JP5815614B2 - Reverberation suppression apparatus and method, program, and recording medium - Google Patents

Description

  The present invention relates to a reverberation suppressing apparatus and method, a program, and a recording medium for suppressing a component due to room reverberation from a signal picked up by a microphone.

  When sound is collected indoors, the reverberant sound component reflected from the wall or floor is picked up simultaneously with the direct sound and the sound deteriorates. For example, in a voice conference by hands-free in a large conference room or a call by a mobile terminal in a place with much reverberation, the sound becomes difficult to hear due to the effect of reverberation.

  Therefore, conventionally, a method for suppressing a reverberation component for the purpose of reducing the sound quality deterioration due to such reverberation has been proposed. For example, a dereverberation apparatus 900 using multi-step linear prediction disclosed in Non-Patent Document 1 is known.

FIG. 17 shows a functional configuration of the dereverberation apparatus 900 and its operation will be briefly described. The dereverberation apparatus 900 includes a whitening unit 910, a multi-step linear prediction unit 920, a reverberation calculation unit 930, an FFT 940, an FFT unit 950, a spectral subtraction unit 960, and an inverse FFT unit 970. The whitening unit 9 1 0 whitens the time-domain sound collected signal by removing frequency characteristics due to speech autocorrelation using linear prediction with a short tap length. The multi-step linear prediction unit 9 20 performs filter steps for predicting reverberation components by performing multi-step linear prediction with a long tap length on a whitened time domain sound pickup signal. The long tap length is a length of about 5 to 6,000 points corresponding to a reverberation time of 600 to 700 ms, assuming that the sampling frequency is 8 kHz, for example. The reverberation calculation unit 9 30 predicts a reverberation component by filtering the collected sound signal in the time domain with the calculated filter coefficient.

  The FFT unit 940 converts the predicted reverberation component into a frequency domain reverberation component that is a frequency domain signal by short-time Fourier transform. The FFT 950 converts the time domain sound collection signal into a frequency domain sound collection signal that is a frequency domain signal by short-time Fourier transform.

  Spectral subtraction unit 960 calculates a gain for suppressing reverberation from the power for each frequency of the frequency domain reverberation component and the power for each frequency of the frequency domain sound collection signal, and multiplies the frequency domain sound collection signal by the gain. A reverberation suppression signal in which reverberation is suppressed is output. The reverberation suppression signal is converted into a time domain reverberation suppression signal by inverse Fourier transform.

Keisuke Kinoshita, Marc Delcroix, Tomohiro Nakaatani, and Masato Miyoshi, “Suppression of Late Reverberation Effect on Speech Signal Using Long-Term Multiple-step Linear Prediction,” IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, No. 4, MAY 2009.

However, since the conventional dereverberation apparatus 900 uses linear prediction with a long tap length, there is a problem that the amount of calculation becomes enormous. In the case of linear prediction by the Levinson-Durbin algorithm, the amount of calculation is on the order of the square of the tap length. In the above example, a calculation amount of 6 × 10 ⁶ order is required.

  The present invention has been made in view of this problem, and an object of the present invention is to provide a reverberation suppressing device and method, a program, and a recording medium for suppressing reverberation with a low amount of computation.

  The dereverberation apparatus of the present invention includes an FFT unit, a power calculation unit, a delay unit, a subtraction unit, an adaptive algorithm unit, a frequency domain filter unit, a dereverberation suppression gain calculation unit, a multiplication unit, and an inverse FFT unit. And. The FFT unit converts the collected sound signal into a frequency domain sound collected signal in the frequency domain. The power calculation unit outputs a power spectrum in which the power of the frequency domain sound collection signal output from the FFT unit is calculated for each frequency. The delay unit outputs a delay power spectrum obtained by delaying the power spectrum output from the power calculation unit by a predetermined delay amount. The subtracting unit subtracts the filtered signal obtained by multiplying the delayed power spectrum output from the delay unit by the filter coefficient from the power spectrum output from the power calculating unit to obtain the estimated power of the direct sound. The adaptive algorithm unit receives the delay power spectrum output from the delay unit and the estimated power of the direct sound, and updates the filter coefficient so as to minimize the estimated power of the direct sound. The frequency domain filter unit outputs a filtered signal obtained by filtering the delayed power spectrum output from the delay unit by the filter coefficient. The reverberation suppression gain calculation unit calculates the reverberation suppression gain for suppressing the reverberation sound by using the power spectrum output from the power calculation unit and the estimated power of the direct sound output from the subtraction unit as inputs. The multiplication unit multiplies the frequency domain collected signal by a dereverberation gain and outputs a dereverberation signal. The inverse FFT unit converts the dereverberation signal output from the multiplication unit into a dereverberation signal in the time domain.

  According to the dereverberation apparatus of the present invention, the filter coefficient is updated in the region of the power spectrum using an adaptive algorithm so as to minimize the estimated power of the direct sound that is the output signal of the subtraction unit. Reverberation suppression can be performed with a calculation amount in the order of the first power. Although specific effects will be described later, the amount of calculation can be significantly reduced as compared with the conventional method.

The figure which shows the function structural example of the dereverberation apparatus 100 of this invention. The figure which shows the operation | movement flow of the dereverberation apparatus. The figure which shows the function structural example of the frequency domain filter part 160. FIG. The figure which shows the function structural example of the adaptive algorithm part 150. FIG. The figure which shows the function structural example of the reverberation suppression gain calculation part 170. FIG. The figure which shows the function structural example of the adaptive algorithm part 250 of the dereverberation apparatus 200 of this invention. The figure which shows the operation | movement flow of the adaptive algorithm part 250. The figure which shows the function structural example of the adaptive algorithm part 350 of the dereverberation apparatus 300 of this invention. The figure which shows the operation | movement flow of the adaptive algorithm part 350. The figure which shows the function structural example of the reverberation suppression gain calculation part 470 of the reverberation suppression apparatus 400 of this invention. The figure which shows the operation | movement flow of the first half of the reverberation suppression gain calculation part. The figure which shows the operation | movement flow of the latter half of the dereverberation gain calculation part. The figure which shows the function structural example of the dereverberation apparatus 500 of this invention. The figure which shows the function structural example of the dereverberation apparatus 600 of this invention. The figure which shows the function structural example of the dereverberation apparatus 700 of this invention. The figure which shows the result of evaluation experiment. The figure which shows the function structure of the conventional dereverberation apparatus 900.

  Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

  FIG. 1 shows a functional configuration example of a dereverberation apparatus 100 of the present invention. The operation flow is shown in FIG. The dereverberation apparatus 100 includes an FFT unit 110, a power calculation unit 120, a delay unit 130, a subtraction unit 140, an adaptive algorithm unit 150, a frequency domain filter unit 160, a dereverberation suppression gain calculation unit 170, and a multiplication unit. 180, an inverse FFT unit 190, and a control unit 195. The dereverberation apparatus 100 is realized by reading a predetermined program into a computer composed of, for example, a ROM, a RAM, a CPU, and the like, and executing the program by the CPU. The same applies to each device described below.

  The FFT unit 110 converts the collected sound signal into a frequency domain sound collected signal in the frequency domain (step S110). The collected sound signal is a signal composed of a direct sound collected by a microphone (not shown) and a reverberant sound reflected by an indoor wall or floor. The collected sound signal is, for example, a digital signal digitized at a sampling frequency of 8 kHz. In FIG. 1, the notation of an A / D converter that digitizes a collected sound signal and a D / A converter that converts a digital signal into a continuous value is omitted. The FFT unit 110 converts the collected sound signal, which has been made discrete by the short-time Fourier transform, into a frequency domain sound collected signal at intervals of, for example, 128 frames (t = 16 ms) and a window size of 32 ms. The window is obtained by taking the square root of the Hanning window.

  The power calculation unit 120 receives the frequency domain sound collection signal output from the FFT unit 110, calculates a power spectrum of the frequency domain sound collection signal, and outputs the power spectrum (step S120). The delay unit 130 outputs a delay power spectrum obtained by delaying the power spectrum output from the power calculation unit 120 by a predetermined delay amount (step S130). The predetermined delay amount is a delay time of about several tens to 100 ms, for example.

  The subtracting unit 140 subtracts the filtered signal obtained by multiplying the delayed power spectrum by the filter coefficient obtained by the adaptive algorithm unit 150 from the power spectrum output from the power calculating unit 120 to obtain the estimated power of the direct sound (step S140). . The adaptive algorithm unit 150 receives the delay power spectrum and the estimated power of the direct sound as input, and updates the filter coefficient so as to minimize the estimated power of the direct sound (step S150).

  The frequency domain filter unit 160 outputs a filtered signal that has been filtered by multiplying the delay power spectrum by a filter coefficient (step S160). The reverberation suppression gain calculation unit 170 receives the power spectrum and the estimated power of the direct sound as input, and calculates a reverberation suppression gain for suppressing the reverberation sound (step S170).

  Multiplier 180 multiplies the frequency domain collected signal by a dereverberation gain and outputs a dereverberation suppression signal (step S180). The inverse FFT unit 190 performs inverse Fourier transform on the dereverberation signal at the same interval and the same number as the FFT unit 110, multiplies the output by a window, and performs overlap addition to convert the signal into a time domain dereverberation signal (step S190). ). The processes in steps S110 to S190 described above are repeated until the operation is stopped while updating the frame (step S196) (No in step S195). The control unit 195 controls this repetitive process. The control unit 195 controls the time series operation of the dereverberation apparatus 100 and is not special.

  In the dereverberation apparatus 100, the delay power spectrum delayed by the delay unit 130 by updating the filter coefficient using an adaptive algorithm so as to minimize the error signal (estimated power of the direct sound) output from the subtraction unit 140. To predict the power spectrum before the delay. That is, the error signal output from the subtracting unit 140 is a signal component that could not be predicted, and the component subtracted by the subtracting unit 140 is a signal component that could be predicted.

  On the other hand, the reverberation in the room is a phenomenon in which when the sound reaches the microphone from the sound source, the sound reflected by the wall or the like arrives at the microphone with a delay from the direct sound. Since the reflected sound always arrives later than the direct sound, the reverberation component can be predicted from the sound that has reached the microphone in the past. On the other hand, since the direct sound reaches the earliest, it cannot be completely predicted from the past arrival sounds. Therefore, the error signal output from the subtractor 140, which is a component that cannot be predicted from the past signal, is the estimated power of the direct sound.

  Therefore, if the frequency domain sound pickup signal is multiplied by the gain obtained by dividing the estimated power of the direct sound by the power of the input signal, the power spectrum of the input signal (sound pickup signal) is transformed into the power spectrum of the direct sound component. Therefore, reverberant sound can be suppressed. Thus, the reverberation suppression apparatus 100 can suppress reverberation using an adaptive algorithm in the power spectrum region. By using an adaptive algorithm, dereverberation can be realized with a calculation amount in the order of the first power of the tap length of the frequency domain filter unit 160.

  Hereinafter, the operation of the dereverberation apparatus 100 will be described in more detail by showing more specific functional configuration examples of the respective units.

[Frequency domain filter section]
FIG. 3 shows a functional configuration example of the frequency domain filter unit 160. The frequency domain filter unit 160 includes a signal buffer unit 161 and a convolution calculation unit 162.

The signal buffer means 161 stores the past M pieces of the delay power spectrum | X (ω, t−D) | ² output from the delay unit 130. However, t is a frame number, ω is a frequency, D is a delay amount given by the delay unit, and X (ω, t) is a frequency domain sound pickup signal output from the FFT unit 110.

The convolution calculation means 162 applies the M tap filter coefficients F (ω, m, t), m = 0,..., M−1 held in the adaptive algorithm unit 150 to the delay power spectrum | X (ω , T−D) | ² , a post-filter signal | Y (ω, t−D) | ² is calculated by the following equation and output to the subtracting unit 140.

[Adaptive algorithm part]
FIG. 4 shows a functional configuration example of the adaptive algorithm unit 150. The adaptive algorithm unit 150 includes an update vector calculation unit 151, a step size multiplication unit 152, an addition unit 153, and a filter coefficient holding unit 154.

The update vector calculation means 151 receives the delay power spectrum | X (ω, t−D) | ² and the output signal E (ω, t) of the subtraction unit 140 as inputs, and uses the adaptive algorithm to update the vector U (ω, m , T). The output signal E (ω, t) of the subtracting unit 140 is an error obtained by subtracting the filtered signal | Y (ω, t−D) | ² from the power spectrum | X (ω, t) | ² as shown in the following equation. Signal. The output signal E (ω, t) of the subtractor 140 is a signal component that could not be predicted from the past signal, and is the estimated power of the direct sound.

  As an adaptive algorithm, for example, the NLMS method (reference 1: Simon Haykin, Adaptive filter theory. Prentice-Hall, 1986) is used. The update vector calculation means 151 calculates the update vector U (ω, m, t) by the following equation.

  Here, n is a filter coefficient number.

  The step size multiplier 152 outputs a filter coefficient obtained by multiplying the update vector U (ω, m, t) by a preset step size α in the range of 0-2. The adding means 153 adds the filter coefficient of the previous frame held in the filter coefficient holding means 154 to the filter coefficient output from the step size multiplier 152, and adds the filter coefficient F (ω, m, t + 1) of the next frame. ).

An adaptive algorithm such as the NLMS method operates so as to minimize the mean square of the error signal E (ω, t) output from the subtraction unit 140. Therefore, a signal obtained by predicting the current frequency domain sound collection signal X (ω, t) as much as possible from the past delay power spectrum | X (ω, t−D) | ^{2 by} D and removing the predicted component. An error signal E (ω, t) is obtained. Therefore, only the components that cannot be predicted from the past signal remain, and the error signal E (ω, t) becomes the estimated power of the direct sound.

[Reverberation suppression gain calculator]
FIG. 5 shows a functional configuration example of the reverberation suppression gain calculation unit 170. The reverberation suppression gain calculation unit 170 includes a dividing unit 171, a maximum value limiting unit 172, and a time smoothing unit 173.

The dividing means 171 receives the power spectrum | X (ω, t) | ² and the estimated direct sound power E (ω, t), calculates the ratio between them, and raises the ratio to the β power to obtain the gain G ′ ( (ω, t) is calculated (formula (5)).

  Here, β is a preset constant, and the larger the value, the stronger the reverberation suppression amount. β is set between approximately 0.5 and 1.0.

  The maximum value limiting means 172 limits the value of the gain G ′ (ω, t) output from the dividing means 171 with 1 as the upper limit as shown in the equations (6) and (7).

  The time smoothing unit 173 time smoothes the limited gain G ″ (ω, t) output from the maximum value limiting unit 172, and outputs a reverberation suppression gain G (ω, t). For example, it is realized by the following equation.

  Here, γ is a smoothing coefficient and is set in advance. γ takes a value in the range of 0 to 1, and the closer to 1, the smoothing is with a longer time constant.

  The reverberation suppression gain G (ω, t) output from the reverberation suppression gain calculation unit 170 is calculated based on a value obtained by dividing the estimated direct sound estimated power (error signal) by the power of the collected sound signal. By multiplying the reverberation suppression gain G (ω, t) by the frequency domain collected signal X (ω, t), an output in which the reverberation component is suppressed can be obtained.

  The dereverberation apparatus 100 converts the collected sound signal into a frequency domain signal using FFT and inverse FFT, and realizes direct power component estimation using an adaptive algorithm. A reverberant sound can be suppressed with a low calculation amount in the order of the first power of long.

  FIG. 6 shows a functional configuration example of the adaptive algorithm unit 250 of the dereverberation apparatus 200 of the present invention. The operation flow is shown in FIG. The dereverberation apparatus 200 has a configuration in which the adaptive algorithm unit 150 of the dereverberation apparatus 100 is simply replaced with the adaptive algorithm unit 250, and thus an overall functional configuration example thereof is omitted.

The adaptive algorithm unit 250 is different from the above-described adaptive algorithm unit 150 (FIG. 4) in that a step size setting unit 255 is provided. The step size setting means 255 updates the step size α according to the value of the update vector U (ω, m, t) output from the update vector calculation means 151. For example, if the update vector U (omega, m, t) is a positive value sets the step size alpha ₁ small value set in advance in (Yes in step S2551) (Equation (9)) (step S2552), update vector U (ω, m, t) may take a negative value the step size alpha ₂ setting a large value set in advance in (No in step S2551) (equation (10)) (step S2553).

  By setting the step size in this way, the direct sound estimation error can be controlled. When the step size is not controlled, the estimation error appears evenly in the plus and minus directions, but by controlling as described above, it is possible to control so that the direct sound estimation error appears in a positive direction. That is, it is possible to prevent the direct sound power from being estimated to be small, and to suppress deterioration in sound quality.

  Thus, according to the dereverberation apparatus 200, in addition to the effect of reducing the calculation amount obtained by the dereverberation apparatus 100, an effect of outputting a high-quality dereverberation sound can be achieved.

  FIG. 8 shows a functional configuration example of the adaptive algorithm unit 350 of the dereverberation apparatus 300 of the present invention. The operation flow is shown in FIG. The dereverberation apparatus 300 is different from the dereverberation apparatuses 100 and 200 in that the adaptive algorithm unit 350 includes a non-negative constraint means 356.

  When the updated filter coefficient F (ω, m, t + 1) output from the adding means 153 becomes a negative value, the non-negative constraint means 356 replaces the value with 0 (step S3562), thereby obtaining a filter coefficient. Control is performed so that F (ω, m, t + 1) does not become a negative value.

  In the present invention, since the reverberation component is estimated by paying attention to the power of a signal having only a positive value, the correct solution is to take a positive value for the filter coefficient. Since the fact that the filter coefficient F (ω, m, t + 1) takes a negative value is an estimation error, it can be corrected to a more accurate filter coefficient by replacing the value with 0. Thus, according to the dereverberation apparatus 300, the filter coefficients can be accurately obtained for the dereverberation apparatuses 100 and 200, and high-quality dereverberation can be performed.

  FIG. 10 shows a functional configuration example of the dereverberation gain calculation unit 470 of the dereverberation apparatus 400 of the present invention. The operation flow is shown in FIG. 11 and FIG. Since the dereverberation apparatus 400 has a configuration in which the dereverberation suppression gain calculation unit 170 of the dereverberation suppression apparatus 100, 200, 300 is replaced with the dereverberation suppression gain calculation unit 470, an example of the overall functional configuration thereof is omitted.

Reverberation suppression gain calculator 470, to the reverberation suppression gain calculator 170 (FIG 5), further differs in that comprises masking level calculation section 474 and the maximum value selection unit 47 6. The masking level calculation means 474 obtains the auditory masking level Q (ω, t) from the estimated power E (ω, t) of the direct sound output from the subtracting unit 140. Auditory masking is a phenomenon in which when the power of frequency K is P (K), the sound component below the auditory masking level that can be calculated as a function of P (K) cannot be heard by human hearing.

  The auditory masking level Q (ω, t) can be obtained, for example, as follows. When obtaining the auditory masking level Q (ω, t) at a certain frequency ω, the estimated power E (ω, t) of the direct sound at that frequency and the provisional auditory masking level Q ′ (ω at the next lower frequency) −1, t) are respectively multiplied by coefficients a and b (step S4744), and the larger value is set as a provisional auditory masking level Q ′ (ω, t) (steps S4745 and S4746). This is repeated in order from the minimum value of the frequency ω until ω = maximum value (No loop in step S4747). The coefficients a and b are constants less than 1 and greater than or equal to 0, which are preset based on auditory masking characteristics. The coefficients a and b may be set to different values depending on the frequency ω.

  When the frequency ω becomes the maximum value (Yes in step S4747, connector A), then, in order from the maximum value of the frequency ω, the provisional auditory masking level Q ′ (ω, t) of the frequency ω and the next higher one A value obtained by multiplying the provisional auditory masking level Q ′ (ω + 1, t) of the frequency by the coefficient c is compared (step S4751), and a larger value is set as the auditory masking level Q (ω, t) (steps S4752, S4753). The coefficient c is a constant less than 1 and greater than or equal to 0 that is preset based on the characteristics of auditory masking. The coefficient c may be set to a different value depending on the frequency ω.

  The auditory masking level Q (ω, t) can be obtained by the above method.

  The maximum value selection means 475 compares the error signal E (ω, t) output from the subtraction unit 140 with the auditory masking level Q (ω, t) (step S4761), and uses the larger value as a new direct sound. The estimated power (error signal E (ω, t)) is output to the dividing means 171.

  By the above method, the reverberation suppression apparatus 400 can suppress the reverberation component that cannot be heard in the sense of hearing by using the auditory masking characteristic. Therefore, the reverberation suppression apparatus 400 can prevent unnecessary suppression of reverberation sound, and has an effect of reducing direct sound deterioration.

  FIG. 13 shows a functional configuration example of the dereverberation apparatus 500 of the present invention. The dereverberation apparatus 500 is obtained by adding the configuration of the adaptive section detection unit 511 to the dereverberation apparatus 100.

The adaptive interval detection unit 511 compares the magnitude of the delay power spectrum | X (ω, t−D) | ² output from the delay unit 130 with a preset threshold value, and the magnitude of the delay power spectrum exceeds the threshold value. Only in this case, control is performed so that the filter coefficient is updated by the adaptive algorithm unit 150.

  By adding this control, it is possible to stop the update of the filter coefficient in the section where the signal level of the collected sound signal is small and susceptible to ambient noise, so that reverberation can be suppressed with higher accuracy. Filter coefficients can be determined.

  The dereverberation apparatus 500 has been described with the configuration in which the adaptive interval detection unit 511 is added to the dereverberation device 100. However, the configuration of the adaptive interval detection unit 511 is included in any of the above-described dereverberation devices 200, 300, and 400. The same effect can be obtained by adding.

  FIG. 14 shows a functional configuration example of a dereverberation apparatus 600 of the present invention. The dereverberation device 600 is obtained by adding the configurations of a band aggregation unit 611 and a band expansion unit 612 to the dereverberation device 100.

The band aggregating unit 611 aggregates the frequency ω of the power spectrum | X (ω, t) | ² output from the power calculating unit 120 and converts the frequency ω ′ to a smaller frequency division number. The frequency ω is divided into a plurality of groups, and the sum of the power spectrum | X (ω, t) | ² output from the power calculation unit 120 is taken for each group, and the value is taken as the power spectrum | X of the new frequency ω ′. It outputs as '(ω', t) | ² (formula (16)).

  Here, Ω (ω ′) is a set of frequencies ω belonging to the group of ω ′.

The power spectrum | X ′ (ω ′, t) | ² in which the frequency ω output from the band aggregating unit 611 is aggregated is supplied to the delay unit 130, the dereverberation suppression gain calculation unit 170, and the subtraction unit 140, respectively. Are processed in units of frequency ω ′.

  The band expanding unit 612 converts the frequency ω ′ of the dereverberation suppression gain G ′ (ω ′, t) obtained by the dereverberation suppression gain calculation unit 170 into a frequency ω that is the frequency division number of the FFT unit 110. The reverberation suppression gain of the aggregated frequency ω ′ to which the frequency ω of the FFT unit 110 belongs is expanded into the frequency ω by copying the reverberation suppression gain of the frequency ω (formula (17)).

  Thus, the amount of calculation can be reduced by collecting the frequency ω and calculating the reverberation suppression gain. Since the auditory characteristics have logarithmic sensitivity to frequency, the higher the frequency, the less the deterioration in audibility even if the frequency aggregation number is increased. You can change it.

  The dereverberation apparatus 600 has been described with the configuration in which the band aggregation unit 611 and the band expansion unit 612 are added to the dereverberation apparatus 100. However, any of the above-described dereverberation apparatuses 200, 300, 400, and 500 has the band. By adding the configurations of the aggregation unit 611 and the band expansion unit 612, the same effect can be obtained.

  FIG. 15 shows a functional configuration example of a dereverberation apparatus 700 of the present invention. The dereverberation device 700 is obtained by adding a configuration of a coefficient multiplication unit 711, a second frequency domain filter unit 712, and a subtraction unit 713 to the dereverberation device 100.

  The coefficient multiplication unit 711 multiplies each of the filter coefficients F (ω, m, t) output from the adaptive algorithm unit 150 by a preset coefficient H (m), and converts the converted filter coefficient F ′ (ω, m , T) is output (formula (18)).

  By setting the coefficient of the portion where m is small among the coefficients H (m) to be smaller than 1, it is possible to reduce erroneous estimation of the reverberation characteristics in the human oral cavity as reverberation components. Since the reverberation characteristic in the human mouth is a response in a short time as compared with the reverberation in the room, it greatly affects the filter coefficient of the portion where m is small. By multiplying the filter coefficient of the portion where m is small by a coefficient H (m) smaller than 1, the influence can be reduced.

The second frequency domain filter unit 712 multiplies the delay power spectrum | X (ω, t−D) | ² output from the delay unit 130 by the filter coefficient after multiplication by the coefficient H (m). The subtraction unit 713 subtracts the output signal of the second frequency domain filter unit 712 from the power spectrum output from the power calculation unit 120. The output of the subtraction unit 713 is input to the reverberation suppression gain calculation unit 170 instead of the error signal.

  According to the reverberation suppression apparatus 700, it is possible to reduce erroneous estimation of the reverberation characteristics in the human mouth as a reverberation component. The dereverberation apparatus 700 has been described with a configuration in which a coefficient multiplication unit 711, a second frequency domain filter unit 712, and a subtraction unit 713 are added to the dereverberation apparatus 100, but the above-described dereverberation apparatuses 200, 300, and 400 are described. , 500, 600, the same effect can be obtained by adding the same functional component.

[Evaluation experiment]
For the purpose of confirming the effect of the present invention, an evaluation experiment was performed comparing the reverberation suppression performance with the conventional method. The evaluation experiment was performed by comparing the conventional method with the method of the present invention, by generating a sound with pseudo reverberation by computer simulation using the mirror image method, and performing the reverberation suppression processing on the sound. Evaluation was performed by PESQ (Perceptual Speech Quality Measure).
The conditions for generating reverberant speech are as follows. Assume that a non-directional sound source and a non-directional microphone are arranged at a distance of 1 m in a 6 m × 10 m × 3 m room, and the microphone sound reception signal is generated by changing the reflection coefficient of the wall surface. As a sound source signal, 8 kHz sampling voice (female voice 10s, male voice 10s) was used.

  FIG. 16 shows the PESQ when compared with the original sound. The horizontal axis represents the reverberation time (ms), and the vertical axis represents the PESQ value. As the reverberation time increases, the PESQ value of the signal before processing decreases. In both the conventional method and the method of the present invention, the lowered PESQ value is improved by about 0.1 in data with a long reverberation time. The amount of improvement is comparable between the conventional method and the method of the present invention.

  On the other hand, comparing the calculation time, the conventional method required about 6 s to process 20 s of speech. The calculation time is the time processed under the condition of Core-i7, 3.4 GHz. In the method of the present invention, the calculation was completed in about 100 ms to process the same 20 s voice. Thus, it was confirmed that the same level of PESQ can be obtained with a calculation amount of about 1/60 of the conventional method, which has a significantly short calculation time.

  Thus, by approximating the indoor impulse response in the region of the power spectrum and using sequential learning with an adaptive filter, it is possible to realize dereverberation suppression performance that is the same as the conventional method with a small amount of computation.

  When the processing means in the above apparatus is realized by a computer, the processing contents of the functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only) Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording media, MO (Magneto Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  This program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

Claims (12)

An FFT unit that converts the collected sound signal into a frequency domain sound collected signal in the frequency domain;
With the frequency domain collected signal as an input, a power calculation unit that calculates and outputs a power spectrum for each frequency of the frequency domain collected signal; and
A delay unit that outputs a delayed power spectrum obtained by delaying the power spectrum by a predetermined delay amount;
From the power spectrum, a subtracting unit for subtracting the filtered signal obtained by multiplying the delayed power spectrum by a filter coefficient to obtain the estimated power of the direct sound;
An adaptive algorithm unit that receives the delay power spectrum and the estimated power of the direct sound as inputs, and updates a filter coefficient so as to minimize the estimated power of the direct sound;
A frequency domain filter unit that outputs a filtered signal obtained by multiplying the delayed power spectrum by the filter coefficient;
A reverberation suppression gain calculator for calculating a reverberation suppression gain for suppressing the reverberation sound, using the power spectrum and the estimated power of the direct sound as inputs,
A multiplier that multiplies the frequency domain collected signal by the reverberation suppression gain and outputs a reverberation suppression signal;
An inverse FFT unit for converting the dereverberation signal into a dereverberation signal in the time domain;
Equipped with,
The adaptive algorithm part is
An update vector calculation means for calculating an update vector using an adaptive algorithm, using the delay power spectrum and the estimated power of the direct sound as inputs,
The signal processing frame number is t, the frequency is ω, the filter order is m, the update vector value is U (ω, m, t), the step size is α,

Step size setting means and
Multiplication means for outputting a filter coefficient obtained by multiplying the update vector by the step size;
Adding means for adding the filter coefficient of the previous frame to the filter coefficient output by the multiplication means;
Filter coefficient holding means for holding the output of the adding means as a filter coefficient of the next frame;
A dereverberation device comprising: In the dereverberation device according to claim 1 ,
The reverberation suppression gain calculation unit is
Masking level calculation means for calculating auditory masking level Q (ω, t) from the estimated power of the direct sound;
A maximum value selecting means for comparing the estimated power of the direct sound and the auditory masking level Q (ω, t), and selecting a larger value as a new direct sound power spectrum in the calculation of the reverberation suppression gain;
A dereverberation apparatus comprising: In the dereverberation device according to claim 1 or 2 ,
Furthermore,
An adaptive interval for controlling the update of the filter coefficient by the adaptive algorithm unit only when the magnitude of the delay power spectrum is compared with a preset threshold and the magnitude of the delay power spectrum exceeds the threshold A dereverberation apparatus comprising a detection unit. In the dereverberation device according to any one of claims 1 to 3 ,
Furthermore,
A band aggregating unit that aggregates the frequency division ω of the power spectrum and converts it to a smaller frequency division ω ′ as shown in the following equation:

Where Ω (ω ′) is a set of frequencies ω belonging to the group of ω ′,
A band expansion unit that converts the frequency division ω ′ of the reverberation suppression gain into the frequency division ω of the FFT unit as shown in the following equation;

A dereverberation apparatus comprising: In the dereverberation device according to any one of claims 1 to 4 ,
Furthermore,
The filter coefficient is F (ω, m, t) , each of F (ω, m, t ) is multiplied by a preset coefficient H (m), and the converted filter coefficient F ′ (ω, m, t) a coefficient multiplier for outputting t);
A second frequency domain filter unit that multiplies the delayed power spectrum by the converted filter coefficient F ′ (ω, m, t);
A second subtraction unit that outputs a signal obtained by subtracting the output signal of the second frequency domain filter unit from the power spectrum to the reverberation suppression gain calculation unit instead of the estimated power of the direct sound;
A dereverberation apparatus comprising: An FFT process for converting the collected sound signal into a frequency domain sound collected signal in the frequency domain;
Power calculation process of calculating and outputting a power spectrum for each frequency of the frequency domain sound collection signal, using the frequency domain sound collection signal as an input,
A delay process for outputting a delayed power spectrum obtained by delaying the power spectrum by a predetermined delay amount;
From the power spectrum, subtracting the filtered signal obtained by multiplying the delayed power spectrum by a filter coefficient to obtain the estimated power of the direct sound; and
An adaptive algorithm process that takes the delayed power spectrum and the estimated power of the direct sound as inputs and updates the filter coefficients to minimize the estimated power of the direct sound;
A frequency domain filtering process for outputting a filtered signal obtained by multiplying the delayed power spectrum by the filter coefficient;
A reverberation suppression gain calculation process for calculating a reverberation suppression gain for suppressing the reverberation sound, using the power spectrum and the estimated power of the direct sound as inputs.
A multiplication process for multiplying the frequency domain collected signal by the reverberation suppression gain and outputting a reverberation suppression signal;
An inverse FFT process for converting the dereverberation signal into a time domain dereverberation signal;
Equipped with a,
The adaptive algorithm process is
An update vector calculation step for calculating an update vector using an adaptive algorithm, using the delay power spectrum and the estimated power of the direct sound as inputs;
The signal processing frame number is t, the frequency is ω, the filter order is m, the update vector value is U (ω, m, t), the step size is α,

Step size setting step and
A multiplication step of outputting a filter coefficient obtained by multiplying the update vector by the step size;
An addition step of adding the filter coefficient of the previous frame to the filter coefficient output in the multiplication step;
A filter coefficient holding step of holding the output of the addition step as a filter coefficient of the next frame;
A reverberation suppression method characterized by comprising: The dereverberation method according to claim 6 , wherein
The above reverberation suppression gain calculation process is:
A masking level calculating step for calculating an auditory masking level Q (ω, t) from the estimated power of the direct sound;
A maximum value selection step of comparing the estimated power of the direct sound and the auditory masking level Q (ω, t), and selecting a larger value as a new direct sound power spectrum for the calculation of the reverberation suppression gain;
A reverberation suppression method characterized by comprising: The dereverberation method according to any one of claims 6 and 7 ,
Furthermore,
An adaptive interval for controlling the update of the filter coefficient by the adaptive algorithm process only when the magnitude of the delay power spectrum is compared with a preset threshold and the magnitude of the delay power spectrum exceeds the threshold A reverberation suppression method comprising a detection process. The dereverberation method according to any one of claims 6 to 8 ,
Furthermore,
A band aggregation process for aggregating the frequency division ω of the power spectrum and converting it to a smaller frequency division ω ′ as shown in the following equation:

Where Ω (ω ′) is a set of frequencies ω belonging to the group of ω ′,
A band expansion process for converting the frequency division ω ′ of the reverberation suppression gain into the frequency division ω of the FFT process as shown in the following equation:

A reverberation suppression method comprising: The reverberation suppression method according to any one of claims 6 to 9 ,
Furthermore,
The filter coefficient is F (ω, m, t) , each of F (ω, m, t ) is multiplied by a preset coefficient H (m), and the converted filter coefficient F ′ (ω, m, t) a coefficient multiplication process for outputting t);
A second frequency domain filtering process of multiplying the delayed power spectrum by the converted filter coefficient F ′ (ω, m, t);
A second subtraction process for outputting a signal obtained by subtracting the output signal of the second frequency domain filter process from the power spectrum to the reverberation suppression gain calculation process instead of the estimated power of the direct sound;
A reverberation suppression method comprising: A program for causing a computer to function as the dereverberation device according to any one of claims 1 to 5 . Computer readable recording medium having recorded one of the program set forth in claim 1 1.

Images