components other than voice can be regarded as white noise. Generally, white noise has less auditory discomfort than noise in which a specific band is emphasized. In addition, since the white noise has a uniform frequency distribution, it does not easily affect the voice analysis result of the voice recognition device.
It is possible to effectively prevent adverse effects such as erroneous recognition due to residual noise unevenly distributed in a specific frequency band.

By the way, the adaptive filter 16 based on the reference input n _b, is intended to generate a pseudo output (pseudonoise) y of the noise component n _a by learning, as shown in FIG. 1, the filter And an adaptive algorithm unit 22. The above reference input n
_b is input to the filter unit 21 via the terminal 16a as input x.
And the adaptive algorithm unit 22. An output y from the filter section 21 is taken out as an output of the adaptive filter 16 via a terminal 16b, and sent to the adder 15 as the subtraction means. The residual e from the adder 15 is
The signal is supplied to the adaptive algorithm unit 22 via the terminal 16c. Adaptive algorithm section 22, by varying the filter coefficients of the filter unit 21 changes the filter characteristics, residual output y obtained input x to filter obtained by subtracting from the primary input s + n _a The adaptive control for minimizing the power of e is performed.

FIG. 2 shows the internal configuration of the adaptive filter 16, and shows a specific example in which a so-called FIR (finite impulse response) filter is used as the filter unit 21. In FIG. 2, the reference input x from the input terminal 16a is composed of delay elements 23 ₁ , 23 ₂ ,.
, Sent to a 23 _L series circuit. Input terminal 16a
Input x ₀ and the delay element 23 from _1, 23 _2, ...
-, each output x ₁ from the _{23 L, x 2, ···,} x L is,
The coefficient multipliers 24 ₀ , 24 ₁ , 24 ₂ ,.
24 _L , respectively, and filter coefficients w ₀ , w ₁ , w
_2, ..., is multiplied by w _L, it is sent to the adder 25. Each filter coefficient w ₀ , w ₁ , w ₂ ,..., W _L
Is corrected by the coefficient correction signal from the adaptive algorithm unit 22, and the output y from the adder 25 is extracted from the output terminal 16b.

As the adaptive algorithm used in the adaptive algorithm unit 22, many methods have been proposed. One specific example is LMS (Least Mean Square,
The least mean square) algorithm will be described.

The input data and the delay elements 23 ₁ in the k-th sample period time of the data sequence of the input x (time k), 23 _2, ···, each delay output data from the 23 _L, respectively x _k0 , X _k1 , x _k2 ,..., X _kL , the input vector X to be FIR filtered
a _k, put the X _k = _{[x k0 x k1 x k2 ··· x} kL ] ^T ··· (1). T in the equation (1) indicates a transposed symbol. With respect to this input vector X _k , assuming that the above filter coefficients (weighting coefficients) are w _k0 , w _k1 , w _k2 ,..., W _kL and the FIR filter output is y _k , the input / output relationship is as follows. (2)
It looks like an expression. _{y k = w k0 x k0 +} w k1 x k1 + ··· + w kL x kL ··· (2) In addition, the filter coefficient vector (weight vector) W
_{If k} is _defined as W _k = [w _k0 w _k1 w _k2 ... w _kL ] ^T (3), the input / output relationship is y _k = X _k ^T W _k (4) It is described as follows. If the desired response and d _k, the error epsilon _k and the output is expressed as _{ε k = d k -y k =} d k -X k T W k ··· (5). Using these, the LMS algorithm is expressed as follows: W _{k + 1} = W _k +2 με _k X _k (6) Μ in the equation (6) is a gain factor that determines the speed and stability of adaptation.

The FIR filter is the second filter in FIG.
Of the spatial path from the microphone 14 to the first microphone 12 is linearly approximated. Therefore,
If the accuracy of the approximation increases, the output y of the adaptive filter approaches the noise component n _a ,
+ By subtracting from n _a can offset the noise component n _a, the residual e becomes closer to the signal component s.

However, the above approximation can only be performed imperfectly because the number of taps of the FIR filter (filter unit 21) cannot be sufficiently large, and the impulse response to be approximated is not necessarily linear. There are many. If the approximation of the impulse response by the adaptive filter 16 cannot be completely performed, a residual noise component having a correlation with the noise component na remains in the residual error e. If this residual noise frequency component is unevenly distributed in a specific band, even a small level may cause unpleasantness in hearing or cause erroneous confirmation when applied to a speech recognition device. Is as described above.

On the other hand, as in the embodiment of the present invention, the randomness of whiteness from the random noise generator 19 is limited to the output (residual e) from the adder 15 as the subtracting means. By superimposing the noise and extracting it from the output terminal 17, the residual noise is prevented from being unevenly distributed in a specific frequency band. The random noise is given by 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. The white noise has a substantially uniform frequency distribution, does not have a harsh feeling like noise unevenly distributed in a specific band, has little auditory discomfort, and has little adverse effect on a voice recognition device. .

The random noise generator 19
The position at which the random noise is added from the adder 18
Is not limited to the insertion position, for example the points P ₂ to P ₅ of FIG. 3
May be any of the positions. That is, in FIG. 3, adders are inserted and connected at the positions of the respective points P _{2 to} P ₅ to add random noise, thereby obtaining the _{second to second} aspects of the present invention.
The fifth embodiment 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 the corresponding parts are denoted by the same reference symbols and description thereof will be omitted.

First, as a second embodiment of the present invention, the first microphone 12 and an adder 15 as subtracting means are provided.
Configuration to add random noise from the point inserted and connected adder (adder 18 of FIG. 1) to P _2, (the random noise generator 19 of FIG. 1) noise generator to the adder between No.

Next, as a third embodiment of the present invention, inserted and connected the adder to the point P ₃ immediately after the adder 15 as the subtraction means include configuration to add random noise to the adder . At this time, a signal in which the random noise is superimposed on the residual e from the adder 15 is supplied to the adaptive algorithm unit 22 of the adaptive filter 16.

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 the point P ₄ , and this is subtracted by the adder 15. There is a configuration in which the signal is supplied as a signal.

In the case of these second to fourth embodiments,
Unlike the case of the first embodiment, the superimposed random noise affects the filter coefficient of the adaptive filter 16. That is, the filter coefficient (weighting coefficient) of the filter unit 21 is updated slightly based on the residual to which the random noise is added, and thus slightly fluctuates. As a result, the frequency component of the filter output of the adaptive filter 16, in addition to the main component of the approximate noise of the noise component n _a, will have a dispersed small component in a wide band. Note the amount of random noise added to the position of each point P ₂ to P ₄ is appropriately adjusted, the random residual noise is white with the subtraction of the filter output from the primary input is included in the residuals after performing Make it be considered as noise. As a result, as in the case of the first embodiment, the voice and the like after the noise reduction processing can be easily heard, and the erroneous recognition when applied to the voice recognition device can be prevented.

Furthermore, as a fifth embodiment of the present invention, configured to superimpose random noise to the reference input position of P ₅ points immediately before the adaptive algorithm unit 22 of the adaptive filter 16 and the like. In the case of the fifth embodiment, the reference input supplied to the adaptive algorithm unit 22 is slightly fluctuated, and the same effects as those of the first to fourth embodiments can be obtained.

FIG. 4 shows a specific configuration of the adaptive filter used in the sixth embodiment of the present invention. The adaptive filter having the configuration shown in FIG. 4 is used as shown in FIG. An adaptive noise reduction device (excluding the adder 18 and the random noise generator 19) can be configured.
In FIG. 4, portions common to those in FIG. 2 are denoted by the same reference numerals, and description thereof is omitted.

In FIG. 4, the adaptive algorithm 22
From the coefficient multipliers 24 ₀ , 24 ₁ , 2 of the filter unit 21
4 _2,..., In the transmission path of each filter coefficient of the 24 _L, adders 28 _0, 28 _1, 28 _2, ...,
28 _L , and these adders 28 ₀ , 28 ₁ ,
The random noise from the random noise generator 29 is sent to and superimposed on 28 ₂ ,..., 28 _L respectively. As a result, each coefficient multiplier 24 ₀ , 24 ₁ , 24 ₂ ,.
-, the 24 _L, each filter coefficient random noise is superimposed _{w 0 ', w 1',} w 2 ', ···, w L' would be sent, respectively.

Therefore, random noise is superimposed on the coefficient corrected by the above-described adaptive algorithm, and as a result, each coefficient for determining the filter characteristic of the filter unit 21 is slightly fluctuated, and the residual filter output is reduced. The frequency component of the noise has minute components dispersed in a wide band. Therefore, the same effects as those of the first to fifth embodiments can be obtained.

The present invention is not limited to only the above embodiment. For example, the specific configuration of the filter unit 21 and the algorithm used in the adaptive algorithm unit 22 are not limited to the FIR filter and the LMS algorithm of the above embodiment. It is not limited to. Further, noise from the noise generator is not limited to white noise.

[0036]

As is apparent from the above description, according to the noise reduction apparatus of the present invention, the input from the microphone for receiving the signal from the signal source is sent to the subtraction means, and the signal from the noise source is sent to the subtraction means. An input from a microphone for receiving noise is sent to the adaptive filter, an output from the adaptive filter is sent to the subtraction means, and a filter characteristic of the adaptive filter is set so as to minimize the power of the output from the subtraction means. In addition to the control, since the residual noise component in the output from the subtraction unit is masked by the random noise from the noise generation unit, the residual noise component in the output from the subtraction unit is masked, and uneven distribution to a specific frequency is performed. Is prevented, voices and the like are easily heard, and erroneous recognition of the voice recognition device is prevented.

[0037]

[Brief description of the drawings]

FIG. 1 is a block circuit diagram showing a schematic configuration of an adaptive noise reduction device according to a first embodiment of the present invention.

FIG. 2 is a block circuit diagram showing a specific example of an internal configuration of an adaptive filter used in the embodiment.

FIG. 3 is a block circuit diagram for explaining second to fifth embodiments of the present invention.

FIG. 4 is a block circuit diagram showing a specific example of an internal configuration of an adaptive filter used in a sixth embodiment of the present invention.

FIG. 5 is a block circuit diagram illustrating a schematic configuration of an adaptive noise reduction device.

[Explanation of symbols]

 11... Signal source 12... (First) microphone 13... Noise source 14... (Second) microphone 15... Adder (subtraction) Means 16 adaptive filter 17 output terminal 18 adder 19 random noise generator 21 filter unit 22 Adaptive algorithm section

──────────────────────────────────────────────────の Continued on the front page (51) Int.Cl. ⁷ Identification code FIG10L 9/00 F (72) Inventor Takeshi Hara 6-7-35 Kita-Shinagawa, Shinagawa-ku, Tokyo Inside Sony Corporation (56) References JP-A-2-244098 (JP, A) JP-A-62-135019 (JP, A) JP-A-1-239596 (JP, A) Patent 2862771 (JP, B2) Spring Study of the Acoustical Society of Japan in 1991 Proceedings of the meeting, 2-5-3, Keizaburo Takagi, et al. “Noise-based speech recognition by two-stage spectral subtraction” p. 59-60 (issued on March 27, 1991) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 11/00-21/06 H03H 21/00 H04R 3/00-3/14

Claims (6)

(57) [Claims] A first microphone provided for receiving a signal from a signal source; and a second microphone provided for receiving a noise from a noise source.
And microphone, the a FIR (finite impulse response) filter unit for time-series output signal from the second microphone is supplied, the time-series output signal from the first microphone
Subtracting means for subtracting the time-series output signal of the FIR filter unit, on the basis of the time-series output signal from the time-series output signal and the subtracting means from said second microphone, the time-series output signal from said subtracting means Adaptive control means for controlling the characteristics of the FIR filter unit so as to minimize the power of the FIR filter unit, noise generation means for generating random noise, and the output of the time-series output signal from the subtraction means.
The random noise from the noise generator.
And a random noise output from the noise generating means.
To a specific band in the time-series output signal from the subtraction means.
A noise reduction device characterized in that unevenly distributed residual noise components are masked . 2. An apparatus for receiving a signal from a signal source.
And a second microphone provided for receiving noise from a noise source.
And the time-series output signals from the second microphone are supplied.
A FIR (finite impulse response) filter unit, noise generating means for generating random noise, a time-series output signal from the first microphone,
Addition to add random noise from noise generation means
Means and the FIR filter from the time-series output signal from the adding means.
Subtraction means for subtracting and outputting the time series output signal of the filter unit
And the time-series output signal from the second microphone and the
Based on the time series output signal from the calculating means.
On the power of the time-series output signal from the stage so as to minimize
Adaptive control means for controlling the characteristics of the FIR filter unit.
The random noise output from the noise generating means.
To a specific band in the time-series output signal from the subtraction means.
That the ubiquitous residual noise component is masked
Characteristic noise reduction device. 3. An apparatus for receiving a signal from a signal source.
And a second microphone provided for receiving noise from a noise source.
And the time-series output signals from the second microphone are supplied.
(Finite impulse response) filter unit and a time series output signal from the first microphone.
Subtraction means for subtracting the time series output signal of the FIR filter unit
If the noise generating means for generating a random noise, a time-series output signal from said subtracting means, the noise generating
Add the random noise output from the means and output
Adding means, a time-series output signal from the second microphone,
Based on the time series output signal from the calculating means.
To minimize the power of the time series output signal from the stage.
Adaptive control means for controlling the characteristics of the FIR filter unit.
The random noise output from the noise generating means.
To a specific band in the time-series output signal from the adding means.
That the ubiquitous residual noise component is masked
Characteristic noise reduction device. 4. An apparatus for receiving a signal from a signal source.
And a second microphone provided for receiving noise from a noise source.
And the time-series output signals from the second microphone are supplied.
(Finite impulse response) filter unit, noise generating means for generating random noise, a time-series output signal from the FIR filter unit,
Add random noise output from noise generation means
Adding means for calculating the time series output signal from the first microphone.
Subtraction means for subtracting and outputting the time series output signal of the addition means
And the time-series output signal from the second microphone and the
Based on the time series output signal from the calculating means.
To minimize the power of the time series output signal from the stage.
Adaptive control means for controlling the characteristics of the FIR filter unit.
The random noise output from the noise generating means.
To a specific band in the time-series output signal from the subtraction means.
That the ubiquitous residual noise component is masked
Characteristic noise reduction device. 5. An apparatus for receiving a signal from a signal source.
And a second microphone provided for receiving noise from a noise source.
And the time-series output signals from the second microphone are supplied.
(Finite impulse response) filter unit and a time series output signal from the first microphone.
Subtract and output the time series output signal of the FIR filter unit
Subtraction means, noise generation means for generating random noise, a time-series output signal from the second microphone,
Add the random noise output from the noise generator
Adding means, and a time series output signal from the adding means and the subtracting means.
Based on the time series output signal of
In order to minimize the power of the series output signal,
Adaptive control means for controlling the characteristics of the filter section, wherein the random noise output from the noise generating means is provided.
To a specific band in the time-series output signal from the subtraction means.
That the ubiquitous residual noise component is masked
Characteristic noise reduction device. 6. A device for receiving a signal from a signal source.
And a second microphone provided for receiving noise from a noise source.
And the time-series output signals from the second microphone are supplied.
(Finite impulse response) filter unit and a time series output signal from the first microphone.
Subtract and output the time series output signal of the FIR filter unit
Subtraction means, and a time-series output signal from the second microphone;
Based on the time series output signal from the calculating means.
To minimize the power of the time series output signal from the stage.
An adaptive control unit for controlling the characteristics of the FIR filter unit; and a noise generation unit for generating random noise, wherein the noise coefficient is added to the filter coefficient of the FIR filter unit.
By adding the random noise output from the
To a specific band in the time-series output signal from the subtraction means.
That the ubiquitous residual noise component is masked
Characteristic noise reduction device.