KR20070050694A - Method and apparatus for removing noise of multi-channel voice signal - Google Patents

Abstract

Disclosed are a method and apparatus for removing noise of a multichannel voice signal. The noise removing method comprises the steps of: (a) removing a noise-specific component according to the environment from a voice signal; (b) separating speech and noise from the speech signal from which the noise intrinsic component has been removed; And (c) a post-processing step of removing residual noise remaining in the separated speech.

According to the present invention, by removing the noise at the input terminal of the multi-channel noise cancellation system, as a result, it is possible to obtain improved performance than the noise processing method in a single channel environment and to improve the performance of the entire system.

Description

Method and apparatus for removing noise of multi-channel voice signal {Method and apparatus for removing noise of multi-channel voice signal}

1 is a schematic diagram of a conventional GSC noise cancellation method.

2 is a block diagram illustrating a configuration of an embodiment of a device for removing noise of a multichannel voice signal according to the present invention.

FIG. 3 is a block diagram illustrating a more detailed configuration of the voice & noise separation unit 220 of FIG. 2.

4 is a block diagram illustrating a more detailed configuration of the time delay compensation unit 300 of FIG. 3.

5 is a block diagram illustrating a more detailed configuration of the eigen filtering unit 340.

6 is a flowchart illustrating an embodiment of a method for removing noise of a multichannel voice signal according to the present invention.

FIG. 7 is a flowchart illustrating a more detailed process of the process of separating voice and noise (step 620) of FIG. 6.

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to speech recognition, and more particularly, to a method and apparatus for removing noise of a multichannel speech signal.

Voice recognition or communication systems perform well in noisy or relatively quiet laboratory environments. However, when it is put to practical use and used in actual field, recognition performance is remarkably degraded by various noise factors. In other words, the speech recognition system has a significant performance deterioration due to background noise and interference signals. Therefore, in order to achieve satisfactory performance in a real environment, a speech preprocessing technology that excludes or mitigates the effects of unwanted signals input to the microphone. This is essential. These speech preprocessing techniques have been studied for noise reduction in the past decades. In particular, in the field of speech signal processing, a single microphone based preprocessing algorithm has been treated as a main noise canceling technique due to its low computational complexity and ease of implementation, and its performance has been gradually improved. However, since it is based on the estimation of accurate noise components, it is disadvantageous that effective performance can be obtained only for stable (unchanging) noise when the information on the noise components is insufficient. Therefore, various microphones have been developed to obtain various information about the voice signal and the noise, and to remove the noise or enhance the voice signal. However, the method of removing noise using multiple microphones has the disadvantage of increasing the computational speed and processing speed. Therefore, as a solution to this problem, the technique of removing noise using two microphones is appropriate.

When applying the Adaptive Noise Canceling (ANC) method among the existing multi-channel noise processing methods, there is an assumption that only a noise signal should exist at the input of the reference microphone (second microphone) among the two microphones. This is difficult and also because the two microphones are not completely separated, the reference microphone also contains a lot of the audio signal we want. In practice, in order for the ANC method to be effective, the noise of the primary and reference microphones must be highly correlated. And to satisfy this, two microphones should be installed close to each other. However, in this case, there are many desired voice signals in the reference microphone (called cross-talk interference), and when applying the ANC method, the ANC removes not only the noise but also the desired signal, which is a major factor in performance degradation.

The next most widely known method is the Generalized Sidelobe Canceller (GSC). 1 is a schematic diagram of a conventional GSC noise cancellation method. In FIG. 1, the GSC includes a confusion pattern generator 100, a pure noise pattern generator 102, an adaptive filter 104, and a subtractor 106. The horn pattern generator 100 divides the sum of the two inputs in half to generate the horn of the original voice + noise, and the pure noise pattern generator 102 removes the voice component by the difference between the two inputs and divides it in half. A pure noise difference component of two inputs is generated, and the adaptive filter unit 104 uses the pure noise pattern generated by the pure noise pattern generator 102 as an input to estimate the noise component in the confusion pattern through adaptive filtering. The subtraction unit 106 finally outputs the speech component by subtracting the noise thus estimated from the confusion.

However, if the conventional GSC noise reduction method is applied as it is, since the paths between the two microphones are not ideally identical in a real environment, the pure noise pattern generator 102 necessarily includes the voice leakage signal in the noise component. . This interferes with estimating the correct noise in the adaptive filter section 104 of the GSC and subtracts the original speech component in the final subtraction step in the subtractor 106, which in turn hinders the performance of the overall noise reduction system. have.

Therefore, there is inevitable cross-talk interference in which there are many voice signals that we want. As a way to solve this problem, since we cannot completely separate the two microphones or obtain a noise signal in the reference microphone, New noise processing methods are needed to deal with this.

SUMMARY OF THE INVENTION The present invention has been made in an effort to provide a method and an apparatus for removing noise of a multichannel speech signal for improving sound quality and improving performance of a speech recognition system by obtaining a cleaner speech signal from which noise components have been removed.

According to the present invention for solving the above technical problem, a method for removing noise of a multi-channel voice signal, (a) removing the noise-specific components according to the environment from the voice signal; (b) separating speech and noise from the speech signal from which the noise intrinsic component has been removed; And (c) a post-processing step of removing residual noise remaining in the separated speech.

The noise inherent component removal in step (a) is preferably performed by frequency analysis of noise to remove noise intrinsic components using at least one of a high pass filter, a low pass filter, and a band pass filter.

In the step (b), y ₁ and y _{2 which} enhance the voice signal component and noise component included in the channel by correcting the time delays of the input signals x ₁ and x ₂ required to arrive at the respective microphones from the sound source signal, respectively. Generating a; (b2) obtaining a data matrix Y for y ₁ and y ₂ for each frame; And (b3) separating the speech and the noise by obtaining the eigen filter B (z) using the data matrix Y.

Step (b1) may include performing cross-correlation on the input signals x ₁ and x ₂ ; Obtaining a time delay of an input signal x ₁ , x ₂ required to arrive at each microphone from a sound source signal using the cross-correlated information; Generating the synchronized signals x ' ₁ and x' ₂ by shifting the input signals x ₁ and x ₂ by a time delayed value; And taking the half of the sum of the sum of x ' ₁ and the sum of x' ₂ and subtraction to obtain y ₁ and y ₂ , respectively.

Step (b3) may include obtaining an eigen filter B (z) using the data matrix Y; Constructing a polynomial using the components of the selected eigenvector, finding a root of the polynomial, moving it to the inside of a unit circle, and constructing a polynomial using the moved root to generate a filter A (z) ; And dividing the eigen filter B (z) by the filter A (z) to obtain an infinite impulse response filter H (z) (= B (z) / A (z)) to separate speech and noise. desirable.

In the step (c), it is preferable to remove the residual noise remaining in the voice signal by using the noise estimation based on one channel. The residual noise removal may include: finding a voice signal section and a noise section by applying voice activity detection (VAD); And subtracting the signal periodically updated from the noise signal in the noise section from the signal generated in step (b). The update is preferably made by weighted sum of the previous noise estimate and the current noise value.

According to an aspect of the present invention, there is provided a noise canceling apparatus for a multichannel speech signal, including: an inherent noise canceling unit for removing a noise-specific component according to an environment from a speech signal; A voice & noise separator for separating voice and noise from the voice signal from which the noise inherent components are removed; And a post-processing unit for removing residual noise remaining in the separated voice. The noise inherent component removal of the inherent noise removal unit preferably removes the noise intrinsic component using at least one of a high pass filter, a low pass filter, and a band pass filter by frequency analysis of the noise.

The voice and noise separation unit corrects the time delay of the input signals x ₁ and x ₂ to arrive at the respective microphones from the sound source signal, thereby generating y ₁ and y ₂ which enhance the voice signal components and noise components included in the channel, respectively. Time delay compensation; Y ₁ , y _{2 for} each frame A data matrix generator for obtaining a data matrix Y with respect to; And an eigen filtering unit for dividing speech and noise by obtaining an eigen filter B (z) using the data matrix Y.

The time delay compensation unit may include a cross correlation unit performing cross-correlation on the input signals x ₁ and x ₂ ; A time delay acquisition unit for obtaining a time delay of an input signal x ₁ , x ₂ required to arrive at each microphone from a sound source signal using the cross-correlated information; A synchronizer configured to shift the input signals x ₁ and x ₂ by a time delayed value to generate synchronized signals x ' ₁ and x'₂; And y ₁ & y ₂ generation units that take half of the sum of the sum of x ' ₁ and x' ₂ and subtract the values of y ₁ and y ₂ .

The eigen filtering unit comprises: an eigen filter generator for obtaining an eigen filter B (z) using the data matrix Y; A to construct a polynomial using the components of the selected eigenvector, find the root of the polynomial, move it to the inside of the unit circle, and construct a polynomial using the moved root to generate the filter A (z). (z) generating unit; And an infinite impulse response filtering unit for dividing the eigen filter B (z) by the filter A (z) to obtain an infinite impulse response filter H (z) (= B (z) / A (z)) to separate speech and noise. It is desirable to have.

The post-processing unit preferably removes residual noise remaining in the voice signal by using a noise estimation based on one channel. The residual noise is removed by applying voice activity detection (VAD) to find a voice signal section and a noise section, and subtracting the signal periodically updated from the noise signal of the noise section from the signal output from the voice & noise separation unit. This is preferred. The update is preferably made by weighted sum of the previous noise estimate and the current noise value.

A computer readable recording medium having recorded thereon a program for executing the invention described above is provided.

Hereinafter, a method and apparatus for removing noise of a multichannel voice signal according to the present invention will be described in detail with reference to the accompanying drawings. Figure 2 is a block diagram showing the configuration of an embodiment of a noise canceling device for a multi-channel voice signal according to the present invention, the inherent noise removal unit 200, voice & noise separation unit 220 and post-processing unit 240 It is made, including.

The inherent noise removal unit 200 removes noise inherent components of the environment from the voice signal. The noise inherent component may remove noise inherent components using at least one of a high pass filter, a low pass filter, and a band pass filter by frequency analysis of the noise.

The voice & noise separator 220 separates voice and noise from the voice signal from which the noise intrinsic component is removed. 3 is a block diagram illustrating a more detailed configuration of the voice and noise separation unit 220, and includes a time delay compensation unit 300, a data matrix generation unit 320, and an eigen filtering unit 340.

The time delay compensator 300 corrects the time delays of the input signals x ₁ and x ₂ to arrive at the respective microphones from the sound source signal, thereby increasing y ₁ and y ₂ respectively by reinforcing the voice signal components and noise components included in the channel. Create 4 is a block diagram illustrating a more detailed configuration of the time delay compensator 300. The cross correlator 400, the time delay acquirer 420, the synchronizer 440, and the y ₁ & y ₂ generator ( 460. The cross correlation unit 400 performs cross-correlation on the input signals x ₁ and x ₂ . The time delay acquisition unit 420 calculates a time delay of the input signals x ₁ and x _{2 that} are required to arrive at the respective microphones from the sound source signal using the cross-correlated information. The synchronizer 440 shifts the input signals x ₁ and x ₂ by a time delayed value to generate synchronized signals x ' ₁ and x' ₂ . The y ₁ & y ₂ generation unit 460 calculates y ₁ and y ₂ by taking half of the sum of the sum of x ' ₁ and x' ₂ and subtraction.

The data matrix generator 320 is y ₁ and y ₂ for each frame. Find the data matrix Y for.

The eigen filtering unit 340 obtains an eigen filter B (z) using the data matrix Y to separate speech and noise. 5 is a block diagram illustrating a more detailed configuration of the eigen filtering unit 340, and includes an eigen filter generator 500, an A (z) generator 520, and an infinite impulse response filter 540. do. The eigen filter generator 500 obtains an eigen filter B (z) using the data matrix Y. The A (z) generator 520 constructs a polynomial (ploynomial) using the components of the selected eigenvector, obtains the root of the polynomial, and moves it to the inside of the unit circle to use the polynomial. Construct filter A (z). The infinite impulse response filtering unit 540 divides the eigen filter B (z) by the filter A (z) to obtain an infinite impulse response filter H (z) (= B (z) / A (z)). Isolate the noise.

The post processor 240 removes residual noise remaining in the separated voice. The post processor 240 preferably removes the residual noise remaining in the voice signal by using a noise estimation based on one channel. The residual noise removal is performed by applying voice activity detection (VAD) to find a voice signal section and a noise section, and subtracting a signal periodically updated from the noise signal of the noise section from the signal output from the voice & noise separation unit. Can be. The update can be made by weighting the previous noise estimate and the current noise value.

FIG. 6 is a flowchart illustrating an example of a method for removing noise of a multichannel voice signal according to the present invention. Referring to FIG. 6, a method for removing noise of a multichannel voice signal according to the present invention will be described.

Removing only noise-specific components (step 600) has the following process. Different noises in different environments have their own components that can represent their respective noise characteristics. Therefore, frequency characteristic analysis is performed using signals in a section in which only a corresponding noise exists in order to characterize these components. Using the analyzed noise characteristics, noise can be removed by implementing a filter that can first remove the corresponding noise signal.

Particularly in a driving vehicle environment, noises generated in the vehicle environment are very large in the low frequency part. In general, noise generated while driving a vehicle is mostly made up of low frequency components generated from wind noise, tire noise, and engine noise. That is, the vehicle noise generally has a peak power between 100 and 800 Hz depending on the driving environment and the vehicle condition. Also below 1 kHz the noise spectral level is reduced to 6 dB / octave, while above 1 kHz the spectral level is rapidly reduced to 12 dB / octave. However, the power of the spectrum of the voice appears in a form similar to noise, so it is difficult to completely separate the speech and noise. However, most of the noise signal components are distributed, and by reducing the low frequency components, a cleaner speech signal can be obtained. In consideration of this, a high-pass filter (high-pass filter) having a cutoff frequency of 200 to 300 Hz may be applied to improve performance. Equation 1 below is an example of a simple high pass filter (cutoff frequency = 240 Hz).

Next, in the separation of speech and noise (step 620), a filter is implemented to extract only a speech signal component using two input signals, which has the following process. 7 is a flowchart illustrating a more detailed process of the separation of the voice and the noise (step 620).

This method separates two signals from the input signal of two microphones by applying signal separation technique if it is assumed that voice and noise signals are independent of each other. That is, the filter is implemented to extract only the voice signal components separately using the principle of the sound source separation technique to obtain only the voice signal from which the noise is removed.

This method divides and processes the signals from the two microphones into frames. That is, it handles the window while shifting while overlapping.

일 경우, 크로스 상관(cross-correlation)은 다음과 같이 구할 수 있다.Noise signals mixed in two channels are correlated with each other. Thus, cross-correlation can be obtained from the signals of these two channels (step 700). This information can be used to determine the delay between two microphone signals that arrive at each microphone from the sound source signal. Step 710) The signals of each of the two channels (microphone)

In this case, cross-correlation can be obtained as follows.

신호를 구한 후(720단계), 수학식 3과 같이 새로운 입력신호

신호를 재구성한다.(730단계)Next, the point having the largest value among the cross-correlation values is obtained to determine how much time is delayed between the two signals. Using the time-delayed values thus obtained, each signal is shifted by the time-delayed value.

After obtaining the signal (step 720), the new input signal as shown in Equation 3

Reconstruct the signal (step 730).

에 대해서

라는 데이터 행렬(data matrix)을 구한다.(740단계)Then each newly configured input signal

about

Obtain a data matrix (S740).

를 만족하는 값으로 정하고, k는 원하지 않는 신호의 개수이다. 또한 N은 한 프레임의 길이를 나타낸다.Where p is

Is set to a value satisfying k , and k is the number of unwanted signals. In addition, N represents the length of one frame.

)를 구한다.Using the obtained data matrix, correlation matrices (

)

 Next, the ratio matrix is obtained using the following equation.

)에 대해서 eigenvalue decomposition 기법을 이용하여 eigenvalues와 eigenvectors를 구한다. 이렇게 구한 값들을 정렬하여 가장 작은 eigenvalues와 이에 해당하는 eigenvectors를 얻는다. 작은 eigenvalues에 해당하는 것들이 음성신호에 관련된 성분들이고, 큰 eigenvalues에 해당되는 것들은 잡음(noise)에 해당되는 성분들이 된다.The Ratio matrix (

Eigenvalues and eigenvectors are obtained using the eigenvalue decomposition technique. Sort these values to get the smallest eigenvalues and the corresponding eigenvectors. The small eigenvalues are the components of the voice signal, and the large eigenvalues are the components of noise.

Using the eigenvetors thus obtained , an eigen filter called B (z) can be obtained (step 750).

In order to improve the frequency response of the eigen filter B (z) , filter A (z) is obtained (step 760), and finally, an Infinite Impulse Response (IIR) filter called H (z) is formed. 770 steps)

The method of obtaining A (z) constructs a polynomial using the elements of the selected eigenvector, finds the roots of the polynomial, and moves this value inwardly into the unit circle. Then we use the moved root to construct the polynomial, which is the coefficient of A (z) . By doing so , the frequency response of H (z) is normalized. That is, the filter is configured in the above manner to filter the noise speech signal for each frame as shown in the following equation to obtain the desired speech signal from which the noise is removed.

는 콘볼루션(convolution) 연산을 의미한다.here

Denotes a convolution operation.

In the third step 640, which is a post-processing process for removing residual noise, the following process is performed. This process obtains the signal from which noise is removed up to 2 steps, but it is a process to remove the remaining noise. This part estimates and subtracts the residual noise in a similar process to Spectral Subtraction, a 1-channel noise reduction method. At this time, the voice activity detection (VAD) method is applied to find a voice and noise section and periodically update the noise signal from the noise signals of the section corresponding to the noise. The residual noise is removed by subtracting the obtained noise signal from the signal passed in step 2.

The noise reduction method using two channels (two microphones) through the above process can be used to improve the sound quality and performance of the speech recognition system by obtaining a clearer speech signal from which noise components are removed.

Table 1 compares the SNR (Signal-to-Noise Ratio) values of the input / output voice-to-noise and the speech recognition rate results when no noise reduction is applied and when the 1 ~ 2 channel noise cancellation is applied. . The database used in this experiment is Car01 DB collected in a car noise environment of 80km high speed driving, and the voice list is composed of Navigation, car accessories, car audio, dialing command and Route guide words.

We used two channels 3 and 5 for the location of the microphone.

Experimental results of Table 1 show the results of previous papers (Sungjoo Ahn and HANSEOK KO, “Background Noise Reduction via Dual-channel Scheme for Speech Recognition in Vehicular Environment”, IEEE Transactions on Consumer Electronics, Vol. 51, No. 1, pp. 22 -27, Feb. 2005.) shows better performance. A post-processing unit (240 of FIG. 2) was added than in the previous paper, and both SNR ratio and speech recognition rate showed improved performance. In addition, even when the inherent noise canceller (200 of FIG. 2) proposed in the present invention is applied to the two-channel GSC-based noise canceling method, the performance is greatly improved.

The present invention can be embodied as code that can be read by a computer (including all devices having an information processing function) in a computer-readable recording medium. The computer-readable recording medium includes all kinds of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable recording devices include ROM, RAM, CD-ROM, magnetic tape, floppy disks, optical data storage devices, and the like.

Although the present invention has been described with reference to the embodiments shown in the drawings, this is merely exemplary, and it will be understood by those skilled in the art that various modifications and equivalent other embodiments are possible. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

According to the method and apparatus for removing noise of a multi-channel voice signal according to the present invention, by removing noise from an input terminal of a multi-channel noise cancellation system, as a result, an improved performance can be obtained than a noise processing method in a single channel environment. It can improve performance.

In addition, it is possible to prevent the performance degradation of the system by removing the noise components that exist in each noise environment. Also, it is suitable to be applied to the sound quality improvement and speech recognition system by showing the superior performance compared to other noise processing methods in the two channel environment. Do. In addition, the voice preprocessing technology such as the present invention provides a foundation for applying a voice recognition and interface technology, which was made only in an office environment, to a noisy environment.

Claims (17)

(a) removing noise inherent components of the environment from the speech signal; (b) separating speech and noise from the speech signal from which the noise intrinsic component has been removed; And and (c) a post-processing step of removing residual noise remaining in the separated speech. The method of claim 1, wherein the noise inherent component of step (a) is A noise reduction method of a multi-channel speech signal characterized by removing noise intrinsic components using at least one of a high pass filter, a low pass filter, and a band pass filter by frequency analysis of the noise. The method of claim 1, wherein step (b) (b1) generating y ₁ and y ₂ by correcting the time delays of the input signals x ₁ and x ₂ to arrive at the respective microphones from the sound source signal by respectively enhancing the voice signal components and the noise components included in the channel; (b2) y ₁ and y _{2 for} each frame Obtaining a data matrix Y for; And and (b3) separating the speech and the noise by obtaining an eigen filter B (z) using the data matrix Y. The method of claim 3, wherein step (b3) Obtaining an eigen filter B (z) using the data matrix Y; Constructing a polynomial using the components of the selected eigenvector, finding a root of the polynomial, moving it to the inside of a unit circle, and constructing a polynomial using the moved root to generate a filter A (z) ; And And dividing the eigen filter B (z) by the filter A (z) to obtain an infinite impulse response filter H (z) (= B (z) / A (z)) to separate speech and noise. Noise reduction method for a multichannel audio signal. The method of claim 3, wherein step (b1) Performing cross-correlation on the input signals x ₁ , x ₂ ; Obtaining a time delay of an input signal x ₁ , x ₂ required to arrive at each microphone from a sound source signal using the cross-correlated information; Generating the synchronized signals x ' ₁ and x' ₂ by shifting the input signals x ₁ and x ₂ by a time delayed value; And And obtaining y ₁ and y ₂ , respectively, by taking half of the sum of the sum of x ' ₁ and x' ₂ and subtraction. The method of claim 1, wherein step (c) A noise canceling method for a multichannel speech signal, characterized in that the residual noise remaining in the speech signal is removed using a 1-channel noise estimation. The method of claim 6, wherein the residual noise cancellation Finding a voice signal section and a noise section by applying voice activity detection (VAD); And And subtracting the signal which is periodically updated with the noise signal in the noise section from the signal generated in the step (b). 8. The method of claim 7, wherein said update is A noise canceling method of a multichannel speech signal, characterized by a weighted sum of a previous noise estimate and a current noise value. An inherent noise removing unit for removing a noise inherent component according to the environment from the voice signal; A voice & noise separator for separating voice and noise from the voice signal from which the noise inherent components are removed; And And a post-processing unit for removing residual noise remaining in the separated voice. 10. The method of claim 9, wherein the noise inherent component removal of the inherent noise removal unit A noise canceling device for a multichannel speech signal, characterized in that the noise is removed by frequency analysis using at least one of a high pass filter, a low pass filter, and a band pass filter. The method of claim 9, wherein the voice and noise separation unit A time delay compensator for correcting the time delays of the input signals x ₁ and x ₂ to arrive at the respective microphones from the sound source signal, thereby generating y ₁ and y ₂ , each of which enhances the voice signal component and the noise component included in the channel; Y ₁ , y _{2 for} each frame A data matrix generator for obtaining a data matrix Y with respect to; And And an eigen filtering unit for dividing speech and noise by obtaining an eigen filter B (z) using the data matrix Y. The method of claim 11, wherein the time delay compensation unit A cross correlator for performing cross-correlation on the input signals x ₁ and x ₂ ; A time delay acquisition unit for obtaining a time delay of an input signal x ₁ , x ₂ required to arrive at each microphone from a sound source signal using the cross-correlated information; A synchronizer configured to shift the input signals x ₁ and x ₂ by a time delayed value to generate synchronized signals x ' ₁ and x'₂; And Wherein x _'1, x' ₂ obtained by adding the value obtained by subtracting half the taking y _1, multi-channel, characterized in that it comprises & y ₁ y ₂ y ₂ generation unit to obtain the removal of the audio signal noise values each device. The method of claim 11, wherein the eigen filtering unit An eigen filter generator for obtaining an eigen filter B (z) using the data matrix Y; A to construct a polynomial using the components of the selected eigenvector, find the root of the polynomial, move it to the inside of the unit circle, and construct a polynomial using the moved root to generate the filter A (z). (z) generating unit; And Infinite impulse response filtering unit for dividing the eigen filter B (z) by the filter A (z) to obtain an infinite impulse response filter H (z) (= B (z) / A (z)) to separate speech and noise Noise canceling device for a multi-channel voice signal characterized in that. The method of claim 9, wherein the post-processing unit A noise canceling device for a multichannel speech signal, characterized in that the residual noise remaining in the speech signal is removed using a 1-channel noise estimation. 15. The method of claim 14, wherein the residual noise cancellation is It is made by applying the voice activity detection (VAD) to find the voice signal section and the noise section, and by subtracting the signal periodically updated the noise signal of the noise section from the signal output from the voice & noise separation unit Noise canceling device for channel voice signal. The method of claim 15 wherein the update is A noise canceling device for a multichannel speech signal, characterized by a weighted sum of a previous noise estimate and a current noise value. A computer-readable recording medium having recorded thereon a program for executing the invention according to any one of claims 1 to 8.

Images