JP4879195B2 - Noise reduction device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reduce a noise component without deteriorating audio itself caused by the action of a filter. <P>SOLUTION: The noise reduction apparatus comprises: a first linear prediction filter to which an output signal of a first microphone is inputted; a second linear prediction filter to which an output signal of a second microphone is inputted; a first adaptive filter to which an output signal from the first linear prediction filter is inputted and of which a filter coefficient is updated so as to minimize an error signal between output signals of the second linear prediction filter and of itself; a second adaptive filter to which the output signal of the first microphone is inputted; a subtractor which outputs a signal resulting from subtracting an output signal of the second adaptive filter from the output signal of the second microphone; and an inverse filter to which the signal outputted from the subtractor is inputted to restore a second audio signal inputted to the second microphone. <P>COPYRIGHT: (C)2008,JPO&amp;INPIT

Description

  The present invention relates to a noise reduction device for reducing noise components and outputting them from voice input to a microphone.

  Under actual circumstances, it is inevitable that noise is superimposed on the voice collected by the microphone. For this purpose, methods for enhancing noise by reducing noise have been studied, and many methods have been proposed so far. Generally, these conventional methods are divided into the following two types based on the number of microphones used. The first is a method that is applied to a system with one microphone, and the spectral subtraction method (SS method) is a representative example. The second is a method applied to a microphone array having a plurality of microphones.

  These methods are not suitable for the environment in which they are applied. The SS method is advantageous in an environment where many noises are incident from many directions. On the other hand, when the incident angles of noise and voice to the microphone are clearly different and there is a sufficient space where a large number of microphones can be installed, the microphone array is more advantageous. This is because the noise removal performance is expected to be improved as much as the spatial information related to voice and noise can be used.

The microphone array systems proposed so far have a problem in that there is a restriction on the distance between the microphones. This problem is solved by a noise reduction method using a microphone array system (see Non-Patent Document 1, for example).
Hirofumi Nakano, Kensaku Fujii, Yoshio Ito, “Noise Reduction Method Using a Two-Microphone Array System,” IEICE Technical Report, EA2006-14, May. 2006 S. Heykin, “Introduction to Adaptive Filters,” Miki Takebe (translation), Hyundai Engineering, 1999 Kensaku Fujii, Go Tada, Meiji Muneyasu, “Preliminary Study for Separation of Men and Women by Microphone Array,” 2003 Spring Nihon Ronshu, 2-8-19, March, 2003

  However, the EE ′ method used in this method has a problem that the operation of the system becomes unstable. Therefore, the present invention aims to provide a new technique applied to a microphone array system in order to solve this problem.

  In order to solve the above-described problems, a noise reduction device according to the present invention is arranged at a position different from the first microphone and a first microphone into which voice and noise are input as a first voice signal and a first noise signal, respectively. And the second microphone to which the voice and the noise are input as the second voice signal and the second noise signal, respectively, the output signal of the first microphone, and the voice from the output signal of the first microphone. The first linear prediction filter whose filter coefficient is updated so as to remove the correlation and the output signal of the second microphone are input, and the filter coefficient is updated with the same value as the filter coefficient of the first linear prediction filter A second linear prediction filter, and an output signal from the first linear prediction filter is input to the second linear prediction filter. The first adaptive filter whose filter coefficient is updated so as to minimize the power of the signal obtained by subtracting one of the output signal and the self output signal from the other, and the output signal of the first microphone are input. In addition, the second adaptive filter whose filter coefficient is updated with the same value as the filter coefficient of the first linear prediction filter, the output signal of the second microphone, and the output signal of the second adaptive filter are changed from one to the other. A subtractor that outputs an error signal obtained by subtraction; and an inverse filter that receives the signal output from the subtractor and restores the second audio signal input to the second microphone.

Further, in the noise reduction device, when the acoustic propagation characteristic from the first microphone to the second microphone is h ^ab and the acoustic propagation characteristic from the second microphone to the first microphone is h ^ba , the inverse filter ^May be a filter having a function of 1 / (1-h ^ab h ^ba ).

In the noise reduction device, the acoustic propagation characteristic ^hab may be the same value as a filter coefficient of the first adaptive filter.

Further, in the noise reduction device, a third linear prediction filter in which an output signal of the second microphone is input and a filter coefficient is updated so as to remove a speech correlation from the output signal of the second microphone; A fourth linear prediction filter in which an output signal of the first microphone is input and a filter coefficient is updated with the same value as a filter coefficient of the third linear prediction filter, and an output signal from the third linear prediction filter is A third adaptive filter whose filter coefficients are updated so as to minimize the power of the error signal obtained by subtracting one of the output signal from the fourth linear prediction filter and its own output signal from the other. In addition, the acoustic propagation characteristic h ^ba may be the same value as the filter coefficient of the third adaptive filter.
Further, in the noise reduction device, the first adaptive filter and the second adaptive filter are updated by any one of a plurality of step sizes, and the output power of the second adaptive filter with respect to the output power of the first subtractor. A large step size may be employed when the power is relatively large, and a small step size may be employed when the output power of the second adaptive filter relative to the power of the first subtractor is relatively small. .

In the noise reduction apparatus, when P _Ej is the output power of the first subtractor, P _Yj is the output power of the second adaptive filter, and E is an arbitrary value, the first subtractor The magnitude of the output power of the second adaptive filter relative to the output power may be determined by 2E P _Yj / (P _Ej + E P _Yj ).
In the noise reduction apparatus, a fourth adaptive filter in which an output signal of the second microphone is input and a filter coefficient is updated with the same value as the filter coefficient of the third adaptive filter, and the first microphone A second subtractor for outputting an error signal obtained by subtracting one of the output signal and the output signal of the fourth adaptive filter from the other, and a plurality of the third adaptive filter and the fourth adaptive filter are provided. If the output power of the fourth adaptive filter is relatively large with respect to the output power of the second subtractor, the larger step size is adopted, and the output of the second subtractor is updated. When the output power of the fourth adaptive filter relative to the power is relatively small, a small step size is adopted. It may be.

  According to the noise reduction device of the present application, it is possible to reduce the noise component without causing deterioration of the sound itself due to the function of the filter.

  Hereinafter, preferred embodiments of the present invention will be described.

(1.2 Basic structure of microphone array system)
Figure 1 shows the basic structure of the microphone array system. In the figure, the microphones A and B are used, and ADF (Adaptive digital filter) indicates an adaptive filter. In the present embodiment, it is assumed that the audio signal s _j comes from the left direction and the noise signal n _j comes from the right direction.

Here, assuming that the acoustic system between the microphones AB is h ^ab , the noise signal n ^b _j input to the microphone B uses the noise signal vector n ^a _j arriving at the microphone A,

It can be expressed as. Similarly, if the acoustic system from microphones B to A is h ^ba , the audio signal s ^a _j input to the microphone A is the audio signal s ^b _j arriving at the microphone B.
Using,

It is indicated. At this time, signals x ^a _j and x ^b _j output from the microphones A and B are respectively

It becomes.
Here, a normal learning identification method is used to update the coefficient H ^ab _j of the adaptive filter ADF (see, for example, Non-Patent Document 2).

Consider the case of using. In the above equation, e _j is the estimation error,

It is expressed. Y _j is the output signal of the adaptive filter,

It is indicated.
At this time, considering the k-th element of the adaptive filter, the numerator of the second term on the right-hand side is

It becomes. Here, among the components of equation (8), the component whose average is 0 has no effect on the coefficient update of the adaptive filter. That is, noise and voice have no correlation, so if you ignore them,

It can be said. Further substituting equations (1) and (7),

And expanded. Similarly, voice and noise have no correlation, so if you ignore them,

It becomes. Here, the first item s ^b _j s ^a _j (k) on the right side indicates that the adaptive filter operates as a linear predictor for speech. In the second term, even if there is no autocorrelation in the speech, H ^ab _j (k) s ^a _j (k) 2 does not become 0, and the adaptive filter coefficient update due to the estimation error according to the third term is disturbed.

(2. Proposed microphone array system)
FIG. 2 shows the basic structure of the microphone array system according to this embodiment. LPF (Linear Prediction Filter) in the figure shows a linear predictor. However, the linear predictor connected to the microphone B is not updated, and the filter has the same coefficient as the filter coefficient of the linear predictor connected to the microphone A. Similarly, the adaptive filter at the bottom of the figure is not updated, and the coefficient of the adaptive filter at the center of the figure is copied. In FIG. 2, the output signals x ^{a ′} _j and x ^{b ′} _j of the linear predictors connected to the microphones A and B are respectively

It can be expressed as. Here, s ^{a ′} _j and n ^{a ′} _j indicate speech components and noise components included in x ^{a ′} j, and s ^{b ′} _j and n ^{b ′} _j indicate speech components and noise components of x ^{b ′} . Applying the learning identification method using these signals,

It is indicated. E ′ _j is the estimation error in the proposed method,

It can be expressed as. Where y ′ _j is the output signal of the adaptive filter,

It becomes.
Using these, the same expression expansion as the previous section,

In addition,

And can be expanded. Here, there is no correlation between voice and noise, so if you ignore them,

It becomes. If the speech correlation can be removed by the action of the linear predictor, the expected value of s ^{b ′} _j s ^{a ′} _j (k) is approximated to zero. Therefore,

Can be rewritten. The second term on the right side is

Thus, the influence on the system identification of the first term can be reduced.
Furthermore, the output e _{j at} this time is

If H ^ab can be approximated to h ^ab , then

It becomes. As a result, it is expected that a voice-only signal from which noise can be removed can be extracted as an output signal.

(3. Confirmation of system operation by simulation)
The convergence characteristics of the adaptive filter ADF are confirmed by simulation. Estimation error as an evaluation of simulation

Is used. However,

And T indicates the number of taps. As simulation conditions,
・ Linear predictor LPF
Update method: Learning identification method Step size μ: 0.1
Number of taps: 128
・ Adaptive filter ADF
Update method: Learning identification method Step size μ: 0.01
Number of taps: 128
・ Input signal audio signal s _j : male audio noise signal with sampling period 8 kHz n _j : colored noise signal with a power ratio of 0 dB to audio signal and acoustic propagation characteristics between microphones AB: 0.9z- ²
And

  As a result of the simulation under these conditions, FIG. 3 is obtained. From the figure, it can be confirmed that the adaptive filter oscillates around the estimated error of about -10dB. This is thought to be because the voice works as a disturbance and the influence cannot be ignored in the section where the voice is loud and the estimation is disturbed. Also, the noise suppression effect of the output signal

Represented by In the above equation e ^s, e ⁿ denotes an audio signal component contained in the output signal e, I denotes the number of input samples. The signal-to-noise ratio of the output signal is 12.3 dB. It can be confirmed from the index that noise can be reduced.

(4. Update control by voice power)
In this embodiment, the voice works as a disturbance. This is clear from equation (20). Therefore, updating in a section where the disturbance is large leads to a deterioration in estimation. Therefore, in this section, a method for controlling the update using the output sound e _j and the filter output y _j in FIG. 2 will be described. In this embodiment, in order to realize the update control, the step size μ is set to

Decide. Here, E is a parameter for controlling the step size, and is set to 0.01 in this embodiment. P _Yj and P _Ej are the power of the output y _{j of the} adaptive filter and the output e _j of the system, respectively. These P _Yj and P _Ej are derived by a block diagram as shown in FIG.

  In the block diagram:

And α ₁ = 1-1 / 1024 and α ₂ = 1-1 / 256. This is because the noise used in the present embodiment is stationary, so that the value of P _Yj is kept constant as much as possible so that the time change is small and small.

And P _Yj and P _Ej are determined. The reason why P _Ej is obtained in this way is to quickly correspond the step size to the input of the audio signal. In the present embodiment, the step size obtained by the equation (28) is

And restrict. The convergence characteristic of the adaptive filter when the step size μ is determined by the above method is confirmed by simulation.

  When simulation is performed under the same conditions as in the previous section, the transition of the estimation error is as shown in FIG. According to the figure, the estimation error changes at about -20 dB, and it can be confirmed that the performance is improved as compared with FIG. The output signal-to-noise ratio is 23.0 dB, and it can be confirmed that the noise is almost inconspicuous by looking at the output signal waveform in FIG.

(5. Examination for voice restoration)
Consider the output signal e _j next obtained. According to the above-described formula (23), the output signal e _j obtained is

Thus, a signal from which noise is removed can be obtained. However, this signal is obtained by modulating the audio signal s ^b _j observed by the microphone B. This means that the sound quality of the obtained output signal is degraded.

Therefore, in this section, a method for restoring the original speech s ^b _j using the output signal e _j obtained by the methods up to the previous section will be described. In the above equation, s ^a _j is

There is a relationship.

Therefore, if an IIR (Infinite Impulse Response) filter having a function 1 / (1-H ^ab _j ^T h ^ba ) is formed and the output signal e _{j of the} system is input,

Thus, an audio signal s ^b _j that can be observed with the microphone B can be obtained.

However, in the above equation, h ^ba is the acoustic propagation characteristic of microphones B to A. Therefore, h ^ba _j between the microphones BA is estimated by applying the method for obtaining the acoustic characteristic H ^ab _j between the microphones AB up to the previous section also between the microphones BA. If the obtained H ^ba can be estimated sufficiently, it can be approximated to h ^ba, and the target filter is created by substituting this. A system for realizing this is shown in FIG. As shown in the figure, by inputting e _j to the inverse filter, it is possible to obtain an output signal o _j that can restore the voice.

A system for estimating H ^ba _j is shown in FIG. The ADF filter coefficient vector in the figure matches the target H ^ba _j . Here, H ^ba _j must be updated in consideration of the audio signal as a reference signal. In other words, if the signal component s ^{b ′ ′} (k) contained in the component x ^{b ′ ′} _j (k) of the signal vector input to H ^ba _j is small, the reference signal is buried in the disturbance. If you update with, the estimation will get worse. However, since s ^{b ′ ′} (k) is a signal that cannot actually be observed, the output signal o _j − _k is obtained by inputting it to the linear predictor used to generate s ^{b ′ ′} _j (k). The update tap is controlled using the signal o ′ _{j -k} . Expressing this update equation for H ^ab _j of K-th element H ^ab _j (k),

It becomes. Where P _s is the average power of the speech, and e ^{b ′} _j is the estimated error signal,

It is indicated. Also, assuming that the LPF characteristic in FIG. 9 is L _j , o ′ _{j −k} in the above equation is

It can be expressed as.

(6. Confirmation of voice restoration by simulation)
The operation of the proposed method is confirmed by simulation. First, to confirm the convergence characteristics of the H ^ba _j in FIG. For the adaptive algorithm, the method shown in the previous section is used and its parameters are:
・ Linear predictor LPF
Update method: Learning identification method Step size μ: 0.1
Number of taps: 128
・ Adaptive filter ADF
Step size μ: 0.01
Number of taps: 128
And Other conditions are the same as the simulation conditions in Section 5.

When simulation is performed under the above conditions, the convergence characteristics are as shown in FIG. From the figure, it can be confirmed that the estimation is up to about -16dB. It can also be confirmed that it takes longer than the convergence of H ^ab _j as compared with FIG. This is thought to be due to the fact that the frequency of updates has decreased due to update control.

Next, the sound restoration effect is confirmed by simulation. In the present embodiment, a sound quality evaluation index VE and a comprehensive evaluation index OE are used as evaluation indexes in addition to the SNR _out of Expression (27). Respectively,

It is indicated. In the above equation, o ^s _j is an audio signal component included in the output signal o _j , and I is the number of input samples.

When simulation is performed under these conditions, FIG. 11 is obtained as the output signal o _j . This indicator VE = 14.7dB, OE = 13.8dB, the SNR _out = 20.3dB. When e _j is evaluated, VE = 2.4 dB, OE = 2.3 dB, and SNR _out = 23.6 dB, which confirms that the proposed method is effective in improving sound quality.

(7. Summary)
In the present embodiment, the algorithm applied to the microphone array system has been described. It was confirmed by simulation that noise can be reduced by estimating the acoustic system between microphones by this method. Since the obtained audio signal has deteriorated, the method of restoring it to the original audio has also been described. It was confirmed that the sound quality of degraded speech can be improved by this method.

  According to the noise reduction device of the present application, the noise component can be reduced without causing deterioration of the voice itself, and thus can be used in the technical field of voice communication.

It is the figure which showed the basic structure of the microphone array system. It is the figure which showed the microphone array system which concerns on this embodiment. It is the figure which showed the operation | movement confirmation of the microphone array system. P _V, is a derivation block diagram of the P _E. It is the figure which showed the estimation error when performing step size control. It is a figure showing a signal waveform inputted to the microphone A side. It is the figure which showed the output signal waveform. It is the figure which showed the audio | voice restoration system. 1 is a diagram showing a system for estimating an acoustic system of microphones B to A. FIG. It is the figure which showed the convergence characteristic of ^Hba . It is the figure which showed the output signal waveform.

Claims (7)

A first microphone into which voice and noise are input as a first voice signal and a first noise signal, respectively;
A second microphone disposed at a different position from the first microphone and receiving the voice and the noise as a second voice signal and a second noise signal, respectively;
A first linear prediction filter in which an output signal of the first microphone is input and a filter coefficient is updated so as to remove a correlation of speech from the output signal of the first microphone;
A second linear prediction filter that receives the output signal of the second microphone and updates the filter coefficient with the same value as the filter coefficient of the first linear prediction filter;
A filter coefficient is provided so as to minimize the power of a signal obtained by subtracting one of the output signal from the second linear prediction filter and the output signal from the second linear prediction filter when the output signal from the first linear prediction filter is input. A first adaptive filter to be updated;
A second adaptive filter in which the output signal of the first microphone is input and the filter coefficient is updated with the same value as the filter coefficient of the first adaptive filter;
A first subtractor that outputs an error signal obtained by subtracting one of the output signal of the second microphone and the output signal of the second adaptive filter from the other;
A noise reduction device comprising: an inverse filter that receives the signal output from the first subtractor and restores the second audio signal input to the second microphone. When the acoustic propagation characteristic from the first microphone to the second microphone is h ^ab and the acoustic propagation characteristic from the second microphone to the first microphone is h ^ba , the inverse filter is 1 / (1-h The noise reduction device according to claim 1, which is a filter having a function of ^ab h ^ba ). The noise reduction device according to claim 2, wherein the acoustic propagation characteristic ^hab is set to the same value as a filter coefficient of the first adaptive filter. A third linear prediction filter in which an output signal of the second microphone is input and a filter coefficient is updated so as to remove a correlation of speech from the output signal of the second microphone;
A fourth linear prediction filter in which an output signal of the first microphone is input and a filter coefficient is updated with the same value as the filter coefficient of the third linear prediction filter;
A filter coefficient that minimizes the power of the error signal obtained by subtracting one of the output signal from the fourth linear prediction filter and its own output signal when the output signal from the third linear prediction filter is input. A third adaptive filter in which is updated,
The noise reduction device according to claim 2 or 3, wherein the acoustic propagation characteristic h ^ba is set to the same value as a filter coefficient of the third adaptive filter.   The first adaptive filter and the second adaptive filter are updated according to any one of a plurality of step sizes, and the step is large when the output power of the second adaptive filter with respect to the output power of the first subtractor is relatively large. The size is adopted, and a small step size is adopted when the output power of the second adaptive filter is relatively small with respect to the output power of the first subtractor. The noise reduction device described in 1. When P _Ej is the output power of the first subtractor, P _Yj is the output power of the second adaptive filter, and E is an arbitrary value, the second adaptation with respect to the output power of the first subtractor The noise reduction device according to claim 5, wherein the magnitude of the output power of the filter is determined by 2E P _Yj / (P _Ej + E P _Yj ). A fourth adaptive filter in which the output signal of the second microphone is input and the filter coefficient is updated with the same value as the filter coefficient of the third adaptive filter;
A second subtractor that outputs an error signal obtained by subtracting one of the output signal of the first microphone and the output signal of the fourth adaptive filter from the other; and
The third adaptive filter and the fourth adaptive filter are updated by any one of a plurality of step sizes, and a large step is obtained when the output power of the fourth adaptive filter relative to the output power of the second subtractor is relatively large. The size is adopted, and a small step size is adopted when the power of the output of the fourth adaptive filter with respect to the power of the output of the second subtracter is relatively small. The noise reduction device described in 1.