KR19990001296A - Adaptive Noise Canceling Device and Method - Google Patents

Abstract

The present invention relates to a noise canceling device for a voice signal, and in particular, a first microphone (10) for receiving voice and background noise so as to improve the sound quality of a noise voice using two microphones; A second microphone 20 for receiving background noise and voice; An adaptive noise estimator (30) for obtaining a relationship between two input signals by using the signal of the second microphone (20) as an input and the signal of the first microphone (10) as a reference signal; The signal of the first microphone 10, the result of the adaptive noise estimator 30, and an adjustment constant are input, and the result of the noise estimator of the previous M / 2 time input from the delay unit 60 is used as a reference signal. An adaptive residual noise estimator 40 for storing the relationship between these signals in an adaptive filter coefficient; An adaptive residual noise attenuation unit 50 for estimating a speech signal attenuated by the residual noise using the filter coefficients of the adaptive residual noise estimator 40 as the input of the result of the adaptive noise estimator 30 is input. The present invention relates to an apparatus and a method for adaptive noise reduction.

Description

Adaptive Noise Canceling Device and Method

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise canceling device for speech signals, and more particularly, to an adaptive noise canceling device and method for improving the sound quality of noise speech using two microphones.

Most noise reduction devices using adaptive filters use two microphones so that one microphone receives only background noise and the other receives both voice and background noise.

The conventional noise canceling apparatus includes a voice detector (3) for detecting a voice section input through the first microphone (1) and the second microphone (2), as shown in FIG. A background noise remover (4) for estimating and removing background noise from the voice detected by the voice detector (3); It consists of an adaptive filter update unit 5 for filtering the noise detected by the voice detector 3.

The conventional noise canceling apparatus configured as described above starts with searching for a section in which only noise exists.

In this noise section, a relation between the input signals input to the two microphones 1 and 2 is obtained by using an adaptive filtering method.

The relationship between the input signals input to the two microphones 1 and 2 can be written as in Equation 1.

[Equation 1]

x1 [n] = x2 [n] * h [n]

Here, x1 [n] is an input signal of the first microphone 1, x2 [n] is an input signal of the second microphone 2, and h [n] is a signal of the first microphone 1 and a second signal. The impulse characteristic between the signals of the microphone 2 is shown.

This method has three problems.

First is the problem of noise section detection.

Since the consonants and noise of speech are difficult to distinguish, and the time-varying noise cannot be distinguished from the speech section including vowels, the detection success rate of the noise section cannot be guaranteed.

Second: It is assumed that only the noise is input to the second microphone 2, but it is difficult to achieve in actual situations.

Therefore, Equation 1 should be modified as follows.

In Expressions 2 and 3 below, it can be seen that voice signals are mixed in two microphones 1 and 2 input signals.

[Equation 2]

x1 [n] = s [n] + n1 [n]

[Equation 3]

x2 [n] = s [n] * g [n] + n2 [n]

Here, n1 [n] is a noise signal input to the first microphone 1, and g [n] means some relation to each voice input to the first microphone 1 and the second microphone 2. .

That is, even for the same sound source, the voice signals input to the two microphones vary according to the position of the microphone and the surrounding environment.

n1 [n] and n2 [n] refer to noise signals input to the first microphone 1 and the second microphone 2, respectively.

Therefore, even if the noise component contained in the first microphone 1 is removed by calculating the relationship between the noise of the two microphone signals as in the conventional method, if the voice signal component is contained in the second microphone 2 as it is now, Since the relationship between is not known, distorting the speech signal component contained in the first microphone 1 results in a reduction in the noise canceling performance.

Third: Since the coefficient of the adaptive filter is updated only in the noise section, an error caused by the difference between the past adaptive filter coefficient and the current adaptive filter coefficient is unavoidable in the speech section to which the adaptive filter is actually applied.

SUMMARY OF THE INVENTION The present invention has been made to solve the problems of the related art as described above, and an object of the present invention is to provide an adaptive noise canceling apparatus and method, which can improve the sound quality of a noise voice using two microphones. .

1 is a block diagram of a conventional adaptive noise canceller;

2 is a block diagram of an adaptive noise canceling apparatus according to the present invention;

3 is a detailed configuration diagram of an adaptive noise estimator of FIG. 2;

4 is a detailed configuration diagram of an adaptive residual noise estimation unit of FIG. 2;

5 is a detailed configuration diagram of the adaptive residual noise attenuation unit of FIG. 2.

*** Explanation of symbols for the main parts of the drawing ***

10,20: Microphone 30: Adaptive noise estimation unit

40: adaptive residual noise attenuation unit 50: adaptive residual noise attenuation unit

60: delay unit

The present invention for achieving the above object, the first microphone 10 for receiving a voice and background noise; A second microphone 20 for receiving background noise and voice; An adaptive noise estimator (30) for obtaining a relationship between two input signals by using the signal of the second microphone (20) as an input and the signal of the first microphone (10) as a reference signal; The signal of the first microphone 10, the result of the adaptive adaptation estimator 30, and the adjustment constant are input, and the result of the noise estimator of the previous M / 2 time input from the delay unit 60 is used as a reference signal. An adaptive residual noise estimator 40 for storing the relationship between these signals in an adaptive filter coefficient; An adaptive residual noise attenuation unit 50 for estimating a speech signal attenuated by the residual noise using the filter coefficients of the adaptive residual noise estimator 40 as the input of the result of the adaptive noise estimator 30 is input. It is characterized by including the configuration.

The adaptive noise estimator 30 includes a finite impulse response filter (hereinafter referred to as a FIR filter) 31 that convolutionally filters N input data and N adaptive filter coefficients from this point in time. ; A first-order linear predictor 32 which first-order linearly predicts and whitens the result filtered from the finite impulse response filter 31 and the input signal; A correlation degree measurer (33) for obtaining a correlation between a result signal output from the first linear predictor (32) and an input signal; A power meter (34) for obtaining the power of the resultant signal and the input signal output from the first linear predictor (32); A step size estimator (35) for obtaining a step size at this point in time using the correlation and power results output from the correlation measurer (33) and the power measurer (34); An adaptive filter updater 36 for obtaining adaptive filter coefficients using the step size obtained by the step size estimator 35; And a filter coefficient memory 37 for storing the adaptive filter coefficients obtained by the adaptive filter updater 36 for use in the finite impulse response filter 31 at the next data input.

The adaptive residual noise estimator 40 generates an input using a signal input from the first microphone 10 and a signal input from the second microphone 20 filtered by the finite impulse response filter 31. A control input constant generator 41 for generating a necessary control input constant; A control input calculator 42 for obtaining a control input by using the control input constant generated by the control input constant generator 41; A finite impulse response filter (43) for obtaining convolution with control input data output from the control input calculator (42) and adaptive filter coefficients; A reference signal calculator (44) for calculating a reference signal by using the past M / 2 among the results output from the adaptive noise estimation unit (30); And a filter coefficient memory 45 for storing the estimated adaptive residual noise coefficient by changing the superimposed input signal obtained by the finite impulse response filter 43 and the reference signal calculated by the reference signal calculator 44. It is characterized by the configuration.

The adaptive residual noise attenuation unit 50 reduces the residual signal by using the filter coefficients stored in the filter coefficient memory 45 of the adaptive residual noise estimation unit 40 and the residual signal obtained by the adaptive noise estimation unit 30. It comprises a finite impulse response filter (51).

The operation principle of the adaptive noise canceling apparatus according to the object of the present invention will be described in detail as follows.

As shown in FIG. 2, the first microphone 10 receives voice and background noise, and the second microphone 20 mainly receives background noise.

In this case, an adaptive filter having a finite impulse response (FIR) filter structure is used to improve the sound quality of the noise speech.

On the other hand, the signal of the first microphone 10 is a reference signal of this adaptive filter, and the signal of the second microphone 20 is an input signal of the adaptive filter.

That is, the relationship between the noise component of the first microphone 10 signal and the noise component of the second microphone 20 signal is updated for each sample using an adaptive filter, and the relationship of the first microphone 10 is updated using this relationship. Eliminate noise components in the signal.

As described above, the signal of the first microphone 10 whose primary noise has been removed through the adaptive filter passes through a residual noise attenuator of the adaptive filter type to remove noise more accurately.

The adaptive process of the adaptive filter is performed for each sample by using the first-order linear prediction technique, the time-varying step size algorithm, and the existing least square mean normalization (NLMS) method proposed by the present invention. Lose.

This adaptive filtering process is referred to as Whitening and Variable stepsize-WVS-NLMS (Normalized LMS) method compared to the conventional least squares average normalization (NLMS) adaptive method.

The time-varying step size algorithm refers to the correlation between the power of the input signal, the power of the output signal, the input signal and the output signal in order to have the optimal value for each sample of the change in the filter coefficient in the least square average normalization (NLMS) method. It means to find the step size using.

In addition, the time-varying step size algorithm determines the step size in inverse proportion to the input power and the output power. Therefore, when there is noise, the output of the adaptive filter is abnormally large. There is an advantage.

First, as shown in FIG. 3, the adaptive noise estimator 30 filters N input data and N adaptive filter coefficients from the present time through a finite impulse response (FIR) filter 31. Find the convolution.

As shown in FIG. 2, the signal has a signal contained in the second microphone 20, and is a signal obtained by estimating a noise component contained in the first microphone 10. The signal is the first microphone 10. It represents the difference from the signal input to, and is used to obtain the result of the adaptive noise estimation unit 30.

Using this result, a whitening variable stepsize-least squared mean normalization (WVS-NLMS) algorithm is applied.

The whitening variable step size-least squared mean normalization (WVS-NLMS) algorithm comprises a first linear prediction step; Correlation measure step; Power measurement step; Step size estimation step; It is divided into five parts of the adaptive filter update phase.

first ; In the first linear prediction step, first-order linear prediction of the input signal and the resultant signal using the first-order linear predictor 32 obtains a pseudo-whitening effect.

The calculation formula is as follows.

[Equation 4]

E (n) = e (n)-k * e (n-1)

[Equation 5]

X (n) = x (n)-k * x (n-1)

Where k uses a constant of 0.95. Index n is a time variable.

second ; In the correlation measurement step, the correlation between the resultant signal output from the first linear predictor 32 and the input signal using the correlation degree measurer 33 is obtained as in Equation 6.

[Equation 6]

Pex [i] = a * Pex [i] + (1-a) * E (n) * X (n)

Here, the constant a uses a value of 0.995. Index i has a value between 0 and the number of taps (N).

third ; In the power measurement step, power is calculated using the power measurer 34 with respect to the input signal and the resultant signal of the first linear predictor 32 as shown in Equations 7 and 8 below.

[Equation 7]

Px = a * Px + (1-a) * X * X

[Equation 8]

Pe = a * Pe + (1-a) * E * E

fourth ; The step size estimating step uses the results of Equations 6 to 8 output from the correlation degree estimator 33 and the power meter 34 in the step size estimator 35, and as shown in Equation 9 below, the step size at this time Find (u).

[Equation 9]

Fifth ; In the adaptive filter updating step, the adaptive filter updater 36 obtains the adaptive filter coefficient W as in Equation 10 using the step size u obtained in Equation 9 above.

[Equation 10]

w (n) = w (n-1) + u * E * X (n) / X (n) ^T X (n)

Here, w (n), w (n-1), and X (n) are N-dimensional vectors corresponding to time n.

The filter coefficient W thus obtained is stored in the adaptive filter coefficient memory 37 to be used for the finite impulse response (FIR) filter 31 at the next data input.

As shown in FIG. 4, the adaptive residual noise estimator 40 makes an input of the adaptive residual noise estimator 40 through the control input constant generator 41.

The control input constant C generated by the control input constant generator 41 is obtained as shown in Equation 11 below.

[Equation 11]

C = min (1.0, Ey / Ee)

Here, min (a, b) specifies a smaller value among a, b, and Ey corresponds to M estimated powers estimated by the adaptive noise estimation unit 30.

That is, M convolutioned values of the signal of the second microphone 20 and the adaptive filter coefficients are collected, squared, and then added together.

Similarly, Ee is the sum of the squares of the M noise values of the adaptive noise estimator and then the sum of them.

On the other hand, since the control input constant (C) has a value smaller than 1.0 to maintain stability, the “min ()” function is used.

The control input calculator 42 obtains the control input z using the control input constant C generated by the control input constant generator 41 as shown in the following equation (12).

[Equation 12]

z (n) = C * d (n) + (1-C) * e (n)

Here, d (n) corresponds to the second microphone 20 signal, and e (n) corresponds to the result of the adaptive noise estimator.

The control input z obtained as in Equation 12 is used as an input of the adaptive residual noise estimator 40. When the control input adjustment constant C is 0, the estimated noise is applied to the signal of the second microphone 20. Corresponds to the subtracted value

This value contains less noise components than the second microphone 20 signal, where the control input adjustment constant C is one.

The reference signal calculator 44 uses the past M / 2 of the results of the adaptive noise estimator 30 to calculate the reference signal r.

As described above, in the whitening variable step size-least-squared-average normalization (WVS-NLMS) algorithm, it is the coefficient H of the adaptive residual noise estimation unit 40 that changes the input y2 and the reference signal r. .

The coefficient H of the adaptive residual noise estimator 40 thus obtained is stored in the filter coefficient memory 45 and used as a filter coefficient when the adaptive residual noise attenuator 50 attenuates the residual noise.

Meanwhile, as shown in FIG. 5, the adaptive residual noise attenuator 50 includes the filter coefficients stored in the filter coefficient memory 45 of the adaptive residual noise estimator 40 and the adaptive noise estimator 30. Using the obtained result value, the filtering is performed through the finite impulse response (FIR) filter 51, thereby attenuating the adaptive residual noise, and estimating the speech signal with the residual noise attenuated.

As described in detail above, the present invention can be used as a preprocessing part of a codec algorithm in which sound quality drops sharply when noise is present in a communication system such as a digital cellular phone, thereby improving sound quality through noise removal.

In addition, in a speech recognition system, if the noise is severe, not only the speech recognition rate drops but also a large number of errors occur even in the detection of a speech section.

The whitening variable step size-least squared mean normalization (WVS-NLMS) algorithm can be used in various system estimation methods.

Claims (14)

A first microphone 10 for receiving voice and background noise; A second microphone 20 for receiving background noise and voice; An adaptive noise estimator (30) for obtaining a relationship between two input signals by using the signal of the second microphone (20) as an input and the signal of the first microphone (10) as a reference signal; The signal of the first microphone 10, the result of the adaptive noise estimator 30, and an adjustment constant are input, and the result of the noise estimator of the previous M / 2 time input from the delay unit 60 is used as a reference signal. An adaptive residual noise estimator 40 for storing the relationship between these signals in an adaptive filter coefficient; An adaptive residual noise attenuation unit 50 for estimating a speech signal attenuated by the residual noise using the filter coefficients of the adaptive residual noise estimator 40 as the input of the result of the adaptive noise estimator 30 is input. Adaptive noise canceling device, characterized in that it comprises a. According to claim 1, The adaptive noise estimation unit 30, A finite impulse response filter 31 which convolutions and filters the N input data and the N adaptive filter coefficients from this point in time; A first-order linear predictor 32 which first-order linearly predicts and whitens the result filtered from the finite impulse response filter 31 and the input signal; A correlation degree measurer (33) for obtaining a correlation between a result signal output from the first linear predictor (32) and an input signal; A power meter (34) for obtaining the power of the resultant signal and the input signal output from the first linear predictor (32); A step size estimator (35) for obtaining a step size at this point in time using the correlation and power results output from the correlation measurer (33) and the power measurer (34); An adaptive filter updater 36 for obtaining adaptive filter coefficients using the step size obtained by the step size estimator 35; And a filter coefficient memory (37) for storing the adaptive filter coefficient obtained by the adaptive filter updater (36) for use in the finite impulse response filter (31) at the next data input. . According to claim 1, The adaptive residual noise estimator 40 is A control input constant generator for generating a control input constant necessary for making an input using a signal input from the first microphone 10 and a signal input from the second microphone 20 filtered by the finite impulse response filter 31. (41) and; A control input calculator 42 for obtaining a control input by using the control input constant generated by the control input constant generator 41; A finite impulse response filter (43) for obtaining convolution with control input data output from the control input calculator (42) and adaptive filter coefficients; A reference signal calculator (44) for calculating a reference signal by using the past M / 2 among the results output from the adaptive noise estimation unit (30); And a filter coefficient memory 45 for storing the estimated adaptive residual noise coefficient by changing the superimposed input signal obtained by the finite impulse response filter 43 and the reference signal calculated by the reference signal calculator 44. Adaptive noise canceling device, characterized in that configured. According to claim 1, The adaptive residual noise attenuation unit 50 is And a finite impulse response filter 51 which reduces the residual signal by using the filter coefficients stored in the filter coefficient memory 45 of the adaptive residual noise estimation unit 40 and the residual signal obtained by the adaptive noise estimation unit 30. Adaptive noise canceling device, characterized in that. An adaptive noise estimation process of obtaining a relationship between two input signals by using a signal of the second microphone 20 as an input and a signal of the first microphone 10 as a reference signal; Using the signal of the first microphone 10 and the result of the adaptive noise estimator 30 and the adjustment constant as inputs, the results of the noise estimator of the past M / 2 time input from the delay unit 60 are used as reference signals. An adaptive residual noise estimation process for storing the relationship between signals in an adaptive filter coefficient; The result of the adaptive noise estimator 30 is input, and the adaptive residual noise estimator 40 includes an adaptive residual noise attenuation process for estimating a speech signal in which the residual noise is attenuated. Adaptive noise reduction method. According to claim 5, The first process of estimating the adaptive noise, Convoluting and filtering the input signal and the adaptive filter coefficients using the finite impulse response filter 31; A first-order linear prediction step of first-order linear prediction using the first-order linear predictor 32 to whiten the input signal and the adaptive filter result signal; A correlation degree measuring step of obtaining a correlation between an input signal and a resultant signal output from the first linear predictor 32 using a degree of correlation measurer 33; A power measurement step of obtaining power from an input signal and a resultant signal of the first linear predictor 32 using a power measurer 34; A step size estimating step of obtaining a step size u at this point in time using the results output from the correlation degree estimator 33 and the power meter 34 in the step size estimator 35; An adaptive filter updating step of obtaining an adaptive filter coefficient W in the adaptive filter updater 36 using the step size u obtained in the step size estimating step; And storing the filter coefficient (W) obtained in the adaptive filter updating step in the adaptive filter coefficient memory 37 to be used in the finite impulse response (FIR) filter 31 at the next data input. Adaptive noise reduction method. According to claim 6, The first linear prediction step, If k uses a constant S0.95 and index n is a time variable, E (n) = e (n)-k * e (n-1) X (n) = x (n)-k * x (n-1) Adaptive noise reduction method, characterized in that represented by the equation of. According to claim 6, The correlation degree measuring step, When the constant a uses a value of 0.995 and the index i has a value between 0 and the number of taps (N), Pex [i] = a * Pex [i] + (1-a) * E (n) * X (n) Adaptive noise reduction method, characterized in that represented by the equation of. According to claim 6, The power measurement step, Px = a * Px + (1-a) * X * X Pe = a * Pe + (1-a) * E * E Adaptive noise reduction method, characterized in that represented by the equation of. According to claim 6, The step size estimation step, Adaptive noise reduction method, characterized in that represented by the equation of. According to claim 6, The adaptive filter updating step, When w (n), w (n-1), and X (n) are N-dimensional vectors corresponding to time n, w (n) = w (n-1) + u * E * X (n) / X (n) ^T X (n) Adaptive noise reduction method, characterized in that represented by the equation of. According to claim 5, The second process of estimating the adaptive residual noise is A control input constant generation step of making an input of the adaptive residual noise estimator 40 through the control input constant generator 41; A control input calculating step of obtaining a control input z by the control input constant C through the control input calculator 42; Obtaining a convolution with control input data output from the control input calculator 42 and adaptive filter coefficients through a finite impulse response filter 43; A reference signal calculating step of calculating a reference signal r using a past M / 2 among the results of the adaptive noise estimation unit 30 through a reference signal calculator 44; In the whitening variable step size-least squared mean normalization (WVS-NLMS) algorithm, the coefficient H of the adaptive residual noise estimator 40 obtained by changing the input y2 and the reference signal r is converted into a filter coefficient memory ( 45). The method of claim 12, The control input constant generation step, min (a, b) specifies a smaller value of a, b, and Ey corresponds to M estimated power estimated by the adaptive noise estimation unit 30, C = min (1.0, Ey / Ee) Adaptive noise reduction method, characterized in that represented by the equation of. The method of claim 12, The control input calculation step, d (n) corresponds to the second microphone 20 signal, and e (n) corresponds to the result of the adaptive noise estimator, z (n) = C * d (n) + (1-C) * e (n) Adaptive noise reduction method, characterized in that represented by the equation of.