KR101966175B1 - Apparatus and method for removing noise - Google Patents

Abstract

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise canceling apparatus and method, and more particularly, to a noise canceling apparatus and method for speech recognition.
An apparatus for eliminating noise in an input signal, the apparatus comprising: a target signal extractor for extracting a first target signal from the input signal using a first separation vector; A target signal removing section for extracting a target signal, a detection section for extracting audio section information of the first target signal, and a second section for calculating a weight from the first noise signal using the speech section information, And a first noise canceller for removing noise from the signal.

Description

[0001] APPARATUS AND METHOD FOR REMOVING NOISE [0002]

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

When a plurality of noise signals are mixed and input to a voice signal, a device for inputting a voice signal for voice recognition or the like performs a voice recognition function by extracting a desired voice signal from the mixed signal.

For example, when controlling the operation of IPTV through voice recognition, in a noisy environment where various kinds of noise sources exist such as music sound or IPTV sound itself, noise is separated from user's voice, The quality of speech recognition performance similar to that of a noise-free environment can be guaranteed.

Accordingly, a beamforming technique is used as a signal separation technique. A typical beam-forming technique may use a method of enhancing a signal input in a specific direction from an input mixed signal over a signal input in another direction and a method of removing only an enhanced specific direction signal. Then, the signal separation function can be performed by removing the noise of the specific direction signal by using the signal from which the specific direction signal is removed.

However, according to the conventional beam forming technique, an appropriate signal separation operation can be performed only when a voice signal generated from a fixed sound source is input. Accordingly, when the sound source moves, a sound source tracking device or a sound source tracking method is additionally required to obtain position information of the sound source.

Embodiments of the present invention provide a noise canceling apparatus that effectively removes noise even when a sound source moves.

An apparatus for eliminating noise of an input signal according to an embodiment of the present invention includes a target signal extracting unit for extracting a first target signal from the input signal using a first separation vector, A target signal removing section for extracting a first noise signal from the first noise signal, a detection section for extracting speech section information of the first target signal, and a speech section for calculating a weight from the first noise signal using the speech section information, And a first noise canceller for removing noise from the first target signal.

A noise reduction method according to an embodiment of the present invention includes extracting a first target signal from the input signal using a first separation vector, extracting a first noise signal from the input signal using a second separation vector Calculating a weight from the first noise signal using the speech interval information, and removing noise from the first target signal using the weight value, .

The noise canceller according to the embodiment of the present invention effectively removes noise without additional sound source tracking device when the sound source moves.

1 is a block diagram of a noise canceling apparatus according to an embodiment of the present invention.
2 is a flow chart of a noise removal method in accordance with an embodiment of the present invention.
FIGS. 3 to 5 are views for explaining the effect of the noise removing apparatus according to an embodiment of the present invention.

Hereinafter, a noise canceling apparatus according to the present invention will be described in detail with reference to the drawings. The suffix " module " and " part " for the components used in the following description are given or mixed in consideration of ease of specification, and do not have their own meaning or role.

Hereinafter, the structure of a noise canceling apparatus according to an embodiment of the present invention will be described with reference to FIG.

1 is a block diagram of a noise canceling apparatus according to an embodiment of the present invention.

The noise eliminating apparatus according to the embodiment of the present invention includes a plurality of microphones mic1, mic2, mic3 and mic4, a target signal extracting unit 101, a target signal removing unit 103, a detecting unit 105, Rejection 107. [0033] FIG. In addition, a cleaner speech signal including the second noise canceller 109 can be extracted. Then, the speech recognition unit 111 can recognize the speech signal extracted from the first noise remover 107 or the second noise remover 109.

A plurality of microphones (mic1, mic2, mic3, mic4) receive mixed signals (not shown). The mixed signal may include a signal generated from a plurality of sound sources. The mixed signal may include all signals that may occur in daily life, such as a voice signal for controlling IPTV, and a sound generated when using an electronic device.

The device acoustic de-rejection (not shown) removes the device sound from the mixed signal. For example, when a microphone of an IPTV receives a mixed signal, the mixed signal may include sound generated by the IPTV itself. By removing the sound generated from the IPTV itself from the mixed signal by the acoustic deactivation (not shown), it is possible to remove the noise having the largest acoustic size among the mixed signals, thereby improving the performance of the speech recognition. As a result, the input signals X1, X2, X3, and X4 may include signals from which the device sound is removed from the mixed signal.

The target signal extracting unit 101 receives a plurality of input signals X1, X2, X3, and X4 and outputs a first target signal U. At this time, the target signal extracting unit 101 receives a plurality of input signals X1, X2, X3, and X4 and selects a first separation vector w1 for separating one audio signal to be subjected to speech recognition among them And extract the first target signal U using the first separation vector w1. The target signal extracting unit 101 may determine the first separation vector w1 using an Independent Vector Analysis (IVA) algorithm. The independent vector analysis algorithm is an extension of the independent component analysis to the frequency domain.

Equation 1 is a general model representation of the independent component analysis problem.

x is an input vector, A is a mixing matrix, s is a sound source vector, y is an output vector, and W is a pseudo-inverse matrix of A. Learning the pseudo inverse matrix (W) to extract the signal generated from a specific sound source is an independent component analysis problem. In the pseudoinverse (W) learning, a cost function can be used. For example, as the cost function is minimized, an output vector y that is as close as possible to the sound source vector s can be extracted.

Since the independent vector analysis algorithm is an independent component analysis algorithm extended to the frequency domain, the independent vector analysis problem is to determine the first separation vector w1.

The target signal extracting unit 101 may perform a whitening operation on the input vector x as a pre-processing for determining the first separation vector w1. By performing the whitening operation, the correlation of the plurality of input signals X1, X2, X3, and X4 can be eliminated. Equation (5) represents the whitening operation process of the input vector (x).

을 의미하고,

는 E 행렬의 전치 행렬을 의미한다.In the covariance matrix of the plurality of input signals X1, X2, X3 and X4, the diagonal matrix of the eigenvalues of the covariance matrix is D and the vowel matrix of the eigenvectors of the covariance matrix is E At this time,

Lt; / RTI >

Denotes the transpose matrix of E matrix.

Equation (11) shows the cost function (J _g ) using the whitened input vector (z) of Equation (5).

k is a frequency index, E is an expectation, g is a nonlinear function, for example, a nonlinear function that can be approximated by g (y) = y ⁴ when based on kurtosis, H is a Hermitian matrix Matrix. Equation (11) shows a cost function according to vector correlation and vector mutual information in independent vector analysis.

Equation (12) is a result of differentiating the cost function of Equation (11), and means a learning rule included in the maximum diagonal algorithm of the first separation vector (w1).

* Denotes a conjugate matrix.

Equation (13) represents the fast independent vector analysis learning rule included in the Newton algorithm as a result of using the first-order and second-order differentiations of Equation (12).

(w ₁ ^k ) ^H z ^k means the first target signal (U). The target signal extracting unit 101 can determine the first separation vector w1 from Equation 13 and calculate the first separation vector w1 from the plurality of input signals X1, X2, X3, X4 using the first separation vector w1 The target signal U can be extracted.

The target signal removing unit 103 receives a plurality of input signals X1, X2, X3 and X4 and outputs a plurality of first noise signals B1, B2 and B3. In other words, the target signal removing unit 103 receives the plurality of input signals X1, X2, X3, and X4 and separates the remaining sound signals excluding one of the sound signals, as opposed to the target signal extracting unit 101 B2, B3 using the second separation vector w2, and the second separation vector w2 using the second separation vector w2. The target signal removing unit 103 may use an independent vector analysis algorithm used by the target signal extracting unit 101 to remove one output signal. However, the target signal removing unit 103 uses a learning rule having a sign opposite to the learning rule expressed in Equation (12) since a plurality of input signals must be separated and one output signal must be removed. Therefore, the learning rule included in the maximum-radius algorithm used by the target signal remover 103 is expressed by Equation (14).

z denotes the whitening operation result of the input vector x, and (w ₂ ^k ) ^H z ^k denotes the first noise signals B1, B2, and B3. The target signal remover 103 can determine the second separation vector w2 using Equation (14). The target signal removing unit 103 may then extract the first noise signals B1, B2, and B3 from the plurality of input signals X1, X2, X3, and X4 using the second separation vector w2 .

The target signal removing unit 103 may be configured to select two input signals out of a plurality of input signals to exclude a target signal from a plurality of noises and to exclude a target signal from two selected input signals, A method of excluding the target signal directly from the whole can be used.

As described above, according to the embodiment of the present invention, the target signal extracting unit 101 and the target signal removing unit 103 can use the independent vector analysis algorithm to effectively recognize speech even when the sound source moves.

The detection unit 105 outputs probability information P in which a speech signal exists for each frame of the first target signal U. [ The first target signal U output from the target signal extracting unit 101 includes not only a single voice signal but also other noise. The detection unit 105 detects presence or absence of a voice signal in each frame of the first target signal U and outputs the detection result to the probability information P, . The detection unit 105 may use a Minima Controlled Recursive Averaging (MCRA) algorithm to output the voice presence probability information P of the first target signal U.

First, the first target signal U may be displayed in the frequency domain through a short-time Fourier transform. In the first target signal U (k, l) in the frequency domain, k denotes a frequency index and l denotes a frame index. The frame index (1) represents the end sections of the signal when dividing the entire signal into a plurality of short sections. At this time, the short-term Fourier transform may be executed in the target signal extracting unit 101 and the target signal removing unit 103. [

(23) is an expression that represents an operation for extracting the energy spectrum P (k, l) of the first target signal U (k, l) in the frequency domain.

는 0보다 크고 1보다 작은 값을 가지는 스무딩 매개 변수를 의미한다. 검출부(105)는 수학식 23을 이용하여 주파수 영역의 제1 목표 신호(U(k,l))의 스무딩 연산 결과인 에너지 스펙트럼(P(k,l))을 출력할 수 있다. 수학식 23은 과거의 에너지 스펙트럼(P(k,l-1))을 입력받아 현재의 에너지 스펙트럼(P(k,l))을 출력하는 재귀 평균 연산을 나타낸다.

Means a smoothing parameter having a value greater than zero and less than one. The detecting unit 105 can output the energy spectrum P (k, l) which is the smoothing operation result of the first target signal U (k, l) in the frequency domain using the equation (23). (23) represents a recursive averaging operation that receives the past energy spectrum P (k, l-1) and outputs the current energy spectrum P (k, l).

(24) is an equation representing an operation for determining the minimum value of the energy spectrum for each frequency using the energy spectrum P (k, l) of the equation (23).

P _tmp (k, l) denotes a temporary value of the energy spectrum for determining the minimum value of the energy spectrum (P _min (k, l)). The detection unit 105 uses the result of comparing the current energy spectrum component P (k, l) and the minimum value P _min (k, l-1) of the past energy spectrum components to obtain the energy spectrum at each frequency The minimum value _Pmin (k, l) can be determined.

(25) expresses an operation for updating the minimum value (P _min (k, l)) and the temporary value (P _tmp (k, l)) of the current energy spectrum using the results of Equations (23) and to be.

The detection unit 105 can update the minimum value P _min (k, l) and the temporary value P _tmp (k, l) of the current energy spectrum using the calculation of Equation (25).

Equation 26 is the ratio (S _r (k, l) of using the result of Equation 25, the minimum value of the energy spectrum _{(P min (k, l)} ) and the current energy spectrum components (P (k, l)) .

The detection unit 105 calculates the ratio S _r (k, l) of the energy spectrum component P (k, l) and the minimum value P _min (k, l) of the energy spectrum at each frequency using the calculation of Equation , l)) can be obtained.

)을 연산하는 식이다.(27) using Equation 26 and the speech presence probability index I (k, l)

.

)을 출력할 수 있다. 이때, 음성 존재 확률 인덱스(I(k,l))는 수학식 26의 결과인 S_r(k,l)이 특정 문턱 값(δ)보다 작은 경우 0이 되고, S_r(k,l)이 특정 문턱 값(δ)보다 큰 경우 1이 된다. 이와 같이, 검출부(105)는 수학식 23 내지 수학식 27에 표현된 연산을 포함하는 최소값 제어 재귀 평균 알고리즘을 이용하여, 제1 목표 신호(U)의 각 프레임이 음성이 존재하는 구간인지 또는 음성이 존재하지 않는 구간인지 여부를 출력할 수 있다.The detection unit 105 uses the calculation of Equation (27) to calculate the voice presence probability

Can be output. In this case, the speech presence probability index (I (k, l)) is a small case 0 than S _r (k, l) is a specific threshold value (δ) that is the result of Equation 26, S _r (k, l) the And becomes 1 when it is larger than the specific threshold value?. In this manner, the detection unit 105 can determine whether each frame of the first target signal U is in a period in which the speech exists or in a state in which the speech is not speech, using the minimum value control recursive averaging algorithm including the operations expressed by the equations (23) It is possible to output whether or not the section is not present.

The first noise removing unit 107 extracts the second target signal Y which is a result of additionally removing the residual noise signal of the first target signal U using the first noise signals B1, B2, and B3. At this time, as the second target signal Y included in the first target signal U is output without distortion, and the residual noise signal is minimized and output, more effective speech recognition can be performed. On the other hand, there are residual audio signals that can not be removed in the first noise signals B1, B2, and B3. At this time, as the residual voice signal included in the first noise signals B1, B2, and B3 is minimized, the second target signal Y can be output without distortion. The first noise removing unit 107 may use a Wiener filter to remove the residual noise signal of the first target signal U. [

First, the first noise removing unit 107 multiplies the vector (b) of the first noise signals B1, B2, and B3 by the interval of the residual voice signal and the residual voice signal It is possible to define them separately.

b is a first noise signal vector, s is a residual speech signal vector included in the first noise signal, and v is a second noise signal vector included in the first noise signal.

On the other hand, the first noise removing unit 107 may extract the second target signal y through the calculation shown in Equation (29). Equation 29 shows an equation for calculating the second target signal y using the first target signal u, the first noise signal b, and the weight parameter w of the Wiener filter.

u refers to the first target signal, b is the first noise signal vector, w is a weighting parameter of the Wiener filter, y is the second target signal, the cost function J _k. The cost function (J _k) may include a coefficient of μ in order to minimize the residual speech signal vector that contains the first noise signal (b).

On the other hand, the second target signal y can be optimized only when the cost function J _k is minimized. In other words, when the cost function J _k is the minimum, the portion in which no speech exists in the second target signal y can be minimized. The weight parameter (w) of the Wiener filter for minimizing the cost function (J _k ) is shown in Equation (30).

At this time, the result of Equation (31) can be substituted into Equation (30) for calculating the weight parameter (w) of the Wiener filter.

Equation (31) represents the covariance of the first noise signal in the interval in which the residual speech signal exists and the covariance of the first noise signal in the interval in which the residual speech signal does not exist as a result of using the definition of Equation (28). In this manner, by applying the covariance calculation result depending on the presence or absence of the residual speech signal to the weight parameter (w) calculation of the Wiener filter, the energy of the first target signal can be minimized in the speech non-existence period. In other words, the energy of the first target signal is minimized in the non-speech region, so that the residual noise signal of the first target signal can be minimized.

)를 출력한다. 제2 잡음 제거부(109)는 제3 목표 신호(

)를 출력하기 위하여 제2 목표 신호(Y)에 최소값 제어 재귀 평균 알고리즘을 적용하여 음성 존재 확률을 출력할 수 있다. 또한, 제2 잡음 제거부(109)는 제2 목표 신호(Y)에 최소 평균 제곱 오차(Minimum-mean square error) 알고리즘을 적용하여 제2 목표 신호(Y)의 잡음을 제거할 수 있다. 또한, 제2 잡음 제거부(109)는 제2 목표 신호(Y)에 끝점 검출(end-point detection) 알고리즘을 적용하여 제2 목표 신호(Y)에 포함된 음성 신호 즉, 제3 목표 신호(

)의 시작점 및 종료점을 검출할 수 있다.After receiving the second target signal Y, the second noise removing unit 109 further removes the noise remaining in the second target signal Y to generate a third target signal

). The second noise removing unit 109 receives the third target signal (

The second target signal Y may be output with a voice presence probability by applying a minimum value control recursive averaging algorithm to the second target signal Y. [ The second noise removing unit 109 may remove the noise of the second target signal Y by applying a minimum-mean square error algorithm to the second target signal Y. [ The second noise removing unit 109 applies an end-point detection algorithm to the second target signal Y to generate a speech signal included in the second target signal Y, that is, a third target signal

Can be detected.

)를 수신하여, 사용자가 IPTV 등을 제어하기 위하여 전송한 음성 신호를 인식할 수 있다. 예컨대, 음성 인식부(111)는 제3 목표 신호(

)에 HMM(Hidden Markov Model) 알고리즘을 적용함으로써, 복수의 마이크로폰(mic1,mic2,mic3,mic4)이 수신한 혼합 신호 중 IPTV 제어를 위한 음성 신호를 인식할 수 있다.
The speech recognition unit 111 receives the third target signal (the second target signal) output from the second noise removing unit 109

), And can recognize a voice signal transmitted by a user to control IPTV or the like. For example, the speech recognition unit 111 receives the third target signal (

), It is possible to recognize a voice signal for IPTV control among mixed signals received by a plurality of microphones (mic1, mic2, mic3, mic4) by applying an HMM (Hidden Markov Model)

Next, a noise removal method according to an embodiment of the present invention will be described with reference to FIG.

2 is a flow chart of a noise removal method in accordance with an embodiment of the present invention.

The plurality of microphones (mic1, mic2, mic3, mic4) receives the mixed signal (S201). For IPTV operation control using voice recognition, IPTV includes a microphone for receiving voice signals. IPTV may include at least one microphone for effective speech recognition. The mixed signal can include all kinds of noise signals that occur in everyday life as well as voice signals for IPTV control.

The device acoustic de-rejection (not shown) removes the device sound in the mixed signal to generate the input signals X1, X2, X3, and X4 (S203). For example, in the case of voice recognition for IPTV operation control, device sound refers to sound that occurs during IPTV operation.

The target signal extracting unit 101 receives a plurality of input signals X1, X2, X3 and X4 and performs an operation of whitening the input signals X1, X2, X3 and X4 (S205). Referring to Equation (5), the target signal extracting unit 101 removes the correlation between the plurality of input signals X1, X2, X3, and X4 through the whitening operation, Can be separated from the plurality of input signals (X1, X2, X3, X4). The first target signal U may comprise, for example, a voice signal that should be recognized for IPTV operation control.

Subsequently, the target signal extracting unit 101 determines a first separation vector w1 from the whitened input signal using the independent vector analysis learning rule (S207). The independent vector analysis learning rule is an extension of the independent component analysis problem to the frequency domain. The target signal extracting unit 101 extracts a cost function according to the vector correlation and vector mutual information of Equation (11), a maximum likelihood algorithm , The Newton algorithm of Equation (13) can be used to determine the first separation vector w1.

Subsequently, the target signal extracting unit 101 extracts the first target signal U from the input signals X1, X2, X3, and X4 using the first separation vector w1 (S209).

The target signal removing unit 103 receives a plurality of input signals X1, X2, X3 and X4 and performs an operation of whitening the input signals X1, X2, X3 and X4 (S205).

Next, the target signal remover 103 determines a second separation vector w2 from the whitened input signal using the independent vector analysis learning rule (S211). Since the independent vector analysis learning rule is the same as that described in step S207, it is omitted here.

Subsequently, the target signal removing unit 103 extracts the first noise signals B1, B2, and B3 from the input signals X1, X2, X3, and X4 using the second separation vector w2 (S213). The target signal removing unit 103 can remove the target signal by extracting the remaining noise except for the voice signal that should be recognized for the IPTV operation control, for example, as opposed to the target signal extracting unit 101. [ Therefore, the target signal removing unit 103 uses the equations (11) and (14) and (13), which are opposite in sign only to the maximum permutation algorithm of Equation (12) used in the target signal extracting unit (101) (w2).

) 연산을 이용하여, 제1 목표 신호(U)의 음성 존재 확률 정보(P)을 출력할 수 있다.The detection unit 105 outputs the voice existence probability information P of the first target signal U from the first target signal U using the minimum value control recursive averaging algorithm (S215). The detection unit 105 calculates an energy spectrum (P (k, l)) extraction operation of Equation 23, a minimum energy spectrum Pmin (k, l) determination operation of Equation 24, an energy spectrum update operation of Equation (K, l) for the current energy spectrum component P (k, l) of Equation 26, the ratio of the minimum energy spectrum Pmin

) Operation, the voice presence probability information P of the first target signal U can be output.

The first noise canceller 107 divides the first noise signal vector b into the presence or absence of a residual voice signal as shown in Equation 28, The covariance of the first noise signal vector b is calculated in step S217, and the covariance of the first noise signal vector b is calculated in the absence of the remaining speech signal in step S219.

The first noise eliminator 107 determines a weight parameter w that minimizes the energy of the first target signal u in the speech absence period using the covariance calculation result of the first noise signal vector b S221). The first noise removing unit 107 may use Equation (30) to calculate the weight parameter w.

The first noise removing unit 107 generates the second target signal y as a result of removing the signal obtained by calculating the weight parameter w from the first target signal u to the first noise signal b ). The first noise remover 107 may use Equation 29 to generate the second target signal y. At this time, if the signal obtained by calculating the weight parameter w from the first target signal u is removed, the distortion of the audio signal included in the first target signal u is minimized, 1 target signal u can be minimized, it is possible to transmit a cleaner voice signal to the voice recognition unit 111

)를 생성한다(S225).For reference, a second noise canceller 109 may be added to the noise canceller according to an embodiment of the present invention to additionally remove the noise included in the second target signal y. The second noise eliminator 109 subtracts the noise of the second target signal y from the third target signal y using the minimum value control recursive averaging algorithm and the minimum mean square error algorithm

(S225).

)를 수신한 후, 음성 신호를 인식한다(S227).
The speech recognition unit 111 receives the second target signal y or the third target signal y

(Step S227).

Next, a structure of a noise removing apparatus according to an embodiment of the present invention will be described with reference to FIG.

FIGS. 3 to 5 are views for explaining the effect of the noise removing apparatus according to an embodiment of the present invention.

Referring to FIG. 3, in order to explain the effect of the noise canceling apparatus according to an embodiment of the present invention, a case where a plurality of microphones 301 receive various sound signals is set as a simulation environment. In the simulation environment, two IPTV speakers 303 are placed in the 180 占 direction at positions 0.3 m away from the plurality of microphones 301 of the IPTV, (S). Various types of noise sources N are placed in a semicircle at intervals of 30 degrees at a distance of 1.5 m from the plurality of microphones 301. Various types of noise sources N may include, for example, automobile engine sounds, music sounds, voice databases, e.g., male and female voice signals included in the TIMIT database.

The noise removal apparatus according to an embodiment of the present invention separates and extracts a speech signal from a mixed signal received by a plurality of microphones 301 in the above simulation environment, and recognizes the speech signal. At this time, the device acoustics removal unit removes the sound generated in the IPTV speaker 303 using an independent component analysis algorithm, and the voice recognition unit 111 recognizes the voice using the HMM algorithm.

Referring to FIGS. 4 and 5, as a result of applying the simulation, when the signal-to-noise ratio (SNR) between the audio signal of the user S and the audio of the device generated by the IPTV speaker 303 is 5 dB, S) and the word recognition rate according to the noise-to-noise ratio (NOISE-SNR) of noise generated in various kinds of noise sources (N). In this case, when the voice signal-to-noise ratio (NOISE-SNR) is 0 (zero), it means that the size of the voice signal is equal to the size of the noise. When the voice signal to noise ratio (NOISE-SNR) The size is larger than the noise size.

)는 제2 잡음 제거부(109)가 출력한 신호를 의미한다.A microphone (mic) input is a mixed signal received by a plurality of microphones (mic1, mic2, mic3, mic4), a device acoustic elimination output is a signal in which a device sound is removed from a mixed signal, a first target signal The second target signal Y is a signal output from the noise removing unit 107 and the third target signal Y

) Denotes a signal output from the second noise removing unit 109.

As a result of the simulation, comparing the output of each stage of the noise reduction method according to an embodiment of the present invention, the noise reduction method according to an embodiment of the present invention is effective for speech recognition. In particular, when the speech recognition environment is bad, Even when the NOISE-SNR is zero, the speech recognition rate is gradually improved by applying the noise reduction method according to an embodiment of the present invention

According to an embodiment of the present invention, the above-described method can be implemented as a code that can be read by a processor on a medium on which the program is recorded.

The above-described noise canceller may be applied to not only the configuration and the method of the embodiments described above but also all or some of the embodiments may be selectively combined so that various modifications may be made. .

Claims (10)

1. An apparatus for eliminating noise in an input signal,
A target signal extractor for extracting a first target signal from the input signal using a first separation vector;
A target signal removing unit for extracting a first noise signal from the input signal using a second separation vector;
A detector for extracting voice interval information of the first target signal; And
And a first noise canceller for calculating a weight from the first noise signal using the speech interval information and removing noise from the first target signal using the weight,
The detection unit
Extracting the speech interval information using a minimum value control recursive averaging algorithm
Noise canceling device. 1. An apparatus for eliminating noise in an input signal,
A target signal extractor for extracting a first target signal from the input signal using a first separation vector;
A target signal removing unit for extracting a first noise signal from the input signal using a second separation vector;
A detector for extracting voice interval information of the first target signal; And
And a first noise canceller for calculating a weight from the first noise signal using the speech interval information and removing noise from the first target signal using the weight,
The first noise removing unit
The first noise signal is covaried using the speech interval information, and the weight value is calculated using the result of the covariance calculation of the first noise signal
Noise canceling device. 3. The method according to claim 1 or 2,
The target signal removing unit
Determining the second separation vector from the input signal using an independent vector analysis algorithm
Noise canceling device. 3. The method according to claim 1 or 2,
The target signal extracting unit
Determining the first separation vector from the input signal using an independent vector analysis algorithm
Noise canceling device. delete A method for eliminating noise in an input signal,
Extracting a first target signal from the input signal using a first separation vector;
Extracting a first noise signal from the input signal using a second separation vector;
Extracting voice interval information of the first target signal;
Calculating a weight from the first noise signal using the speech interval information; And
And removing noise from the first target signal using the weight,
The step of extracting the voice section information
And extracting the speech interval information using a minimum value control recursive averaging algorithm
Noise canceling method. A method for eliminating noise in an input signal,
Extracting a first target signal from the input signal using a first separation vector;
Extracting a first noise signal from the input signal using a second separation vector;
Extracting voice interval information of the first target signal;
Calculating a weight from the first noise signal using the speech interval information; And
And removing noise from the first target signal using the weight,
The step of computing the weights
Covariance-calculating the first noise signal using the speech interval information; And
And calculating the weight using the result of the covariance calculation
Noise canceling method. 8. The method according to claim 6 or 7,
The step of extracting the first noise signal
And determining the second separation vector from the input signal using an independent vector analysis algorithm
Noise canceling method. 8. The method according to claim 6 or 7,
The step of extracting the first target signal
Determining the first separation vector from the input signal using an independent vector analysis algorithm
Noise canceling method. delete

Images