JPH09261133A - Reverberation suppression method and its equipment - Google Patents

Abstract

PROBLEM TO BE SOLVED: To suppress a reverberation without knowing information of an original voice of a talker and without the need for advance measurement of a sound transfer characteristic at a position of the talker. SOLUTION: The equipment divides a reverberation signal into short time blocks, a linear prediction coefficient is obtained from each block (42), the coefficient is convoluted into the reverberation signal to obtain a residual signal (43), a pitch frequency is eliminated from the residual signal (46), an outline shape of a frequency amplitude characteristic of a reverberation path in the vicinity of the reverberation path is obtained (52), the outline shape is convoluted onto the signal from which the pitch frequency is eliminated (47), an AR coefficient is obtained from the convoluted signal for a long time block (48), the reverberation voice is given to an FIR filter having the AR coefficient as a filter means to obtain the reverberation elimination signal (50).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice / acoustic apparatus, such as a TV conference apparatus, a voice conference apparatus, a voice recognition apparatus, etc., which needs to pick up a sound with reverberation removed even in a reverberant environment.
The present invention relates to a reverberation suppressing method and apparatus applied to a voice / acoustic signal processing system.

[0002]

2. Description of the Related Art Reverberation is a phenomenon in which a sound signal emitted from a sound source becomes a reverberant sound in a target acoustic system (for example, a room sound field) due to reflection of a room wall or the like. . Such reverberation enriches the sound with respect to a music signal or the like, but causes a decrease in clarity with respect to a voice signal. The effect of this reverberation is completely expressed by the acoustic transfer characteristics that represent the state of sound transmission between the sound source and the sound receiver in the room. The dereverberation method means a method of restoring the sound source signal by convolving the signal picked up by the sound receiver with the inverse characteristic of the acoustic transfer characteristic from the sound source to the sound receiver. In the present specification, it is assumed that the signal is a discrete signal, and the time representation of the signal is represented by, for example, x (k) with an integer parameter k representing time, and its frequency representation is expressed by X (z ).

FIG. 4 is a diagram for explaining how a voice generated by a speaker in a room is picked up by receiving the influence of the acoustic transfer characteristic H (z) as reverberation. 1 is indoor space, 2
Is a speaker, 3 is a sound receiver (for example, a microphone), and 4 is an output end.

Assuming that the sound signal generated by the speaker 2 is X (z), the signal X (z) reaches the sound receiver 3 under the influence of the acoustic transfer characteristic H (z). Signal Y (z) received by the sound receiver 3
Is output from the output terminal 4. Here, the signal Y (z) received by the sound receiver 3 is

[0005]

[Equation 1] Is expressed as The acoustic transfer characteristic H (z) includes all the information such as the reflection by the wall in the room. Even within the same room space, the characteristics differ if the spatial positions of the speaker 2 and the sound receiver 3 are different. Usually, a signal Y (z) in which X (z) has passed through H (z) is called a reverberation signal with reverberation or a reverberation signal.

The reverberation suppressing method is to add the reverberation generated by the speaker 2 by removing the H (z) component from the signal Y (z) received by passing through the acoustic transfer characteristic H (z). Signal X before
It means a method of restoring (z).

[Reverberation Suppression by Inverse Filter] FIG. 5 is a diagram for explaining one example of a conventional dereverberation method. 11
Is a reverberant speech Y (z), 12 is an acoustic transfer characteristic measuring unit, 13 is an inverse filter calculation unit, 14 is an inverse filter convolution unit, 15
Is the signal X '(z) after dereverberation.

First, in the acoustic transfer characteristic measuring section 12, a sound source such as a speaker capable of inputting a known input signal is placed at the position of the speaker, and the acoustic transfer characteristic is measured using the speaker and the sound receiver. This will be described with reference to FIG. 6, the elements in FIG. 4 described above are designated by the same reference numerals. Reference numeral 5 is an input terminal, and 6 is a speaker. First, the speaker 6 is installed at the same position as the speaker 2 in FIG. The acoustic transfer characteristic measuring unit 12 of FIG. 5 inputs, for example, a whiteness signal W (z) to the input end 5. At this time, the signal Y (z) observed at the output terminal 4 is

[0009]

[Equation 2] Becomes Here, since W (z) is already known, the acoustic transfer characteristic measuring unit 12 is, for example,

[0010]

(Equation 3) As a result, the acoustic transfer characteristic H (z) is measured. The acoustic transfer characteristic measuring unit 12 in FIG. 5 measures the acoustic transfer characteristic H (z) in advance by the method described above. Next, in the inverse filter calculation unit 13, I (z) = 1 as an inverse filter of H (z).
Calculate / H (z). In the inverse filter convolution unit 14,
This I (z) is convolved with the reverberation signal Y (z), and X '(z) = Y
(z) Outputs I (z).

If I (z) is exactly 1 / H (z), then X '(z) = Y (z) I (z) = H (z) X (z) / H (z ) =
X (z), and the original signal X (z) can be obtained as a dereverberation signal.

However, this method has a problem that it is necessary to measure the acoustic transfer characteristics at the position where the speaker 2 is generated in advance. In addition, there is a problem in that it is necessary to remeasure the acoustic transfer characteristics each time the position of the speaker 2 changes. Further, since the acoustic transfer characteristic H (z) has a non-minimum phase component, the inverse filter I (z) may not be calculated stably in some cases.

[Reverberation Suppression by Minimum Phase Inverse Filter] As a method for solving this problem, the following second conventional method which does not require measurement of acoustic transfer characteristics has been proposed.

In this system, when the logarithmic value of the reverberant voice Y (z) = H (z) X (z) is taken,

[0015]

(Equation 4) Focusing on the fact that it can be expressed separately, using the pre-recorded utterance signal X ₂ (z) of the same speaker instead of logX (z),

[0016]

(Equation 5) Is a method for trying to obtain H '(z). Speaker 2
In general, the frequency amplitude characteristics of X (z) and X ₂ (z) are similar, but the phase characteristics do not match. Therefore, in the above equation, the operation is performed focusing only on the amplitude component, that is, the absolute value. That is,

[0017]

(Equation 6) Is calculated. Here, | * | means an absolute value. H '
The phase component of (z) is determined using the Hilbert transform relationship, assuming that H '(z) is the minimum phase function. According to this method, the approximate value H '(z) can be determined without measuring the acoustic transfer characteristic H (z), and H'
By convolving the inverse characteristic 1 / H '(z) of (z) into the reverberation signal Y (z), the dereverberation signal X' (z) can be obtained.

This will be described with reference to FIG. In FIG. 7, 21 is the reverberant sound Y (z), 22 is the original sound recording unit, 2
3, 25 are absolute value logarithmic calculation units, 24 are original speech X ₂ (z), 2
6 is a subtraction unit, 27 is a Hilbert transform unit, 28 is an IDFT
(Inverse Discrete Fourier Transform) calculation unit, 29 is a window function processing unit, 30 is a DFT (Discrete Fourier Transform) calculation unit, 31
Is an exponential function conversion unit, 32 is an inverse filter calculation unit, 33 is an inverse filter convolution unit, and 34 is a dereverberation signal.

First, the original voice S (z) of the speaker to be uttered in the room is recorded in advance in the original voice recording section 22. At this time, the content of the speaker does not have to match the content of the reverberation signal. Next, the reverberant speech Y (z) and the original speech S (z) are passed to the absolute value logarithmic calculation units 23 and 25, respectively, and the absolute value logarithmic value l
og ′ | Y (z) | and log | X ₂ (z) | are calculated. Next, in the subtraction unit 26, log | Y (z) | and log | X
_From the difference of ₂ (z) |, the absolute value logarithmic value of the estimated value H ′ (z) of the acoustic transfer characteristic H (z) is

[0020]

(Equation 7) Is calculated as log | H ′ (z) | is passed to the Hilbert transform unit 27 and transformed into logH ′ (z) including the phase characteristic. In the Hilbert transform unit 27, specifically, the IDFT calculation unit 28 first performs an inverse discrete Fourier transform of log | H ′ (z) | to calculate h ′ (n).

Next, in the window function processing unit 29, it is 1 when n = 0, 2 when n = 1 to N / 2, and n = N / 2 to N.
A window function of 0 up to -1 is applied to h '(n). Here, N represents the length of the signal. The signal hm '(n) subjected to the window function in the window function processing unit 29 is passed to the DFT calculation unit 30, and logH' (z) is calculated by performing a discrete Fourier transform of hm '(n). . logH '(z)
Is passed to the exponential function conversion unit 31, and H ′ (z) = exp
(Log H '(z)) is calculated. As a result, H '(z) having the minimum phase characteristic is estimated. This result is passed to the inverse filter calculation unit 32, and the inverse filter I (z) = 1 /
H ′ (z) is calculated, and the inverse filter convolution unit 33 is further added.
Passed to. In the inverse filter convolution unit 33, I (z) and Y (z) are convolved, and X'of the dereverberation signal 34 is executed.
(z) = I (z) Y (z) is output. Where I (z) = 1
/ H '(z), Y (z) = H (z) X (z), so X'
(z) = H (z) / H '(z) X (z). Considering that H (z) is generally a non-minimum phase function and H '(z) is a minimum phase function, X' (z) has a frequency amplitude characteristic of X.
It matches (z), but the phase characteristics do not match only the non-minimum phase component. However, since there are generally few non-minimum phase components,
X (z) is a signal with reverberation suppressed.

In the above-described method, the signal X '(z) after dereverberation does not match X (z) by only the non-minimum phase component.
This has the advantage that it is not necessary to measure the room transfer characteristic H (z). Further, since H '(z) is the minimum phase function, there is an advantage that the inverse filter I (z) can be stably obtained.

However, this method has a demerit that the measurement of the room transfer function is not required, but the original voice of the same speaker who speaks in the reverberation room must be recorded in advance. This causes the original voice of the specific speaker to be recorded every time the speaker 2 changes.

[0024]

As described above, in the conventional reverberation suppressing apparatus, it is necessary to measure the acoustic transfer characteristics at the position where the speaker speaks in advance. Alternatively, there is a problem that the original voice of the same speaker who speaks must be recorded in advance.

The present invention has been made in view of the above problems, and does not require pre-measurement of the acoustic transfer characteristic at the position of the speaker, and further suppresses reverberation without knowing the information of the original voice of the speaker. The purpose is to do.

[0026]

SUMMARY OF THE INVENTION A reverberation suppressing method and apparatus according to the present invention has an acoustic transfer characteristic H (z) at a speaker position.
Instead of measuring, the method of measuring the acoustic transfer characteristic H ₂ (z) at an arbitrary position different from the speaker position in the same room, focusing on the components common to both, and obtaining the dereverberation filter. Is.

[0027]

In the present invention, first, the reverberation signal is divided into short time intervals, and the linear prediction coefficient (Linier P
redictive coefficient (LPC coefficient), and the linear prediction coefficient is convoluted with the reverberation signal to obtain the residual signal. The residual signal is convolved with the outline of the frequency-amplitude characteristic of the acoustic transfer characteristic measured in advance in the same room, and the effect of the pitch remaining in the residual signal is removed. From the residual signal corrected in this way, AR (Autoregr
essive: autoregressive) coefficient is obtained and this is used as a dereverberation filter. The present invention requires the measurement of the acoustic transfer characteristics in advance, but since the measurement position does not have to be the speaker's utterance position, there is no need to measure again even if the speaker moves the utterance position, and further, It has the advantage of not requiring the information of the original voice.

First, the AR model of the acoustic transfer function, which is the basis of the dereverberation method of the present invention, will be described below.
What is AR model?

[0029]

(Equation 8) Is a model represented by, and has an advantage that a resonance system can be expressed with a small number of parameters. here,
a _n is a coefficient called an AR coefficient, C is a constant, and P is the order of the AR model. Since this AR model is a minimum phase function, strict AR modeling cannot be performed for the indoor transfer characteristic that is a non-minimum phase function. However, the error between the true acoustic transfer characteristic H (z) and the transfer function of the AR model,

[0030]

[Equation 9] Of by determining AR coefficients so as to minimize the square expected, the true acoustic transfer function H (z) the AR model H _P (z)
Can be approximated by Now, a specific method of obtaining the AR coefficient based on the criterion of the square error minimum will be described. now,
When the impulse response h (k) of the true acoustic transfer characteristic H (z) in the time domain is described using the AR coefficient of the equation (8),

[0031]

(Equation 10) Becomes Here, e (k) is called a formula error (prediction residual). The squared expected value of this equation error,

[0032]

[Equation 11] The AR coefficient a _n that minimizes

[0033]

(Equation 12) Is obtained as a _n satisfying the above condition and is called a least squares solution.
Inverse characteristics of the thus AR model transfer characteristic with AR coefficients determined by H _P (z)

[0034]

(Equation 13) FIR (Finite duration im having
The result of convolving the pulse-response) filter with the true acoustic transfer characteristic H (z) is the minimized equation error.

[0035]

[Equation 14] Becomes Here, the equation error e (k) with respect to the AR coefficient a _n of the least-squares solution is a signal of the sum of the impulse that is the input signal and the whiteness error that occurs when the true acoustic transfer function is represented by the AR coefficient. Is known to be. That is, the characteristic A having the AR coefficient a _n of the least squares solution as the filter coefficient
As a result of convolving (z) with H (z), the frequency amplitude characteristic is flat, and the shape is close to an impulse even in the time domain. Hereinafter, the characteristic A (z) having the AR coefficient as the filter coefficient will be referred to as the AR characteristic.

In the present invention, this property is used to obtain A obtained by AR modeling the indoor transfer characteristic H (z).
(z) is used as a dereverberation filter.

Further, when the sound source is a white signal, the AR coefficient has an advantage that it can be estimated only from the reverberation signal without knowing the acoustic transfer characteristic H (z). Now, when the original signal is the white signal W (z) and the time domain signal is w (k), the output of the AR model is

[0038]

(Equation 15) Can be expressed as here,

[0039]

(Equation 16) When the AR coefficient a _n is determined so as to minimize the squared expected value of, the e (k) matches w (k), and the AR coefficient a _n is (1
This agrees with the case of the formula (2).

In general, since the voice generated by the speaker is not a white signal but a chromatic signal, the AR coefficient cannot be determined by assuming the original signal as a white signal as described above. Hereinafter, a method of determining the AR coefficient of the indoor transfer characteristic from the chromatic audio signal will be described.

Now, the voice signal X (z) is a vocal tract transfer function (envelope) D (z) that determines chromaticity, a white driving source (original voice) S (z), and a room transfer characteristic H (z). Is assumed to be represented by a macroscopic outline (envelope) G (z) and a fine structure F (z). That is, the reverberant voice Y (z) is

[0042]

[Equation 17] It can be expressed by. Here, the general shape G (z) of the indoor transfer function H (z) is almost equal to the general shape of the indoor transfer characteristic at other positions in the same room.

First, consider whitening the audio signal X (z) in order to estimate the AR coefficient. That is, D (z) representing chromaticity is removed from the reverberant voice Y (z), and R (z) = G
If (z) F (z) S (z) = H (z) S (z), then (1)
The AR coefficient of H (z) can be obtained by replacing y (k) in the equation 6) with R (z). Now, in order to remove D (z), specifically, Y (z) is written in a short time frame (for example, 16
LPC analysis may be performed every (ms). However, when obtaining the AR coefficient from the residual signal R (z) after the LPC analysis,
Two problems arise. The first is that the residual signal R (z) is not completely white and a peak remains as an effect of the pitch of the driving sound source S (z). Second, the LPC analysis shows the outline G of the room transfer characteristic H (z) along with Y (z) to D (z).
(z) is also removed, and the residual signal R (z) = F (z) S
It is a point that becomes (z). When the AR coefficient is obtained from this, it is not the AR coefficient of H (z) but the AR of F (z) S (z).
The coefficient will be obtained. Therefore, in the present invention, the notch filter N (z) is applied to the strongest peak remaining in the residual signal R (z). Furthermore, from the room transfer function at a position that can be measured in advance, A of the same order as in LPC analysis is used.
The rough shape G '(z) is calculated as the R coefficient, and R'
(z) = G '(z) F (z) N (z) S (z) Then, the AR coefficient is obtained. Thus, A (z) H (z) X (z) = X '
(z) is required. This suppresses the peak due to resonance and reduces reverberation.

[0044]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of a dereverberation suppressing method and apparatus according to the present invention will be described below with reference to FIG. 1 showing the configuration and operation. In FIG. 1, 41 is a Y (z) reverberant voice, 42 is an LPC coefficient calculation unit, 43 is a residual signal calculation unit,
44 is a frequency peak calculation unit, 45 is a notch filter calculation unit, 46 is a frequency peak correction unit, 47 is a rough shape correction unit, 4
8 is an AR coefficient estimation unit, 49 is a dereverberation filter convolution unit, 50 is a dereverberation signal X '(z), 51 is an acoustic transfer characteristic measurement unit of H ₂ (z) at another position in the same room, 52 Is the rough shape estimation unit.

First, in the present invention, the acoustic transfer characteristic measuring unit 5
1, the acoustic transfer characteristics H of the reverberation path at other positions in the same room
Measure ₂ (z) in advance. Next, the outline estimation unit 52
Then, with respect to the measured acoustic transfer characteristic H ₂ (z), LP with the same order as used when obtaining the residual signal from the reverberant speech is used.
C (Linear Predictive Coding) analysis is performed to find its outline G '(z).

Next, the LPC coefficient calculating section 42 obtains the LPC coefficient D '(z) of the reverberation signal Y (z) of the reverberant voice 41 actually collected. Next, the residual signal calculation unit 43 convolves the inverse characteristic of the LPC coefficient D '(z) with the reverberation speech signal to obtain the residual signal R (z). Here, as described above, L
The PC coefficient D '(z) has a general shape D (z) that represents the chromaticity of speech.
And the acoustic transfer characteristic outline G (z) are included, the convolved output R (z) is R (z) = F (z) S (z). Next, in the frequency peak calculation unit 44, R
Find the peak due to the influence of the pitch component of the voice remaining in (z). Specifically, the maximum amplitude of the frequency amplitude characteristic of the residual signal R (z) and its frequency are searched and recorded. Further, the average value of the frequency amplitude characteristic is calculated and recorded. The notch filter calculation unit 45 generates a notch filter that sets the maximum amplitude value to the average amplitude value based on the result of the frequency peak calculation unit 44. The frequency peak correction unit 46 convolves the output R (z) of the residual signal calculation unit 43 with the notch filter N (z) to obtain a corrected residual signal R '(z) with a flat frequency amplitude characteristic.
= N (z) R (z) is obtained. Further, in the outline correction unit 47,
Other transfer characteristic outline G ′ estimated by the outline estimating unit 52.
(z) is convolved with R '(z), and the corrected residual signal R "(z) =
G '(z) R' (z) = G '(z) N (z) R (z) = G' (z) F
(z) N (z) S (z) is obtained. The AR coefficient estimator 48 calculates the time domain signal r ″ (n) of R ″ (z) as y in equation (16).
The AR coefficient is estimated by replacing (n). Signal N (z) S
Considering that (z) is a substantially white signal in which the influence of pitch is removed, and G ′ (z) is almost equal to G (z), the AR coefficient a obtained from R ″ (z) _n is H (z)
It can be seen that the AR coefficient of the R modeled _HP (z) is almost the same. The dereverberation filter convolution unit 49 convolves the AR characteristic A (z) with the reverberation signal Y (z) to obtain a signal X '(z) after dereverberation processing.

In order to clarify the effect of the present invention, 16
A computer simulation was performed using reverberant voice in which a male voice for 5 seconds picked up by kHz sampling was convoluted with acoustic transfer characteristics measured at a sound source / sound receiving point distance of 3 m in a room with a reverberation time of 250 ms.

First, with respect to the reverberation signal, the frame length is 16 m.
An 18th-order LPC analysis was performed with s to obtain the residual signal R (z). A contour G '(z) obtained as an 18th-order AR coefficient from the acoustic transfer characteristics measured in advance at a point 1 m away from the sound source position was convoluted with this, and notch filtering was performed to obtain a corrected residual signal. Next, the 300th-order AR coefficient a _n was estimated from this corrected residual signal.

FIG. 2 shows the result of convolving the estimated AR characteristic A (z) with the original acoustic transfer characteristic and the frequency-amplitude characteristic of the original acoustic transfer characteristic. Reference numeral 61 is the frequency amplitude characteristic of the original true acoustic transfer characteristic, and 62 is the frequency amplitude characteristic of the acoustic transfer characteristic after the dereverberation process. It can be seen that in the acoustic transfer characteristic 62 after the dereverberation process, the peaks near 1 kHz are suppressed and the frequency characteristic is flattened as a whole, compared with the original acoustic transfer characteristic 61.

FIG. 3 shows an impulse response waveform 71 and an impulse response waveform 72 of the original true acoustic transfer characteristic. 1
It can be seen that the amplitude of the reflected sound is suppressed from 5 ms to 25 ms.

[0051]

The reverberation suppressing method and apparatus according to the present invention obtain a residual signal from a reverberant voice signal picked up in a target acoustic system (for example, in a room). It is characterized in that the removal processing and the convolution processing of the approximate shape of the acoustic transfer characteristics measured at other positions are corrected, an AR coefficient is obtained from the corrected residual signal, and this is used as a dereverberation filter coefficient. This is a reverberation suppression method. When the present invention is applied to a reverberant voice signal picked up in a reverberant room, the characteristics are flattened in the frequency domain,
Even on the time axis, the shape becomes close to an impulse, and the clarity of the audio signal is improved. Furthermore, unlike the conventional dereverberation method, the present invention does not need to know the characteristics of the original signal and the acoustic transfer characteristics between the speaker and the sound receiver, so that the change of the speaker and the speaker position Greatly improves the tracking performance for deflection.

[Brief description of drawings]

FIG. 1 is a diagram illustrating a dereverberation processing method and apparatus according to the present invention.

FIG. 2 is a diagram showing an effect of the present invention by a frequency amplitude characteristic.

FIG. 3 is a diagram showing an effect of the present invention with an impulse response waveform.

FIG. 4 is a diagram illustrating a reverberation audio signal.

FIG. 5 is a diagram illustrating a conventional reverberation suppression processing method using an inverse filter.

FIG. 6 is a diagram illustrating an example of a method of measuring acoustic transfer characteristics.

FIG. 7 is a diagram illustrating a conventional dereverberation processing method using a minimum phase inverse filter.

[Explanation of symbols]

 41 Reverberant Speech 42 LPC Coefficient Calculator 43 Residual Signal Calculator 44 Frequency Peak Calculator 45 Notch Filter Calculator 46 Frequency Peak Corrector 47 General Shape Corrector 48 AR Coefficient Estimator 49 Reverberation Suppression Filter Convolution 50 50 Dereverberation Signal 51 Acoustic transfer characteristic measurement unit 52 General shape estimation unit

Claims (4)

[Claims] 1. Reverberation for reconstructing the sound source signal by estimating the inverse characteristic of the reverberant path from the reverberant signal after the sound source signal has passed through the reverberant path and convolving the inverse characteristic of the reverberant path with the sound source signal. A suppressor, which divides the reverberation signal into short time sections and calculates an LPC coefficient in each section to obtain a linear prediction coefficient, and convolves the linear prediction coefficient with the reverberation signal to obtain a residual signal. Residual signal calculation means, frequency peak correction means for removing a pitch frequency from the residual signal, acoustic transfer characteristic measurement means for measuring a reverberation path near the reverberation path, and frequency amplitude of the measured reverberation path. A rough shape estimating means for obtaining a rough shape of the characteristic, a rough shape correcting means for convolving the rough shape with a signal from which the pitch frequency has been removed by the frequency peak correcting means, and a long time for the convolved signal. An AR coefficient estimating unit for obtaining an AR coefficient in a section, and a dereverberation filter convolution unit for inputting the reverberation signal to an FIR filter having the AR coefficient as a filtering unit to suppress reverberation. apparatus. 2. A frequency peak calculating means for searching for a peak due to the influence of the pitch component of the voice of the residual signal obtained by the residual signal calculating means, and a maximum amplitude value as an average amplitude based on the result of this frequency peak calculating means. The reverberation suppressing apparatus according to claim 1, further comprising notch filter calculating means for inputting the value into the frequency peak correcting means. 3. Reverberation for reconstructing the sound source signal by estimating the reciprocal characteristic of the reverberant path from the reverberant signal after the sound source signal has passed through the reverberant path, and convolving the reciprocal characteristic of the reverberant path with the sound source signal. A method of suppressing, wherein the reverberation signal is divided into short time sections, linear prediction coefficients are obtained in each section, and the linear prediction coefficient is convoluted with the reverberation signal to obtain a residual signal, The step of removing the pitch frequency from the difference signal, measuring the reverberation path in the vicinity of the reverberation path, the step of obtaining the outline of the frequency amplitude characteristics of the measured reverberation path, the signal from which the pitch frequency is removed A step of convolving the outline, a step of obtaining an AR coefficient of the convoluted signal in a long period of time, and a step of inputting the reverberation signal to an FIR filter means having the AR coefficient as a filter means. Therefore, a reverberation suppressing method comprising: a step of suppressing reverberation. 4. A notch to be input in the process of searching for a peak due to the influence of the pitch component of the voice of the residual signal and the process of removing the pitch frequency from the residual signal by setting the maximum amplitude value to the average amplitude value based on this peak. The dereverberation method according to claim 3, further comprising a filtering process.