WO2014097470A1

WO2014097470A1 - Reverberation removal device

Info

Publication number: WO2014097470A1
Application number: PCT/JP2012/083266
Authority: WO
Inventors: 博秋河崎; 健作藤井
Original assignee: Toa株式会社
Priority date: 2012-12-21
Filing date: 2012-12-21
Publication date: 2014-06-26

Abstract

[Problem] To achieve accurate reverberation removal even while being of a simple configuration. [Solution] According to this reverberation removal device (10), by way of input of a collected sound signal X(z) from a microphone (20) into a first linear prediction error filter (50), the original sound component S(z) is removed from the collected sound signal X(z). Then, on the basis of a post-processing signal X'(z) by way of the first linear prediction error filter (50), a second linear prediction error filter (60) identifies an inverse transfer function 1/R(z) which is a reciprocal of a transfer function R(z) of an acoustic system (40) from an audio source (30) to the microphone (20). Then, the inverse transfer function 1/R(z) of the acoustic system (40) which has been identified by the second linear prediction error filter (60) is copied to a reverberation removal filter (70). By inputting the collected sound signal X(z) from the microphone (20) into the reverberation removal filter (70), reverberation components by the acoustic system (40) are removed from the collected sound signal X(z) so that only the original sound component S(z) is extracted.

Description

Reverberation removal device

The present invention relates to a dereverberation apparatus, and in particular, to remove a dereverberation component by an acoustic system from a collected sound signal that is an output signal of a sound collecting means that captures sound from a sound source via an acoustic system having reverberation characteristics. Relates to the device.

Conventionally, as this type of dereverberation device, for example, there is one disclosed in Patent Document 1. According to this prior art, first, the reverberation signal after the sound source signal passes through the reverberation path is divided into short time sections. Then, a linear prediction coefficient (Linier Predictive Coefficient: LPC) is obtained for each section. Further, a residual signal is obtained by convolving the linear prediction coefficient with the reverberant signal. Then, the pitch frequency is further removed from the residual signal. On the other hand, an acoustic transfer characteristic of an acoustic system including a reverberation path (for example, the same room) is measured in advance, and its outline is obtained. And the outline of this acoustic transfer characteristic is convolved with the residual signal after pitch frequency removal, whereby the residual signal is corrected. After that, an AR (Autoregressive) coefficient in a long time section is obtained from the corrected residual signal. The AR coefficient is set in the FIR filter as a dereverberation filter coefficient, so that the FIR filter functions as a dereverberation filter. That is, when a reverberation signal is input to the FIR filter as the reverberation suppression filter, a reverberation component included in the reverberation signal is suppressed.

JP-A-9-261133

However, in the above-described conventional technique, it is necessary to measure the acoustic transfer characteristics of the acoustic system in advance, and thus extra time and effort are required for that purpose. In addition, since the acoustic transfer characteristic measured in advance is not that of the speaker's utterance position (that is, because the reverberation signal that is the target of reverberation suppression has a different path), Reverberation suppression using a simple acoustic transmission characteristic is inevitably lacking in accuracy.

Therefore, the present invention has an object to provide a dereverberation apparatus that does not need to measure acoustic transfer characteristics in advance unlike the prior art, and that can remove dereverberation more accurately than the prior art. And

In order to achieve this object, the present invention provides a reverberation removal that removes a reverberation component by an acoustic system from a collected sound signal that is an output signal of a sound collecting means that captures sound from a sound source via an acoustic system having reverberation characteristics. The device is assumed. Under this assumption, the original sound component removing means for removing the original sound component which is the original sound component (from the sound source) from the collected sound signal, and after the original sound component is removed from the collected sound signal by the original sound component removing means Inverse transfer function computing means for obtaining the inverse of the transfer function of the acoustic system based on the original sound component removed signal, and signal processing means for performing processing based on the inverse transfer function obtained by the inverse transfer function computing means on the collected sound signal And.

According to the present invention configured as described above, in addition to the original sound component that is a component of the original sound emitted from the sound source, the original sound component is repeatedly reflected in the collected sound signal that is the output signal of the sound collecting means. Reverberation components (specifically, various noise components in addition to this) are included. In order to remove the reverberation component from such a collected sound signal, the original sound component removing means first removes the original sound component from the collected sound signal, that is, separates it. Then, based on the original sound component removal signal after the original sound component is removed (separated) from the collected sound signal, in other words, the original sound component removal signal including only the reverberation component, the reciprocal of the transfer function of the acoustic system, That is, the inverse transfer function is obtained by the inverse transfer function calculating means. By taking such a procedure, the inverse transfer function of the acoustic system can be obtained accurately without being affected by the original sound component. And a signal processing means performs the process based on the reverse transfer function of this acoustic system with respect to a sound collection signal. As a result, only the reverberation component is accurately removed from the collected sound signal, and only the original sound component is accurately reproduced (reproduced).

That is, according to the present invention, the inverse transfer function of the acoustic system is obtained based only on the collected sound signal by the sound collecting means, and thus the reverberation component is removed, so-called blind dereverberation is realized. Although this blind dereverberation is generally considered difficult, according to the present invention, the original sound component is removed from the collected sound signal as described above, and the acoustic system is based on the signal after the removal of the original sound component after the removal. Thus, the inverse transfer function of the acoustic system can be accurately obtained without being affected by the original sound component, and thus accurate dereverberation can be realized. In particular, accurate dereverberation can be achieved as compared to the above-described conventional technique in which dereverberation is performed using an acoustic transfer function that is not at the speaker's utterance position. Moreover, unlike the prior art, it is not necessary to measure the acoustic transfer function in advance.

The original sound component removing means in the present invention may include a first adaptive filter to which a sound collection signal from the sound collection means is input. The inverse transfer function calculation means may include a second adaptive filter to which a signal processed by the first adaptive filter is input. Furthermore, the signal processing means may include a third adaptive filter to which a sound collection signal from the sound collection means is input. In this case, the first adaptive filter has a relatively high responsiveness that can follow the fluctuation of the original sound component, and the level of the signal processed by itself or the signal processed by the second adaptive filter is minimized. To work. By operating the first adaptive filter in this manner, the transfer function of the first adaptive filter is the inverse characteristic of the original sound component (strictly speaking, for example, a transfer function for generating the original sound component like a vocal tract transfer function). Is almost equivalent to That is, the reverse characteristic of the original sound component is identified by the first adaptive filter, that is, an inverse filter of the original sound component is formed. Therefore, when the collected sound signal is input to the first adaptive filter, the original sound component is removed from the collected sound signal. Then, the second adaptive filter has a relatively low response that is incapable of following the fluctuation of the original sound component, and operates so that the level of the processed signal by itself is minimized. As the second filter operates in this way, the transfer function of the second adaptive filter becomes substantially equivalent to the inverse of the transfer function of the acoustic system. That is, the inverse transfer function of the acoustic system is identified by the second adaptive filter, that is, the inverse filter of the acoustic system is formed. Further, the third adaptive filter realizes the processing based on the above-described inverse transfer function of the acoustic system by copying the filter coefficient of the second adaptive filter to the third adaptive filter, so-called copying. That is, by copying the filter coefficient of the second adaptive filter to the third adaptive filter, the transfer function of the third adaptive filter becomes substantially equivalent to the inverse transfer function of the acoustic system, that is, the inverse filter of the acoustic system is It is formed. Then, when the collected sound signal is input to such a third adaptive filter, the reverberation component is removed from the collected sound signal.

The first adaptive filter referred to here has a relatively small number of taps (tap length). By setting the number of taps to be relatively small in this way, the first adaptive filter has a relatively high response. The second adaptive filter has a smaller number of taps than the number of taps of the first adaptive filter. As a result, the second adaptive filter has a lower responsiveness than the responsiveness of the first adaptive filter. The number of taps of the third adaptive filter is naturally the same as the number of taps of the second adaptive filter.

Also, the step size for adjusting the update amount of the first adaptive filter is set to be relatively large. This also makes the first adaptive filter have high responsiveness. The step size of the second adaptive filter is set to be smaller than the step size of the first adaptive filter. This also causes the second adaptive filter to have a response lower than the response of the first adaptive filter. The responsiveness of the third adaptive filter is equivalent to the responsiveness of the second adaptive filter.

Such a first adaptive filter is preferably a linear prediction error filter, for example. Similarly, the second adaptive filter is desirably a linear prediction error filter. That is, it is known that the linear prediction error filter is suitable for forming an inverse filter, and that the processing (algorithm) basically does not cause a delay. Therefore, by adopting such a linear prediction error filter as a first adaptive filter (for forming an inverse filter of the original sound component) and a second adaptive filter (for forming an inverse filter of the acoustic system), Real-time performance is realized. In addition, it is also known that the linear prediction error filter has a relatively small amount of calculation. Therefore, it is possible to reduce the burden on the CPU (Central Processing Unit) and DSP (Digital Signal Processor) that constitute the linear prediction error filter. In other words, as the CPU, DSP, etc., it is possible to adopt an inexpensive device whose capability is not necessarily high. Note that the third adaptive filter is a linear prediction error filter configured to conform to the second adaptive filter.

It is a block diagram which shows schematic structure of one Embodiment of this invention. It is a block diagram which shows the specific structure of the 1st linear prediction error filter in the embodiment. It is a block diagram which shows the specific structure of the 2nd linear prediction error filter in the embodiment. It is a figure which shows the experimental result in the same embodiment. It is a figure which expands and shows a part of FIG. It is a block diagram which shows equivalently the transfer function of the acoustic system in the embodiment. It is a figure which shows the experimental result different from FIG. 4 and FIG.

An embodiment of the present invention will be described with reference to FIGS.

The dereverberation apparatus 10 according to the present embodiment includes a microphone 20 as sound collection means as shown in FIG. The microphone 20 captures a sound emitted from the sound source 30, a so-called original sound S (z) (z; a variable in z conversion) via an acoustic system (acoustic space) 40 having reverberation characteristics. The collected sound signal X (z) that is an output signal of the microphone 20 is input to a first linear prediction error filter (Linear Prediction Error LP; LPEF) 50 as a first adaptive filter.

The first linear prediction error filter 50 is for removing only the component of the original sound S (z) from the collected sound signal X (z) input from the microphone 20. Details of the first linear prediction error filter 50 will be described later. Then, the signal X ′ (z) processed by the first linear prediction error filter 50 is input to a second linear prediction error filter 60 as a second adaptive filter.

The second linear prediction error filter 60 is based on the signal X ′ (z) processed by the first linear prediction error filter 50, in short, of the collected sound signal X (z) obtained from the microphone 20. Based on the processed signal X ′ (z) after the component is removed, the inverse transfer function 1 / R (z) that is the inverse of the transfer function R (z) of the acoustic system 40 from the sound source 30 to the microphone 20. Is determined, that is, identified. Then, the inverse transfer function 1 / R (z) of the acoustic system 40 identified by the second linear prediction error filter 60, strictly speaking, a filter coefficient a _n (n: tap) described later of the second linear prediction error filter 60 will be described. ) Is copied to the dereverberation filter 70 as the third adaptive filter. Details of the second linear prediction error filter 60 will be described later in detail.

On the other hand, the collected sound signal X (z) is input to the dereverberation filter 70 from the microphone 20. The dereverberation filter 70 performs a filtering process according to the above-described inverse transfer function 1 / R (z) of the acoustic system 40 on the collected sound signal X (z), so that the acoustic system 40 is obtained from the collected sound signal X (z). The noise component due to, especially the reverberation component, is removed. As a result, the signal from which the reverberation component has been removed, in other words, the signal from which only the component of the original sound S (z) has been extracted is used as the post-processing signal S ′ (z) by the dereverberation filter 70. 70, and thus the output of the entire dereverberation apparatus 10 according to the present embodiment. Details of the dereverberation filter 70 will also be described later in detail.

The first linear prediction error filter 50 removes the component of the original sound S (z) from the collected sound signal X (z) from the microphone 20 as described above. The level of the signal E (z) after processing by the filter 60, that is, a so-called prediction error is operated to be minimized. Specifically, the transfer function H _S (z) of the first linear prediction error filter 50 is expressed by the following equation (1).

In this equation 1, M is the number of taps. The number of taps M is a relatively small value. For example, if the sampling frequency f of the collected sound signal X (z) is f = 16 kHz, M = 16 to 64 (1 ms to 4 ms in terms of time). Is done. By setting the tap number M to a relatively small value in this way, the responsiveness of the first linear prediction error filter 50 becomes high, and it becomes possible to quickly follow the fluctuation of the original sound S (z) component. Become. In Equation 1, b _m is a filter coefficient of the m-th tap. By this filter coefficient b _m are updated accordingly, i.e. above the prediction error E (z) is that is the filter coefficients b _m so as to minimize is updated, the transmission of the first linear prediction error filter 50 The function H _S (z) is substantially equivalent to the inverse characteristic 1 / _S (z) of the original sound S (z) (H _S (z) ≈1 / S (z)). The filter coefficient b _m is updated by, for example, a known learning identification method (Normalized Least Mean Square: NLMS). The step size μ _S for adjusting the update amount of the filter coefficient b _m is set to a relatively large value, for example, μ _S = about 0.1 to 0.5. Thus also by the step size mu _S is a relatively large value, the response of the first linear prediction error filter 50 becomes high. Of course, an appropriate algorithm other than the learning identification method may be employed as an algorithm for updating the filter coefficient b _m .

Such a first linear prediction error filter 50 has a transversal configuration as shown in FIG. 2, for example. According to this configuration, the adder 502 to which the sound collection signal X (z) from the microphone 20 is input is provided. In addition, the collected sound signal X (z) is sequentially delayed by the

M delay elements

504, 504,. The outputs of the

delay elements

504, 504,... Are multiplied by the corresponding filter coefficients b _m by the corresponding

multipliers

506, 506,. Are summed by

adders

508, 508,. The sum P (z) by these

adders

508, 508,... Is a prediction signal that uses the collected sound signal X (z) as a target signal, and this prediction signal P (z) is input to the adder 502 described above. Here, it is subtracted from the collected sound signal X (z) as the target signal. A signal after subtraction by the adder 502, that is, a residual signal that is a difference between the sound pickup signal X (z) as a target signal and the prediction signal P (z) is processed by the first linear prediction error filter 50. It is output as a rear signal X ′ (z). Thereby, the first linear prediction error filter 50 having the transfer function H _S (z) expressed by Equation 1 is realized.

The transfer function H _S (z) of the first linear prediction error filter 50 is substantially equivalent to the inverse characteristic 1 / S (z) of the original sound S (z) as described above. That is, the first linear prediction error filter 50 identifies an inverse characteristic of the original sound S (z), and so forms an inverse filter of the original sound S (z). Therefore, when the collected sound signal X (z) is input to the first linear prediction error filter 50, the component of the original sound S (z) is removed from the collected sound signal X (z), and almost only the reverberation component is present. Remain. Then, the processed signal X ′ (z) by the first linear prediction error filter 50 including only substantially the reverberation component is input to the second linear prediction error filter 60 as described above.

The second linear prediction error filter 60 identifies the inverse transfer function 1 / R (z) of the acoustic system 40 as described above. The second linear prediction error filter 60 is also the first linear prediction error filter. As in the case of 50, the operation is performed so that the prediction error E (z) is minimized. Specifically, the second linear prediction error filter 60 is similar to the transfer function H _S (z) of the first linear prediction error filter 50 expressed by Equation 1, and the transfer function H expressed by Equation 2 below. _R (z).

In Equation 2, N is the number of taps. The tap number N is a relatively large value. For example, if the above sampling frequency f is f = 16 kHz, N = 1024 to 2048 (64 ms to 128 ms in terms of time). Thus, by setting the tap number N to a relatively large value, the responsiveness of the second linear prediction error filter 60 becomes low, and it is impossible to follow the fluctuation of the original sound S (z) component, for example. On the other hand, it is suitable for accurately identifying the inverse transfer function 1 / R (z) of the acoustic system 40. Then, in the number 2, a _n is the filter coefficient of the n taps th. By this filter coefficients a _n are updated accordingly, i.e. the prediction error E (z) is that is the filter coefficient a _n so as to minimize the update, the transfer function H of the second linear prediction error filter 60 _R (z) is substantially equivalent to the inverse transfer function 1 / _R (z) of the acoustic system 40 (H _R (z) ≈1 / R (z)), that is, the inverse transfer function 1 / R of the acoustic system 40. (Z) is identified. Incidentally, the filter coefficients a _n are also, for example, is updated by the above-mentioned learning identification method. Then, the step size mu _R for adjusting the amount of update of the filter coefficient _{a n} is a relatively small value, are for example, _μ R = 0.001 ~ 0.01. Thus also by the step size mu _R is a relatively small value, the response of the second linear prediction error filter 60 is low. The updating algorithm of the filter coefficients a _n of the second linear prediction error filter 60 may, as appropriate algorithms except learning identification method may be employed.

Such a second linear prediction error filter 60 has, for example, a transversal configuration as shown in FIG. 3 similar to the configuration of the first linear prediction error filter 50 shown in FIG. That is, according to this configuration, the adder 602 to which the signal X ′ (z) processed by the first linear prediction error filter 50 is input is provided. In addition, the processed signal X ′ (z) is sequentially delayed by

M delay elements

604, 604,. Then, each of the

delay elements

604, 604, ... are output, the multiplier 606,606 corresponding to each, after ... have multiplied the filter coefficients _{a n} corresponding to the respective through another (N-1 pieces as above Are summed by

adders

608, 608,. The sum P ′ (z) by the

adders

608, 608,... Is a prediction signal having the processed signal X ′ (z) by the first linear prediction error filter 50 as a target signal, and this prediction signal P ′ ( z) is input to the above-described adder 602, where it is subtracted from the processed signal X ′ (z) as the target signal. Then, a signal after subtraction by the adder 602, that is, a residual signal that is a difference between the processed signal X ′ (z) as the target signal and the prediction signal P ′ (z) is the second linear prediction error filter 60. Is output as a prediction error E (z). Thereby, the second linear prediction error filter 60 having the transfer function H _R (z) expressed by Equation 2 is realized.

The second linear prediction error filter 60 identifies the inverse transfer function 1 / R (z) of the acoustic system 40 as described above, that is, forms the inverse filter of the acoustic system 40. Then, the transfer function H _R of the second linear prediction error filter 60 _(z), the filter coefficients a _n of strictly said second linear prediction error filter 60 is copied to the dereverberation filter 70 as described above. The second filter coefficient a _n copies from the linear prediction error filter 60 to dereverberation filter 70, for example, every 1 sampling, i.e. at a constant cycle, is carried out. Incidentally, the copy may be performed for each of the plurality sampling, irregularly, for example when the filter coefficients a _n occurs above a certain variation, it may be performed.

The dereverberation filter 70 basically has the same configuration as that of the second linear prediction error filter 60 shown in FIG. 3 (except that the dereverberation operation itself does not perform an adaptive operation), that is, its transfer function H _F (z) is the above-mentioned. This is equivalent to the transfer function H _R (z) of the second linear prediction error filter 60 expressed by Equation 2 (H _F (z) = H _R (z)). Therefore, when the collected sound signal X (z) from the microphone 20 is input to such a dereverberation filter 70, the reverberation component is removed from the collected sound signal X (z), and the component of the original sound S (z). Only is taken out. The signal S ′ (z) (≈S (z)) after processing by the dereverberation filter 70 is the output of the entire dereverberation apparatus 10 according to the present embodiment.

The dereverberation effect by the dereverberation apparatus 10 according to this embodiment was actually confirmed. Specifically, in a conference room having a width of about 5.8 m, a depth of about 3.2 m, and a height of about 2.7 m, white noise as the original sound S (z) is output from a speaker as the sound source 30; This is received by the microphone 20 separated from the speaker by about 1.0 m. Then, by observing the collected sound signal X (z) by the microphone 20, the impulse response before dereverberation is confirmed, and the signal after processing by the dereverberation filter 70 (the output of the entire dereverberation apparatus 10) is observed. Then, I confirmed the impulse response after dereverberation. Note that the sampling frequency f at this time is f = 16 kHz. The number of taps M of the first linear prediction error filter 50 is M = 16, and the step size μ _S is μ _S = 0.2. Moreover, the tap number N of second linear prediction error filter 60 is N = 2048, the step size mu _R is a mu _R = 0.004. The configuration of the dereverberation filter 70 conforms to the configuration of the second linear prediction error filter 60.

The results shown in FIGS. 4 and 5 were obtained by this experiment. 4A shows an impulse response (X (z)) before dereverberation, and FIG. 4B shows an impulse response (S ′ (z)) after dereverberation. FIG. 5A is an enlarged view of a portion surrounded by a broken line frame A in FIG. 4A, and FIG. 5B is surrounded by a broken line frame B in FIG. 4B. It is the figure which expanded the part. 4 and 5, the component indicating the maximum amplitude (when time is zero) is the component of the original sound (z), which is a so-called direct sound component. A component behind the direct sound component is a reverberation component.

As is apparent from FIGS. 4 and 5, according to the dereverberation apparatus 10 of the present embodiment, the original sound S (z) component (the amplitude thereof) which is a direct sound component is the same as that before dereverberation, that is, nothing. It is not influenced by. And it turns out that only the reverberation component is reduced more effectively than before reverberation removal. Similarly, it was confirmed that a good dereverberation effect was obtained when human speech was adopted as the original sound S (z).

In particular, according to the dereverberation apparatus 10 of the present embodiment, a very good dereverberation effect can be obtained in an environment where flutter echoes (Naruto) occur. This is due to the following reason.

That is, the transfer function R (z) of the acoustic system 40 is expressed by the following mathematical formula 3. In Equation 3, K is the number of taps. Α _k is a k-th tap filter coefficient, and β _k is another k-tap filter coefficient.

The denominator in Equation 3 corresponds to the minimum phase component, and a typical example is a flutter echo. On the other hand, the numerator in Equation 3 corresponds to a non-minimum phase component, and a typical example thereof includes an irregular reflection echo. Then, the second linear prediction error filter represented by the above equation 2 for identifying the transfer function R (z) of the acoustic system 40 represented by the equation 3 and its inverse transfer function 1 / R (z). When the transfer function H _R (z) of 60 is compared, the minimum phase component (the denominator in Equation 3) of the inverse transfer function 1 / R (z) of the acoustic system 40 and the second linear prediction error filter 60 Transfer functions H _R (z) correspond to each other perfectly. This means that the minimum phase component is reliably identified by the second linear prediction error filter 60. Therefore, the minimum phase component including the flutter echo is removed very well. Note that the non-minimum phase component (the numerator in Equation 3) of the inverse transfer function 1 / R (z) of the acoustic system 40 is not subjected to special processing and is ignored.

Moreover, if the transfer function R (z) of the acoustic system 40 is equivalently expressed in a block diagram, it is as shown in FIG. According to FIG. 6, the original sound S (z) from the sound source 30 is multiplied by a filter coefficient β ₀ by a multiplier 402 and then input to an adder 404. The output of the adder 404 is used as the output of the acoustic system 40, that is, input to the microphone 20. In addition, the original sound S (z) from the sound source 30 is sequentially delayed by the

K delay elements

406, 406,. The outputs of the

delay elements

406, 406,... Are multiplied by the corresponding filter coefficients β _k by the corresponding

multipliers

408, 408,. Are input to the

units

410, 410,... Further, the output of the adder 404 is sequentially delayed by K

separate delay elements

412, 412,. The outputs of the

delay elements

412, 412,... Are multiplied by the corresponding filter coefficients α _k by the corresponding

multipliers

414, 414,. It is input to…. Each adder 410 subtracts the multiplication result by the multiplier 414 from the multiplication result by the multiplier 408. Then, the subtraction results by the

adders

410, 410,... Are summed together and input to the above-described adder 402, where they are added to the original sound S (z). Thereby, an equivalent circuit of the transfer function R (z) of the acoustic system 40 expressed by Equation 3 is realized.

The configuration of the equivalent circuit of the transfer function R (z) of the acoustic system 40 shown in FIG. 6 and the second linear prediction error filter 60 (shown in FIG. 3 for identifying the inverse transfer function 1 / R (z) are shown. When the configuration of the transfer function H _R (z)) is compared, a portion corresponding to the above-described minimum phase component (the denominator in Equation 3) in the equivalent circuit of the transfer function R (z) of the acoustic system 40 (FIG. 3) 6 and the configuration of the second linear prediction error filter 60 shown in FIG. 3 are mutually opposite. Also from this, it can be understood that the minimum phase component is reliably identified by the second linear prediction error filter 60 and that the minimum phase component including the flutter echo is satisfactorily removed.

Furthermore, in order to further improve the effect of the dereverberation apparatus 10 of the present embodiment, strictly, the following device may be made in order to obtain the effect earlier.

That is, a signal X ′ (z) after processing by the first linear prediction error filter 50 that is an input signal to the second linear prediction error filter 60 and a prediction error E () that is an output signal of the second linear prediction error filter 60. and z), depending on the step size mu _R of the second linear prediction error filter 60 is to be controlled appropriately. Specifically, the step size mu _R is controlled on the basis of the following equation 4.

In Equation 4, μ ₀ is an initial value of the step size μ _R and is arbitrarily set. E ₀ is a target value of the prediction error E (z), which is also set arbitrarily. Furthermore, P _S is the power of the input signal X '(z), it is approximated by the following equation 5. P _E is the power of the output signal E (z) and is approximated by Equation (6). Note that the coefficient ρ in Equations 5 and 6 is determined as in Equation 7.

By thus step size mu _R of the second linear prediction error filter 60 is controlled, identified speed (convergence rate of the inverse transfer function 1 / R of the acoustic system 40 according to the second linear prediction error filter 60 (z) ) Will improve. The effect was actually confirmed by simulation.

Specifically, white noise is used as the original sound S (z). Then, a normal random number that exponentially decays is used as the reverberation, and the reverberation time is set to 1.024 s. Then, the initial value μ ₀ of the step size μ _R in Equation 4 is set to μ ₀ = 0.2, and the target value E ₀ of the prediction error E (z) is set to E ₀ = 0.1. Further, for the comparison, the step size mu _R is prepared that is fixed to the _mu R = 0.008. Other conditions are the same as in the previous experiment described with reference to FIGS. 4 and 5.

The result shown in FIG. 7 was obtained by this simulation. (A) in FIG. 7 of them is a comparative control of the results step size mu _R is fixed, and FIG. 7 (b) is a result of the control of the step size mu _R. As can be seen from the results, in the configuration step size mu _R is fixed, 1 minute (= 96 × 10 ⁴ sampling) as it takes time for the for identification operation by the second linear prediction error filter 60 is stabilized by the step size mu _R is controlled, time to identification operation by the second linear prediction error filter 60 is stabilized is shortened to 15 seconds (= 24 × 10 ⁴ sampling) degree. That is, by the step size mu _R of the second linear prediction error filter 60 is controlled, that the identification rate by the second linear prediction error filter 60 is improved, has been confirmed. In the experiment using the actual dereverberation apparatus 10, the same effect as the simulation was obtained.

As described above, according to the dereverberation apparatus 10 of the present embodiment, only the reverberation component of the collected sound signal X (z) from the microphone 20 can be accurately removed, and only the original sound S (z) can be accurately extracted. it can. Such a dereverberation apparatus 10 is extremely good in reverberation particularly in an environment where the flutter echo described above occurs, or in an environment where the speaker and the microphone are separated by about 1 m as in a teleconference system. Demonstrate the removal effect. This is particularly effective for devices that are susceptible to reverberation, such as hands-free telephones, intercoms, and boundary microphones.

Also, according to the dereverberation apparatus 10 of the present embodiment, unlike the above-described conventional technology, it is not necessary to measure the acoustic transfer characteristics in advance. Furthermore, according to the dereverberation apparatus 10 of the present embodiment, the inverse transfer function 1 / R (z) of the acoustic system 40 is identified based only on the sound collection signal S (z) from the microphone 20, and thus dereverberation is performed. In other words, so-called blind dereverberation is realized. This blind dereverberation is generally considered difficult, but according to the dereverberation apparatus 10 of the present embodiment, the component of the original sound S (z) from the collected signal X (z) of the microphone 20 as described above. Is removed, and the inverse transfer function 1 / R (z) of the acoustic system 40 is identified based on the processed signal X ′ (z) after the removal, so that the influence of the component of the original sound S (z) can be reduced. Without being received, the inverse transfer function 1 / R (z) of the acoustic system 40 is accurately identified, so that accurate noise removal is realized. In particular, a better dereverberation effect can be obtained as compared with the prior art lacking accuracy of dereverberation suppression. This greatly contributes to the improvement of sound quality.

In addition, according to the dereverberation apparatus 10 of the present embodiment, a good dereverberation effect can be obtained as described above by the three linear prediction error filters 50, 60 and 70 including the dereverberation filter 70. Since these linear prediction error filters 50, 60 and 70 basically do not cause a delay, a so-called real-time property is realized. Further, since the linear prediction error filters 50, 60 and 70 have a relatively small amount of calculation, the burden on the CPU and DSP (not shown) constituting them is reduced. In other words, as the CPU, DSP, etc., it is possible to adopt an inexpensive one whose capability is not necessarily high.

Note that the content described in the present embodiment is one specific example for realizing the present invention, and does not limit the scope of the present invention.

For example, in the present embodiment, the first linear prediction error filter 50 is a transversal type, but is not limited thereto. The first linear prediction error filter 50 may be realized by a configuration other than the transversal type such as a lattice type. However, since the transversal type has a simpler configuration than, for example, the lattice type, by adopting such a transversal type, the CPU, DSP, or the like constituting the first linear prediction error filter 50 is used. The burden is further reduced. In other words, a more inexpensive one can be used as the CPU or DSP.

Similarly, the second linear prediction error filter 60 is not limited to the transversal type but may be realized by a configuration other than the transversal type such as a lattice type. However, it goes without saying that the configuration of the noise removal filter 70 is determined in accordance with the configuration of the second linear prediction error filter 60.

Furthermore, instead of the first linear prediction error filter 50 as the first adaptive filter, an adaptive filter of another configuration (algorithm) may be employed. However, as described above, by employing the first linear prediction error filter 50 as the first adaptive filter, real-time performance is realized, and the burden on the CPU, DSP, and the like constituting the real-time property is reduced.

Similarly, the second linear prediction error filter 60 may be replaced with an adaptive filter having another configuration instead. Also in this case, the configuration of the dereverberation filter 70 is determined in accordance with the configuration of the second linear prediction error filter 60.

DESCRIPTION OF SYMBOLS 10 Reverberation removal apparatus 20 Microphone 30 Sound source 40 Acoustic system 50 1st linear prediction error filter 60 2nd linear prediction error filter 70 Reverberation removal filter

Claims

In the dereverberation apparatus for removing the reverberation component by the acoustic system from the collected sound signal that is the output signal of the sound collecting means that captures the sound from the sound source via the acoustic system having the reverberation characteristics
An original sound component removing means for removing an original sound component that is a component of the sound from the collected sound signal;
An inverse transfer function computing means for obtaining an inverse of the transfer function of the acoustic system based on the original sound component removed signal after the original sound component is removed from the collected sound signal by the original sound component removing means;
Signal processing means for performing processing based on the inverse transfer function obtained by the inverse transfer function calculating means on the collected sound signal;
A dereverberation apparatus comprising:
The original sound component removing means includes a first adaptive filter to which the collected sound signal is input,
The inverse transfer function calculating means includes a second adaptive filter to which a signal processed by the first adaptive filter is input,
The signal processing means includes a third adaptive filter to which the collected sound signal is input,
The first adaptive filter has such a high responsiveness as to follow the fluctuation of the original sound component and operates so that the level of the signal processed by itself or the signal processed by the second adaptive filter is minimized,
The second adaptive filter has such a low response that it cannot follow the fluctuation of the original sound component and operates so that the level of the signal after processing by itself is minimized,
The third adaptive filter realizes processing based on the inverse transfer function by copying the filter coefficient of the second adaptive filter.
The dereverberation apparatus according to claim 1.
The number of taps of the first adaptive filter is smaller than the number of taps of the second adaptive filter.
The dereverberation apparatus according to claim 2.
The step size of the first adaptive filter is larger than the step size of the second adaptive filter;
The dereverberation apparatus according to claim 2 or 3.
A part or all of the first adaptive filter, the second adaptive filter, and the third adaptive filter are linear prediction error filters.
The dereverberation apparatus according to any one of claims 2 to 4.