JP4928918B2 - Signal processing apparatus using adaptive filter - Google Patents

Description

  The present invention relates to a signal processing apparatus using an adaptive filter, and in particular, a microphone array process for extracting a target signal voice by suppressing interference noise using a microphone array for voice input in a voice recognition apparatus, a video conference apparatus, or the like. The present invention relates to a signal processing apparatus suitable for the apparatus.

  A microphone array processing apparatus that extracts a target signal from a target sound source by suppressing a noise by using a microphone array composed of a plurality of arranged microphones and processing an output signal of the microphone array is a speech recognition apparatus or Development is underway for the purpose of voice input in video conferencing equipment. Among them, for a microphone array processing apparatus using an adaptive beamformer that can obtain a large noise suppression effect with a small number of microphones, for example, Reference 1: “Electronic System and Digital Processing” edited by IEICE, Reference 2: Heykin: “ As described in “Adaptive Filter Theory (Plentice Hall)”, various methods such as a generalized sidelobe canceller, a frosted beam former, and a reference signal method are known.

  The adaptive beamformer processing is basically processing for suppressing interference noise using a filter in which a directional beam having a blind spot in the arrival direction of interference noise is formed. In this adaptive beamformer processing, in order to avoid the problem that the target signal is partly removed and distorted when the target signal is input, or the target signal is distorted when the target signal and interference noise are input simultaneously, The adaptation speed is reduced when a signal is input. In general, when the output signal power of the adaptive beamformer is large, processing for stopping the adaptation is performed assuming that the target signal is input. However, the output signal power increases not only when the target signal is input, but also when conditions such as interference noise change, so in order to stop adaptation completely when the output signal power is large in this way. There's a problem.

  On the other hand, it is disclosed in Reference 3: Julie E. Greenburg et.al .: “Evaluation of an adaptive beamforming method for hearing aids”, pp.1662-1676, Jarnul of Acous. Soc. Of Am. 91 (3), 1992. In the microphone array for hearing aids, the filter update calculation formula (1) based on the NLMS (Normalized Least Mean Square) method is improved as an adaptive algorithm of the adaptive filter, and the reference signal power and By normalizing the adaptation speed based on the sum of output signal powers, adaptation is not completely stopped when the output signal power is increased, and distortion of the target signal is suppressed.

W _{n + 1} = W _n −2 μe _n X _n / P _r (1)
_{W n + 1 = W n -2μe} n X n / (P r + P o) (2)
Here, n is the time, _{W n} is the filter coefficient, mu is a step ^{_{size, e n (= d n -W}} n T X n) is the error signal, _{d n} is desired response, the latest of _{X n} is the reference signal x A reference signal vector in which past samples are arranged as many as the number of tap points of the filter as a vector, _Pr is the reference signal power, and _Po is the output signal power. The reference signal x is generally a signal in which interference noise, a part of the target signal, and background noise are superimposed. In both cases of equations (1) and (2), the adaptive speed of the adaptive filter is determined by the magnitude of the second term on the right side.

  FIG. 14 shows the configuration of a conventional adaptive filter. This adaptive filter includes a filter unit 101 that performs a filter operation on the reference signal x, a subtracter 102 that subtracts the output signal y of the filter unit 101 from the desired response d, and a filter coefficient of the filter unit 101 using the formula (1) or ( The filter update unit 103 performs the update calculation shown in 2).

In the normal NLMS method and the improved NLMS method described in Document 3, the reference signal power _Pr becomes very small in the situation where neither interference noise nor a target signal exists, so that the equations (1) and (2) As can be understood, since the adaptation speed becomes very large, there is a problem that the adaptive filter adapts to background noise. When the adaptation progresses with respect to background noise, the filter coefficient moves away from the filter coefficient suitable for removing the signal that is originally desired to be removed. That is, when the interference noise is a non-stationary signal such as voice, the filter coefficient is moved away from the original value in a portion where the interference noise power is small, and the filter coefficient is converged in a portion where the interference noise power is large. In this repetition process, a high level of interference noise is output until the filter coefficient converges.

  For this problem, for example, an echo canceller controls adaptation based on the level of the reference signal. In other words, adaptation is stopped when the reference signal level is equal to or lower than the threshold value, and adaptation is performed only when the reference signal level is higher than the threshold value, thereby reducing the influence of noise. However, in the adaptive beamformer, as described above, since the interference signal, a part of the target signal and the background noise are superimposed as the reference signal, as described above, when it is adapted only when the reference signal level is large, If there is interference noise below the threshold level, it cannot be removed.

  As described above, in order to suppress the phenomenon that the target signal is distorted when the target signal is input in the adaptive filter, the conventional method for performing the process of stopping the adaptation when the output signal power is large is used only when the target signal is input. In addition, when the situation changes, the output power increases and the adaptation stops, so that there is a problem that a correct adaptive operation cannot be performed.

  Further, in the conventional NLMS method and the improved NLMS method, the adaptation speed becomes very large in the situation where there is no interference noise and no target signal and the reference signal power is small. In order to avoid this problem, the adaptation is stopped when the level of the reference signal is below the threshold value, and the adaptation is performed only when the reference signal is larger than the threshold value. When there is interference noise below the threshold level in the adaptive beamformer, there arises a problem that it cannot be removed.

  The present invention has been made to solve the above-described problems of the prior art, and is an adaptive filter that enables a reliable adaptive operation without distorting a target signal or adapting to background noise. An object of the present invention is to provide a signal processing apparatus using the above.

  In order to solve the above-described problem, the adaptive filter according to the present invention has a non-zero step size, which is a coefficient that determines the update speed of the adaptive filter, when the output signal power is equal to or less than a predetermined threshold value. The basic feature is that the adaptive mode is controlled so that the step size decreases as the output signal power increases when the output signal power exceeds the threshold value. It is possible to prevent adaptation to noise.

  In other words, when the target signal is input to the adaptive filter, the output signal power becomes larger than the threshold value, and the target speed is distorted because the step size is reduced with respect to the increase in the output signal power. It wo n’t end up. On the other hand, when only the interference noise is input to the adaptive filter and only the background noise is input, the output signal power is equal to or less than the threshold value, and the adaptive filter operates as an adaptive filter with a fixed step size. In this case, since the step size is a non-zero, almost constant value, the adaptation speed is proportional to the input level, so that it can adapt quickly to high level interference noise and to low level background noise. Is prevented.

  More specifically, the adaptive filter according to the present invention includes a filter unit that performs a filter operation on at least one channel reference signal and outputs a one-channel signal, and subtracts the output signal of the filter unit from a desired response. And a filter updating means for updating the filter means based on the reference signal and the output signal of the subtracting means. The filter updating means has a predetermined threshold for the output signal power of the subtracting means. When the output signal power is below the threshold value, the coefficient that determines the update speed of the filter means is set to a non-zero, almost constant value, and when the output signal power exceeds the threshold value, the coefficient is used as the output signal. It is characterized by decreasing with increasing power.

  Another adaptive filter according to the present invention includes a filter unit that processes a plurality of channels of input signals and outputs a signal of one channel, and the filter unit is subjected to a predetermined constraint regarding a characteristic change range of the filter unit. A filter update means for updating, and the filter update means compares the output signal power of the filter means with a predetermined threshold value, and determines the update speed of the filter means when the output signal power is below the threshold value. The coefficient is set to a non-zero substantially constant value, and when the output signal power exceeds a threshold value, the coefficient is decreased with respect to the increase of the output signal power.

  More specifically, the filter update means in the adaptive filter according to the present invention compares the output signal power of the adaptive filter with a predetermined threshold value, and updates the filter means when the output signal power is less than or equal to the threshold value. A control means for controlling the speed to be a non-zero, substantially constant value and controlling the coefficient to decrease with respect to an increase in the output signal power when the output signal power exceeds a threshold value, and the control means Update calculation means for performing update calculation of the filter means in accordance with the update calculation formula including the calculated coefficient.

  The signal processing apparatus according to the present invention arranges a blocking filter that removes a component related to a target signal from a signal of a plurality of channels before the above-described adaptive filter, and inputs an output signal of this blocking filter to the adaptive filter as a reference signal.

  Another signal processing apparatus according to the present invention includes a first adaptive filter and a second adaptive filter having a higher adaptive speed than the first adaptive filter, and the first adaptive filter is a second adaptive filter. The update of the filter is controlled based on the output signal. In this way, the convergence speed is further improved.

  Here, in the first adaptive filter, the output signal power is compared with a predetermined threshold value. When the output signal power is equal to or lower than the threshold value, the adaptive speed becomes a substantially constant value, and the output signal power is When the threshold value is exceeded, the update of the filter is controlled so that the adaptive speed changes according to the output signal power of the first and second adaptive filters. More specifically, the first adaptive filter compares its output signal power with a predetermined threshold, and when the output signal power is less than or equal to the threshold, a coefficient that determines the update rate of the filter is non-zero. A substantially constant value is set, and when the output signal power exceeds a threshold value, the coefficient is decreased with respect to an increase in the ratio of the output signal powers of the first and second adaptive filters.

  According to the present invention, it is possible to provide a signal processing device using an adaptive filter that can achieve both prevention of distortion of a target signal and adaptation to noise.

Embodiments of the present invention will be described below with reference to the drawings.
(First embodiment)
With reference to FIG. 1, an embodiment in which the adaptive filter according to the present invention is applied to an adaptive noise canceller will be described. The adaptive filter includes a filter unit 11 configured by a delay line tap filter that performs a filter operation on the reference signal x to obtain an output signal y, and an adaptive filter by subtracting the output signal y of the filter unit 11 from the desired response d. The subtractor 12 that obtains the error signal e that is the output signal of the filter 11 and the filter update unit 13 that updates the filter characteristic (filter coefficient) of the filter unit 11 based on the reference signal x and the error signal e. Hereinafter, the process of updating the filter coefficient of the filter unit 11 is simply referred to as filter update.

  Here, the filter update unit 13 is different from the conventional adaptive filter shown in FIG. In other words, the filter update unit 13 performs an filter update operation that actually performs a filter update operation according to the adaptive mode controlled by the adaptive mode control unit 14 and the adaptive mode control unit 14 that switches and controls the adaptive mode according to the power of the error signal e. And an arithmetic unit 15. Details of the filter update calculation will be described later.

  In the present embodiment, the most common LMS (Least Mean Squre) method will be described as an example of the adaptive algorithm of the adaptive filter. However, if the algorithm can control the adaptive speed, the projection LMS method or the RLS (Recursive Least Squre) method is also used. However, other algorithms may be used.

  In the adaptive noise canceller, for example, as shown in FIG. 2, the microphone M1 is placed in the vicinity of the target signal source, the output signal of the microphone M1 is set as the desired response d, and another microphone M2 is placed in the vicinity of the interference noise source. The output signal of M2 is used as the reference signal x. In this case, the target signal and the background noise are input to the microphone M1, and the interference noise and the background noise are input to the microphone M2. The microphone arrangement and the like are set so that the target signal is not input to the microphone M2 as much as possible, but it is difficult to completely prevent the target signal from being input. That is, part of the target signal component is actually input to the microphone M2 in addition to the interference noise and the background noise.

  In general, such an adaptive noise canceller removes a component correlated with the reference signal x from the desired response d, so that interference noise having a large correlation between the microphones M1 and M2 is removed. Actually, when the target signal component is mixed in the reference signal x as described above, not only the interference noise but also the target signal may be removed and the target signal may be distorted. Need to be late.

Conventionally, the filter update calculation of the adaptive filter by the LMS method is performed based on the above formula (1) or (2). In this embodiment, the power of the error signal e (hereinafter, output) of the adaptive filter is shown. The adaptive mode is switched according to _{Po (referred} to as signal power), and the filter update calculation is performed based on the following equation.
^{_{e n = d n -W n T}} X n (3)
When _{P o} ≦ P _{a
W n + 1 = W n -2μe} n X n P a (4)
When P _o > P _{a
W n + 1 = W n -2μe} n X n P a / P o (5)
Here, P _a is a predetermined adaptive mode switching threshold, W _n is n times updated filtering coefficient of the filter unit 11, mu is a step size, e _n is the output signal (error signal of the adaptive filter ), the d _n desired response, X _n is a reference signal vector and the past samples from the most recent sample arranged by tapping of the specified number of filter portion 11 vector of the reference signal x. The values of the various constants to be set in advance, for example, 100 the number of taps of the filter unit 11, the step size mu 0.001, but 0.001 times the average output signal power threshold _{P a,} depending on the circumstances It is desirable to decide experimentally.

Next, the filter update operation in the present embodiment will be described with reference to FIG. 3, in the second term on the right side representing the update amount of filter update calculation formula (2) (4) (5), applied to the product _e n _{X n} of the reference signal _{X n} and the error signal _{e n} coefficients (filter coefficients The change in the magnitude of the coefficient that determines the update speed is shown in the figure, and the magnitude of this coefficient is distinguished from the step size μ and will be referred to as the effective step size for convenience. FIG. 3A shows the effective step size when the adaptive stop is performed at _a threshold value Pa or less in the formula (2), which is a filter update calculation formula in the conventional improved NLMS method, and the reference signal power P _r. to a have the meanings indicated with respect to the sum of the output signal power P _o, (b) the output signal power P _o the effective step size when the filter update calculation formula based on the embodiment (4) (5) It shows.

In the NLMS method, as described in the prior art, when there is no interference noise and no target signal, the reference signal power P _r or the sum of the reference signal power P _r and the output signal power P _o is low with respect to the background noise. adaptive Te proceeds, in order to prevent the filter coefficients will be away from the value for removing interference noise to be removed originally stopped adaptation when P _r or P _{r +} P _o is equal to or less than the threshold value P _a is doing. In this case, since 0 is the effective step size as shown in FIG. 3 (a), the P _r or P _{r +} P _o is for the interfering noise that becomes equal to or smaller than the threshold value P _a, not at all adapted become.

In contrast, in the present embodiment, in the filter updated by equation (4) when the output signal power P _o is less than or equal to the threshold P _a, when the output signal power P _o is greater than the threshold value P _a is the formula ( The adaptive mode is switched so that the filter is updated according to 5). Equation (4) is a filter updating calculation formula based on the LMS method step size fixed, equation (5) is a filter update calculation formula based on the NLMS method, FIG effective step size for the output signal power P _o (b ).

That is, in the present embodiment, when P _o ≦ P _a , the effective step size is _a non-zero constant value of 2 μ × P _a , and when P _o > P _a , the effective step size is 2 μ * P _a / P _o. become inversely proportional to _{P o.} Note that the effective step size in the case of P _o ≦ P _a need not be a strictly non-zero constant value, but may be an almost constant value.

Accordingly, even in the filter update by the formula (4) in the case where the output signal power P _o is less than or equal to the threshold P _a, it does not stop perfectly adapted, can be adapted to interfere noise, moreover formula (4) Step Since it is a fixed size filter update, it is unlikely to adapt to low-level background noise. On the other hand, the filter update by the formula (5) in the case where the output signal power P _o is greater than the threshold P _a, so that the output signal power P _o in FIG. 3 (b) is shown in more portions threshold P _a to, the effective step size is inversely proportional to the output signal power P _o adaptation comes late, it is possible to reduce the distortion of the target signal.

Next, the flow of the adaptive filter process described above will be described with reference to the flowchart shown in FIG.
First, step size μ, adaptive mode switching threshold value P _a , and filter length K are set as initial settings, all filter coefficients are set to 0, and the number of filter updates n is set to 0 (step S11).

Next, after constructing the reference signal vector _Xn from a plurality of samples of the reference signal x,
An adaptive filter output signal (error signal) _{e n} calculated by Equation (3) to calculate the power _{P o} of the output signal _{e n.} Calculation of the output signal power P _o, for example, from the latest value of the signal e _n to use a mean value of the past values of up to several minutes taps of the filter section 11 (step S12).

Then, by comparing the output signal power _{P o} and the threshold _{P a} in step S13, in the case of _{P o} ≦ _{P a} flow proceeds to step _S14, in the case of P o> _{P a} flow proceeds to step S15. In step S14, the filter is updated according to equation (4), and the process proceeds to step S16. In step S15, the filter is updated according to the equation (5), and the process proceeds to step S16. In step S16, n is incremented and the process returns to step S12. The above processing is repeated for the number of input data.

Thus, in this embodiment, the adaptive mode of the adaptive filter in response to the output signal power P _o by switching control between the LMS method and NLMS method step size fixed, the distortion reduction and background noise target signal It is possible to achieve both prevention of adaptation.

In the present embodiment, the effective step size in the case of P _o > P _a is set to μ * P _a / P _o as shown in the equation (5), but is inversely proportional to the output signal power P _o , but it is not necessarily accurate. It is not necessary to make it in inverse proportion, and it is sufficient to decrease the output signal power _Po with respect to the increase. In addition, instead of reducing the effective step size in the case of P _o > P _a only depending on the increase in the output signal power P _o , Equation (2), which is a filter characteristic update calculation formula in the prior art, is used. As shown, it may be decreased with respect to an increase in the sum of the reference signal power _Pr and the output signal power _Po .

(Second Embodiment)
In the present embodiment, a case where the adaptive filter according to the present invention is extended to multiple channels will be described.

  As shown in FIG. 5, this adaptive filter performs a filter operation on a reference signal x of a plurality of channels (M channels) by unit filter units 21-1 to 21-M, and then adds the sum y of the output signals to an adder. 26, a subtractor 22 that subtracts the output signal y of the filter unit 21 from the desired response d to obtain an error signal e that is an output signal of the adaptive filter, a reference signal x, and an error signal e. The filter update unit 23 updates the characteristic (filter coefficient) of the filter unit 21 based on the above.

  The filter update unit 23 includes an adaptive mode control unit 24 that performs switching control of the adaptive mode according to the power of the error signal e, and a filter update calculation that actually performs a filter update calculation according to the adaptive mode controlled by the adaptive mode control unit 24. Part 25.

  Next, the operation of the adaptive filter of this embodiment will be described. The number of channels is M, and the number of taps of each of the unit filter units 21-1 to 21-M per channel is K. As the adaptive algorithm of the adaptive filter, the most general LMS method will be described as an example. However, as long as the algorithm can control the adaptive speed, the projection LMS method or the RLS method may be used, as in the first embodiment.

The multi-channel filter unit 21 includes M unit delay line tap filters each including a delay line with a tap as shown in FIG. 6 and a coefficient multiplier that multiplies each tap output by a coefficient. M is configured. The filter characteristic of the multichannel filter unit 21 is a vector in which k-th coefficients w _{m, k} of the _m- th channel delay line tap filter are arranged.
W _n = (w _1,1 , w _1,2 ,..., W _{1, K} ,
..., w _M-1,1 , ..., w _{M-1, K} ) ^T (6)
In FIG. 5, the reference signal x is represented by the following equation in which samples from the latest sample xm _{, n} of the m-th channel reference signal to the number of filter taps are arranged.

_Xn = (x1 _{, n} , x1 _{, n-1} ,..., X1 _{, n-K + 1} ,
..., xM _{, n} , xM _{, n-1} , ..., xM _{, n-K + 1} ) ^T (7)
When the above W _n and X _n are used, the filter update calculation is performed based on the following expression, just like Expressions (3), (4), and (5).

^{_{e n = d n -W n T}} X n (8)
When _{P o} ≦ P _{a
W n + 1 = W n -2μe} n X n P a (9)
When P _o > P _{a
W n + 1 = W n -2μe} n X n P a / P o (10)
Here, P _a is a predetermined adaptive mode switching threshold, e _n is the output signal (error signal) of the adaptive filter, P _o is the output signal power, the μ is the step size. The values of the various constants to be set in advance, for example, the unit filter unit 11-1 to 11-M each number of taps 100, a step size mu 0.001, 0 of the average output signal power threshold _{P a.} Although it is 001 times, it is desirable to determine experimentally according to the situation.

Also in this embodiment, as in the case of the first embodiment, the adaptive mode is switched between LMS and NLMS with a fixed step size by the output signal power _Po of the adaptive filter. The fixed size LMS method prevents adaptation to noise, and the NLMS method of Equation (10) reduces the distortion of the target signal.

  Since the processing flow of the adaptive filter of this embodiment is the same as that of the first embodiment, it will not be described again.

As described above, in the present embodiment, the adaptive signal of the adaptive filter is switched and controlled between the LMS method and the NLMS method with a fixed step size according to the output signal power _Po in the multi-channel input adaptive filter. It is possible to achieve both reduction of distortion and prevention of adaptation to background noise.

In this embodiment, the effective step size in the case of P _o > P _a is set to μ * P _a / P _o as shown in the equation (10), but is inversely proportional to the output signal power P _o , but it is not necessarily accurate. It is not necessary to make it in inverse proportion, and it is sufficient to decrease the output signal power _Po with respect to the increase. In addition, instead of reducing the effective step size in the case of P _o > P _a only depending on the increase in the output signal power P _o , Equation (2), which is a filter characteristic update calculation formula in the prior art, is used. As shown, it may be decreased with respect to an increase in the sum of the reference signal power _Pr and the output signal power _Po .

(Third embodiment)
Next, an embodiment in which the adaptive filter described in the first or second embodiment is applied to an adaptive beamformer that extracts only a target signal coming from a specific direction while suppressing interference noise will be described with reference to FIG. explain.

  FIG. 7 shows a configuration called generalized sidelobe cancellation (GSC), which is a general example of an adaptive beamformer. The target signal is removed from an input signal s of a plurality of channels (M channels), and the M-1 channel. It comprises a blocking filter 41 that outputs a signal, the adaptive filter 42 described in the first or second embodiment, and an adder 43 that adds an M-channel input signal s.

  Since M−1 = 1 when M = 2, the one-channel input adaptive filter described in the first embodiment is used as the adaptive filter 41, and the second implementation is performed as the adaptive filter 41 when M> 2. The multi-channel input adaptive filter described in the embodiment is used. Here, the description will be made assuming that M> 2, but when M = 2, the number of channels is simply reduced and the same processing is simply performed.

  There are various types of GSC depending on the configuration of the blocking filter for removing the target signal. Here, the simplest Griffith-Jim type GSC will be described as an example. Since the adaptive filter of the present invention does not depend on the method of the blocking filter, it can also be applied to GSCs other than this Griffith-Jim type.

The Griffith-Jim type GSC blocking filter is a filter that takes the difference between adjacent channels of the M channel input signal as shown in FIG. 8, and when the target signal is input to all channels in the same phase, The signal from which the signal has been deleted is output for M-1 channels. The output signal of the mth channel blocking filter is time n, and the mth channel input signal sm _{, n} .
x _{m, n} = s _{m + 1, n} −s _{m, n}
(M = 1,..., M−1) (11)
It is.

When the output signal of the blocking filter 41 is input as the reference signal x to the adaptive filter 42 described in the second embodiment and the filter is updated, the update arithmetic expression represents the filter coefficient of the adaptive filter 42 as W _n = (w ₁ , ₁ , w ₁ , ₂ ,..., W _{1, K} ,..., W _M−1,1 ,..., W _{M−1, K} ) ^T , and the reference signal vector X _n = (x _{1, n} , x _{1, n−1} ,..., X _{1, n−K + 1} ,..., XM _{, n} , xM _{, n−1} , ..., xM _{, n−K + 1} ) ^T
y _n = Σx _{M, n} (12)
^{_{e n = y n -W n T}} X n (13)
When _{P o} ≦ P _{a
W n + 1 = W n -2μe} n X n P a (14)
When P _o > P _{a
W n + 1 = W n -2μe} n X n P a / P o (15)
It is expressed as

  When such an adaptive beamformer is applied to, for example, the above-described microphone array processing apparatus and speech is extracted as a target signal, only the blocking filter 41 is caused by causes such as voice propagation fluctuations and the sound source not being a complete point sound source. Therefore, it is difficult to completely remove the target signal.

  On the other hand, if the adaptive filter described in the first or second embodiment is arranged as the adaptive filter 42 after the blocking filter 41, the adaptive filter can generate the target signal when the target signal is input as described above. Since it is possible to reduce distortion and prevent adaptation noise from being applied to background noise when there is no interference noise and target signal, the performance can be further improved.

That is, when the output signal power P _o in the adaptive filter 42 in the present embodiment is less than or equal to the threshold P _a performs filter updated by equation (14), when the output signal P _o is greater than the threshold P _a formula The adaptive mode is switched so that the filter is updated by (15). Since equation (14) LMS method of the step size fixed, less adaptation to low background noise of power, also does not stop completely adaptation can adapt for following interference noise threshold P _a.

  On the other hand, in the update by Expression (15) when the output signal power is large, the distortion of the target signal can be reduced because the step size is inversely proportional to the output signal power.

  Next, the flow of processing of the adaptive beamformer according to the present embodiment will be described with reference to the flowchart shown in FIG.

First, step size μ, adaptive mode switching threshold value P _a , number of channels M, and filter length K are set as initial settings, for example, all the filter coefficients are set to 0, and the number n of filter updates is set to 0 (step S21). ).

Next, the difference signal x _{m, n} between the input channels is calculated by the equation (11) (step S22).

Then, the sum _{y n} of the input signal calculated by equation (12) (step S23).

Then, the difference signal _{x m,} constitutes a reference signal vector _{X n} from _n, the error signal _{e n} is calculated by the equation (13) (step S24).

Next, calculate the power _{P o} of the error signal _{e n,} comparing the _{P o} and the threshold _{P a} in step S25, in the case of _{P o} ≦ _{P a} flow proceeds to step _S26, if the P o> _{P a} Advances to step S27.

  In step S26, the filter is updated according to equation (12), and the process proceeds to step S28. In step S27, the filter is updated according to equation (13), and the process proceeds to step S28. Next, in step S28, n is incremented and the process returns to step S22. The above processing is repeated for the number of input data.

  As described above, in this embodiment, by applying the adaptive filter of the present invention to an adaptive beamformer, particularly GSC, it is possible to greatly improve the noise suppression performance when a non-stationary signal such as speech is input. .

In the present embodiment, the effective step size in the case of P _o> P _a as shown in equation (15), but was inversely proportional to μ * P a _/ P _o as the output signal power P _o, necessarily accurately inversely proportional It is not necessary to reduce the output signal power _Po . In addition, instead of reducing the effective step size in the case of P _o > P _a only depending on the increase in the output signal power P _o , Equation (2), which is a filter characteristic update calculation formula in the prior art, is used. As shown, it may be decreased with respect to an increase in the sum of the reference signal power _Pr and the output signal power _Po .

(Fourth embodiment)
Next, an embodiment in which the adaptive filter of the present invention is applied to a frosted beamformer that minimizes power under constraint conditions will be described.

  The frosted beamformer of the present embodiment has the configuration shown in FIG. 10, and after filtering operations are performed on the input signal x of a plurality of channels (M channels) by the unit filter units 51-1 to 51-M, the output signals thereof are output. The filter unit 51 calculates and outputs the sum y by the adder 56, and the filter update unit 53 updates the characteristic (filter coefficient) of the filter unit 51 based on the output signal y. The filter unit 51 is configured by, for example, a multi-channel delay line tap filter shown in FIG. The number of channels is M and the number of taps is K. The filter update unit 53 includes an adaptive mode control unit 54, a filter update calculation unit 55, and a constraint condition setting unit 57.

  In the filter update unit 53, a constraint condition for constraining the characteristic change range of the filter unit 51 is set in the constraint condition setting unit 55 prior to processing, and an input is performed under the conditions set by the constraint condition setting unit 57. Based on the signal x, the filter update calculation unit 55 performs update calculation of the filter unit 51. In this filter update calculation, the adaptive mode control unit 54 controls the adaptive speed based on the power of the output signal y.

  Here, the processing of the frost type beam former will be described. The frost type beam former is described in detail in, for example, Reference 4: OLFrost, III, “An Algorithm for Linearly Constrained Adaptive Array Processing”, Proceeding of The IEEE, Vol. 60, No. 8, pp. 926-935 (1972). As described, the filter update is performed according to the following equation using the LMS algorithm.

^{_{y n = W n T X n}} (16)
_{W n + 1 = P [W} n -μy n] + F (17)
Where n is the time, W _n = (w _1,1 , w _1,2 ,..., W _{1, K} ,..., W _M−1,1 ,..., W _{M−1, K} ) ^T is the filter coefficient, _Xn = (x1 _{, n} , x1 _{, n-1} , ..., x1 _{, n-K + 1} , ..., xM _{, n} , xM _{, n-1} , ..., xM _{, n-K + 1} ) ^T is reference signal vector, P is a projection matrix onto the subspace defined by constraint, F is the translation vector from subspace to satisfying the constraint conditions space, mu is a step size, y _n is the filter output.

Projection matrix P and translation vector F are
P = I−A (A ^T A) ⁻¹ A ^T (18)
F = A (A ^T A) ⁻¹ G (19)
Thus, it is obtained before the adaptive processing. Here, A and G is a matrix that defines the constraint condition of the filter, for example, the number of constraints and J, A is _{A = [a 1, a 2} , ..., a j, ..., a J], G Is a vector with a value of 1 for the first J / 2 + 1 and 0 for the other elements. a _j is, for example,
i = K * (m−1) + (K / 2 + 1) − (J / 2 + 1) + j (20)
(M = 1, 2,..., M)
The value of the element of the number i calculated in (1) is 1, and the others are 0 vectors. In the above formula, division is rounded down.

In the present embodiment, the update calculation formulas (16) and (17) of the frost type beam former are changed, and the filter is updated based on the following formula.
^{_{y n = W n T X n}} (21)
When _{P o} ≦ P _{a
W n + 1 = P [W} n -μy n P a] + F (22)
When P _o > P _{a
W n + 1 = P [W} n -μy n P a / P o] + F (23)
Here, _{P o} is the power of the output signal _{y n,} the _{P a} is a predetermined threshold. As in the previous embodiment, LMS method step size fixed case and so as to control switching of the adaptation speed, the output signal power P _o is less than or equal to the threshold in response to the output signal power P _o, NLMS method greater Switch to.

  Thus, as in the case of GSC, when the target signal is input, the adaptive speed is slowed down by the NLMS method of Equation (21) to suppress distortion of the target signal, and the target signal and interference noise are not input. Is less adapted to background noise by the LMS method with a fixed step size according to equation (21).

Next, the processing flow of this embodiment will be described with reference to FIG.
First, the step size by default mu, adaptive mode switching threshold P _a, and set the number of channels M, the filter length K, all the values of the filter, for example, 0, placing the number n of the filter update with 0 (step S31) .

  Next, the matrices A and G that define the constraint conditions are constructed, and the matrices P and F are obtained by the equations (18) and (19) (step S32).

Then configure the reference signal vector _{X n,} the filter output _{y n} is calculated by the equation (21) (step S33).

Next, calculate the power _{P o} of the filter output signal _{y n,} comparing the _{P o} and the threshold _{P a} in step S34, in the case of _{P o} ≦ _{P a} flow proceeds to step _S35, P o> _{P a} In this case, the process proceeds to step S36.

  In step S35, the filter is updated according to equation (22), and the process proceeds to step S37. In step S36, the filter is updated according to equation (23), and the process proceeds to step S37. In step S37, n is incremented and the process returns to step S33. The above processing is repeated for the number of input data.

  As described above, by applying the adaptive filter of the present invention to the frosted beamformer, it is possible to greatly improve the noise suppression performance when an unsteady signal such as speech is input.

(Fifth embodiment)
Next, as a fifth embodiment, the adaptive filter according to the present invention is used as a first adaptive filter, and this is combined with a second adaptive filter having a higher adaptive speed, and the output signal of the second adaptive filter is used. An embodiment of a signal processing apparatus in which the performance of the first adaptive filter is further improved by controlling the adaptive speed of the first adaptive filter will be described. This further improves the convergence speed of the adaptive filter of the present invention described in the first to fourth embodiments, and in particular, in a noise canceller or beamformer when a stationary signal is interference noise. It is possible to achieve both reduction in signal distortion and improvement in convergence speed.

  FIG. 12 shows the configuration of the signal processing apparatus according to the present embodiment. This signal processing apparatus is composed of a first adaptive filter 1 and a second adaptive filter 2, and the same input signal is given to both.

  The first adaptive filter 1 has the same configuration as the adaptive filter of any of the first to fourth embodiments, and performs a filter operation on the reference signal x to obtain an output signal y1, A subtractor 12-1 that subtracts the output signal of the filter unit 11-1 from the desired response d to obtain an error signal e1, and an adaptation that updates the filter coefficient of the filter unit 11-1 based on the reference signal x and the error signal e1. It comprises a mode control unit 14-1 and a filter update calculation unit 15-1. The output signal of the first adaptive filter 1 becomes the output of the signal processing device.

  The second adaptive filter 2 has a configuration of an adaptive filter based on a general NLMS algorithm, and performs a filter operation on the reference signal x to obtain an output signal y2, and a filter unit from the desired response d. The subtractor 12-2 obtains an error signal e2 by subtracting the output signal y2 of 11-2, and a filter update calculation unit 15-2 that performs a filter update calculation based on the reference signal x and the error signal e.

  Here, the adaptive speed of the second adaptive filter 2 is set higher than that of the first adaptive filter 1, and the adaptive mode control unit 14-1 performs the first adaptation based on the output signal power of the second adaptive filter 2. The adaptive speed of the adaptive filter 1 is controlled. The adaptation speed can be set to a larger value, for example, by increasing the step size value or decreasing the number of filter taps. The entire processing is performed by the following equation, where time is n, the filter coefficient of the first adaptive filter 1 is W1, and the filter coefficient of the second adaptive filter 2 is W2.

^{_{e1 n = d n -W1 n T}} X n (24)
^{_{e2 n = d n -W2 n T}} X n (25)
_{W2 n + 1 = W2 n -2μ} 2 e2 n X n / P r (26)
When P1 _o ≦ _{P a
W1 n + 1 = W1 n -2μ} 1 e1 n X n P a (27)
When P1 _o > P _{a
W1 n + 1 = W1 n -2μ} 1 e1 n X n P a
/ P1 _o * (P1 _o / P2 _o ) α (28)
_Here, e1 _n, e2 n output signals of the adaptive filters 1, _{P r} is the reference signal power, d is desired response, μ _1, μ ₂ are each the step size of the adaptive filters 1. The value of μ ₂ is larger than P _a * μ ₁ , for example, P _a * μ ₁ = 0.000003 and μ ₂ = 0.2. Α in the equation (28) is an acceleration coefficient of the adaptive speed, and is set to 1.5, for example.

The process of the adaptive filter 2 in Expression (26) is a filter update calculation expression in a general NLMS method, and the adaptation is fast because the step size μ ₂ is larger than the step size μ ₁ of the adaptive filter 1. Therefore, in the second adaptive filter 2, although the distortion of the target signal is large, the interference noise can be quickly removed.

  On the other hand, since the adaptive speed of the first adaptive filter 1 is relatively slow, when the interference noise is input, the adaptive speed is determined by the power of the interference noise that is output without being removed before the initial convergence. Becomes slower.

Therefore, as shown in Expression (28), when the filter of the first adaptive filter 1 is updated according to the ratio P1 _o / P2 _o of the output signal power of the first adaptive filter 1 and the second adaptive filter 2. Try to increase the change of. When interference noise is input, the output signal power P2 _o of the second adaptive filter 2 that converges _quickly is smaller than the output signal power P1 _o of the first adaptive filter 1, so the power ratio P1 _o / P2 _o is 1. It becomes larger and the adaptation of the first low past filter 1 becomes faster. On the other hand, when the target signal is input, the power ratio P1 _o / P2 _o becomes a value close to 1, so that the adaptive speed of the first adaptive filter 1 remains low, and the distortion reduction of the input signal is maintained.

Next, the processing flow of this embodiment will be described with reference to FIG.
First, the step size mu ₁ by _default, mu _2, the threshold P _a of the adaptive mode _switching, the acceleration factor of the adaptive rate alpha, set the filter length K, all the values of the filter coefficients, for example, 0, the time n 0 (Step S41).

Next, a reference signal vector X _n is constructed from the reference signal x _n , the error signals e1 _n and e2 _n are calculated by the equations (24) and (25), and the powers P1 _o and e2 _n of the error signals e1 _n and e2 _n are calculated. P2 _o is calculated (step S42).

Next, the reference signal power _Pr is obtained from the reference signal vector _Xn , and the adaptive filter 2 is updated according to the equation (26) (step S43).

Next, compare the power P1 _o and the threshold _{P a} of the error signal e1 _n at step S44, in the case of P1 _o ≦ _{P a} flow proceeds to step _S45, in the case of P1 o> _{P a} flow proceeds to step S46.

  In step S45, the filter is updated according to equation (27), and the process proceeds to step S47. In step S46, the filter is updated according to equation (28), and the process proceeds to step S47. In step S47, n is incremented and the process returns to step S42. The above processing is repeated for the number of input data.

  As described above, in the present embodiment, the adaptive filter according to the present invention, in which the adaptive speed is switched and controlled based on the output signal power, is used as the first adaptive filter 1, and the second adaptive filter in which the adaptive speed is faster than this. By changing the adaptive speed of the first adaptive filter 1 in accordance with the output signal of 2, the convergence speed can be improved while reducing the distortion of the target signal.

  Thereby, in the adaptive noise canceller and the adaptive beamformer, it is possible to greatly improve the interference noise suppression performance particularly when stationary interference noise is input. The adaptive beamformer in this embodiment can be applied to both the GSC type described in the third embodiment and the frost type described in the fourth embodiment.

  In the present embodiment, the case where the reference signal is one channel has been described. However, it is also possible to configure to handle a plurality of channels of reference signals as in the second embodiment.

The block diagram which shows the structure of the adaptive filter which concerns on the 1st Embodiment of this invention. Diagram for explaining the principle of adaptive noise canceller The figure for comparing and explaining the present invention and the conventional filter update operation The flowchart which shows the flow of the adaptive filter process in 1st Embodiment. The block diagram which shows the structure of the adaptive filter made multi-channel based on the 2nd Embodiment of this invention Explanatory drawing of the multi-channel filter part in 2nd Embodiment The block diagram which shows the structure of the adaptive beamformer which is the signal processing apparatus which concerns on the 3rd Embodiment of this invention. The figure which shows the structure of an example of the blocking filter used by 3rd Embodiment. The flowchart which shows the flow of the process in 3rd Embodiment. The block diagram which shows the structure of the frost type | mold beam former which is a signal processing apparatus concerning the 4th Embodiment of this invention. The flowchart which shows the flow of the process in 4th Embodiment. The block diagram which shows the structure of the signal processing apparatus which concerns on the 5th Embodiment of this invention. The flowchart which shows the flow of the process in 5th Embodiment. Block diagram showing the configuration of a conventional adaptive filter

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... 1st adaptive filter 2 ... 2nd adaptive filter 11, 11-1, 11-2 ... Filter part 12, 12-1, 12-2 ... Subtractor 13 ... Filter update part 14, 14-1 ... Adaptive Mode control unit 15, 15-1, 15-2 ... Filter update calculation unit 21 ... Filter unit 21-1 to 21-M ... Unit filter unit 22 ... Subtractor 23 ... Filter update unit 24 ... Adaptive mode control unit 25 ... Filter Update calculation unit 26 ... adder 41 ... blocking filter 42 ... adaptive filter 51 ... filter unit 51-1 to 51-M ... unit filter unit 52 ... subtractor 53 ... filter update unit 54 ... adaptive mode control unit 55 ... filter update calculation 56: Adder 57: Restriction condition setting unit

Claims (1)

A first filter unit that performs a filter operation on the reference signal to obtain a first output signal; a first subtractor that subtracts the first output signal from a desired response to obtain a first error signal; A first filter update calculation unit that updates a filter coefficient of the first filter unit based on the first effective step size, the reference signal, and the first error signal, and the first effective step size . A first adaptive filter including an adaptive mode control unit for controlling;
A second filter unit that performs a filter operation on the reference signal to obtain a second output signal; and a second subtraction that subtracts the second output signal from the desired response to obtain a second error signal. And a second adaptive filter comprising a second effective step size, a second filter update operation unit that updates a filter coefficient of the second filter unit based on the reference signal and the second error signal And
The second adaptive filter is set to have a higher adaptive speed than the first adaptive filter,
The adaptive mode control unit calculates a first power of the first error signal and a second power of the second error signal, compares the first power with a predetermined threshold, and When the first power is less than or equal to the threshold value, the first effective step size is controlled to be a non-zero constant value, and when the first power exceeds the threshold value, the first effective step size is controlled. A signal processing device, wherein the size is controlled to become a value that increases as a ratio of the first power to the second power increases .

Images