JP4631933B2 - Echo suppression method - Google Patents

Description

The present invention relates to an echo suppression technique in a situation where loudspeaking and sound collection are performed at the same time, such as a hands-free telephone and a voice recognition device while reproducing a car stereo.

FIG. 26 shows a configuration example of the prior art that suppresses echoes due to acoustic coupling between the sound collector and the loudspeaker in the hands-free telephone. In FIG. 26, the voice signal (referred to as a far-end signal) of the other party to be applied to the input terminal 10 is amplified as far-end voice from the speaker 2. A necessary sound, for example, a voice of a speaker (referred to as a near-end voice) enters the microphone 1, and a far-end voice amplified by the speaker 2 leaks through the space. This leakage is called echo. An acoustic transmission system including the speaker 2 and the microphone 1 from the far-end signal to the output signal of the microphone 1 is called an echo path.

It is only the near-end sound that is desired to be sent as a near-end signal from the output terminal 9, and it is desired to remove the echo of the far-end sound that has been loudened and leaked from the speaker 2. If the far-end speech leaks greatly, the far-end speech that has been delayed will be heard as an echo on the other side, making it difficult to speak at the far end. Conventionally, a technique that is generally applied to this problem is a technique that uses a linear echo canceller. The linear echo canceller is described in Non-Patent Document 1, for example.

In FIG. 26, the linear echo canceller 3 simulates the transfer function of the echo path. Using this simulated transfer function, a signal simulating an echo leaking into the sound received by the microphone 1, that is, an echo replica signal, is generated from an input signal (far-end signal) to the speaker 2. A near end signal is generated by subtracting from the received sound signal of 1. The voice detector 5 detects whether there is near-end voice by inputting the output of the microphone 1, the output of the linear echo canceller 3, the output of the subtractor 4, and the far-end signal, respectively. When there is sound, 0 or an extremely small value is output as the sound detection result, and when there is no near-end sound, a large value is output.

The operation of the linear echo canceller 3 will be described with reference to FIG. FIG. 27 is a diagram illustrating a configuration example of the linear echo canceller 3. The configuration example of the linear echo canceller illustrated in FIG. 27 includes an adaptive filter 30 and a multiplier 35. The adaptive filter 30 receives the far-end signal input from the terminal 31 and outputs the result of the linear filter from the terminal 32. Here, the filter coefficient for performing the linear filter operation in the adaptive filter 30 is updated every moment. The update is performed using a correlation operation so as to minimize the subtraction result received from the terminal 33. As a result, in the subtraction result applied to the terminal 33, the component having a correlation with the far-end signal is minimized. That is, the far-end signal echo is removed.

The multiplier 35 is inserted in the path from the terminal 33 to the adaptive filter 30 in order to control the amount of filter coefficient update in the adaptive filter 30. If there is no multiplier 35 and the filter coefficient is updated in the adaptive filter 30 when there is near-end speech, the filter coefficient is temporarily disturbed, and the amount of echo removal is reduced. The multiplier 35 sends the result obtained by multiplying the subtraction result received from the terminal 33 by the voice detection result from the voice detection unit 5 received from the terminal 34 to the adaptive filter 30. When there is near-end speech, the speech detection result is 0 or an extremely small value, so that update of the filter coefficient in the adaptive filter 30 is suppressed, and the filter coefficient is not disturbed. As a result, a high echo removal amount can be obtained.

As described above, the conventional linear echo canceller can remove the echo of the far-end signal using the linear adaptive filter. Various configurations such as an FIR type, an IIR type, and a lattice type can be used for the adaptive filter.

A paper by Eberhard HANSLER “The hands-free telephone problem: an annotated bibliography update” (Source: annals of telecommunications, 1994, p360-367) A paper by Xiaojiang Lu and Benoit Champagne “Acoustical Echo Cancellation Over A Non-Linear Channel” (Source: International Workshop on Acoustic Echo and Noise Control 2001) A. "A Speech Enhancement System Based On Negative Beamforming And Spectral Subtraction" by Alvarez et al. (Source: International Workshop on Acoustic Echo and Noise Control 2001) John J. Shynk's paper “Frequency-Domain and Multirate Adaptive Filtering” (Source: IEEE Signal Processing Magazine, January 1992, pp.14-37)

Also in the prior art, when nonlinear elements such as distortion are sufficiently small in the echo path, the echo can be sufficiently suppressed. However, in an actual device, nonlinear elements such as distortion in a speaker are large. The transfer function of the echo path including distortion is non-linear and cannot be completely simulated by a linear echo canceller. In particular, when a loud speaker is used with a small speaker used in a cellular phone or the like, since the distortion of the speaker is large, the echo is suppressed only by about 20 dB. In this case, since the echo is sent to the far end as a near end signal and can be heard by the far end speaker, it becomes difficult to speak.

An object of the present invention is to perform sound collection in which echo is sufficiently suppressed even when distortion in the echo path is large.

A first echo suppression method of the present invention is a method for suppressing echoes due to acoustic coupling between a sound collector and a loudspeaker. In this method, the output of the echo canceller is output from the output signal of the sound collector or the output signal of the sound collector. When any one of the subtracted signals is the first signal and the output signal of the echo canceller is the second signal, the echo leakage is estimated from the first signal and the second signal. A value is calculated, and the first signal is corrected based on the calculated estimated value.

According to a second echo suppression method of the present invention, in the first echo suppression method, echo leakage from the first signal and the second signal is performed for each frequency component of the first and second signals. An estimation value of the degree of dust is calculated, and a process of correcting the first signal based on the calculated estimation value is performed.

According to a third echo suppression method of the present invention, in the second echo suppression method, instead of using the output signal of the echo canceller as the second signal, a signal in which distortion of the output signal of the echo canceller is emphasized is used. The second signal is assumed.

The first echo suppressor of the present invention is an apparatus that suppresses echo due to acoustic coupling between a sound collector and a loudspeaker, and outputs an echo canceller from an output signal of the sound collector or an output signal of the sound collector. When any one of the subtracted signals is the first signal and the output signal of the echo canceller is the second signal, the echo leakage is estimated from the first signal and the second signal. A conversion unit including a calculation unit that calculates a value and a correction unit that corrects the first signal based on the calculated estimated value;

The second echo suppressor of the present invention is an apparatus that suppresses echoes due to acoustic coupling between the sound collector and the loudspeaker, and outputs the echo canceller from the output signal of the sound collector or the output signal of the sound collector. A dividing means for frequency-dividing the first and second signals when any one of the subtracted signals is a first signal and the output signal of the echo canceller is a second signal; Calculation means for calculating an estimated value of echo leakage from the first signal and the second signal for each frequency component of the second signal, and based on the estimated value calculated for each frequency component And a converting unit including correcting means for correcting the first signal for each frequency component and combining means for combining the first signal corrected for each frequency component.

The third echo suppressor of the present invention emphasizes the distortion of the output signal of the echo canceller by the distortion generator instead of using the output signal of the echo canceller as the second signal in the second echo suppressor. Let the signal be a second signal.

A fourth echo suppression device of the present invention is a device that suppresses echoes due to acoustic coupling between a sound collector and a loudspeaker, and includes a conversion domain echo canceller as an echo canceller, and the above-described echo signal is output from the output signal of the sound collector. A signal obtained by subtracting a signal obtained by subtracting the output signal of the transform domain echo canceller in the transform domain echo canceller is a first signal, and a signal for each frequency component before inverse linear transform in the transform domain echo canceller is a second signal. Calculation means for inputting each signal as a signal and calculating an estimated value of echo leakage from the first signal and the second signal for each frequency component of the first and second signals, and each frequency Correction means for correcting the first signal for each frequency component based on the estimated value calculated for each component, and a combining method for combining the first signal corrected for each frequency component Having a conversion unit including and.

A fifth echo suppressor of the present invention is a device for suppressing echoes due to acoustic coupling between a sound collector and a loudspeaker, wherein the output signal of the sound collector and the output signal of the loudspeaker are subband analysis filters. 1 is obtained by subtracting the output signal of the sound collector developed in each band or the output signal of the echo canceller from the output signal of the sound collector developed in each band. A calculation means for calculating an estimated value of echo leakage from the first signal and the second signal when the signal and the output signal of the echo canceller are the second signal, and the calculated estimated value And a subband synthesizing filter that synthesizes the first signal corrected by each converter.

A sixth echo suppressor of the present invention is an apparatus for suppressing echoes due to acoustic coupling between a sound collector and a loudspeaker, and the output signal of the sound collector and the output signal of the loudspeaker are converted by a Fourier transformer. Either the output signal of the sound collector developed in each band, or the signal obtained by subtracting the output signal of the echo canceller from the output signal of the sound collector developed in each band is the first signal. When the output signal of the echo canceller is the second signal, a calculation means for calculating an estimated value of echo leakage from the first signal and the second signal, and a calculated estimated value A conversion unit including a correction unit that corrects the first signal based on each band; and an inverse Fourier transformer that synthesizes the first signal corrected by each conversion unit.

Here, as the estimated value, the ratio of the amount according to the amplitude or power of the first signal to the amount according to the amplitude or power of the second signal in a period in which no near-end speech is detected, or this A value obtained by smoothing the ratio can be used, and the amplitude of the first signal with respect to the amount obtained by smoothing the amount according to the amplitude or power of the second signal in a period in which near-end speech is not detected. A value obtained by smoothing the ratio of the amount obtained by smoothing the amount corresponding to the power can also be used.

The smoothing time constant of the amount corresponding to the amplitude or power of the first and second signals is controlled to be smaller when the first and second signals increase than when the first and second signals decrease. You can do it. Further, the time constant for smoothing the ratio may be controlled to be long or infinite when near-end speech is detected, and to be shortened otherwise, and no near-end speech is detected. It is controlled so that it is much larger when it is detected, and when the near-end speech is not detected, it is smaller when the ratio increases than when it decreases. May be.

The correction may be performed by estimating the amount of echo included in the first signal from the estimated value and the second signal, and subtracting the estimated amount of echo from the first signal. It is good to estimate the ratio of the near-end signal included in the first signal from the estimated value and the first and second signals, and multiply the first signal by the estimated ratio. Also good.

If the echo canceller is a linear echo canceller, the harmonic component contained in the far-end signal appears almost as it is in the output of the echo canceller. Moreover, even if this echo canceller is a nonlinear echo canceller, not only a few harmonic components are included in the far-end signal. On the other hand, the near-end signal that is the input signal of the sound collector includes not only the near-end speech but also harmonics generated by the echo and nonlinear components of the far-end signal due to acoustic coupling between the sound collector and the loudspeaker. In the state where the component is included but there is no near-end speech, the near-end signal is almost only the harmonic component generated by the echo and the nonlinear component of the far-end signal. For this reason, the amount corresponding to the amplitude or power of the first signal (near-end signal itself or the output of the echo canceller is subtracted from the near-end signal) in a state where there is no near-end sound is set to the second signal (echo The quotient divided by the amount corresponding to the amplitude or power of the canceler output signal itself or the signal with enhanced distortion) is an estimate of the gain of the nonlinear component of the echo in the echo path, that is, an estimate of the echo leakage. Become. Accordingly, the amount of echo contained in the first signal is estimated from the estimated value and the second signal and subtracted from the first signal, or from the estimated value and the first and second signals. By estimating the ratio of the near-end signal included in the first signal and multiplying the first signal by this estimated ratio, the nonlinear component of the echo can be removed from the first signal.

According to the present invention, even when distortion in the echo path is large, sound can be collected with sufficiently suppressed echo.

Next, embodiments of the present invention will be described with reference to the drawings.

Referring to FIG. 1, an embodiment of the present invention uses a linear echo canceller (or a linear echo canceller) from an output signal of a microphone 1 or an output signal of a speaker 2 in order to suppress an echo due to acoustic coupling between the microphone 1 and the speaker 2. Any one of the signals obtained by subtracting the output signal 3 by the subtractor 4 may be the first signal, the output signal of the linear echo canceller 3, or the distortion included in the output signal (not shown). The near-end signal obtained by inputting the signal emphasized by the generation unit as the second signal and the voice detection signal of the voice detection unit 5 that detects the presence or absence of the near-end voice and removing the echo from the first signal. Is generated and output to the output terminal 9. Here, the conversion unit 100 calculates an estimated value of echo leakage from the first signal and the second signal, and corrects the first signal based on the calculated estimated value. This process is preferably performed for each frequency component of the first and second signals.

Referring to FIG. 2, the conversion unit 100 includes a frequency division unit 160 that divides the first signal applied from the input terminal 162 into M for each frequency, and a frequency division unit that divides the second signal applied from the input terminal 163 into M for each frequency. 161, M correction units 166m (m = 1 to M) provided for each frequency, and a frequency synthesis unit 164. The frequency division unit 160 and the frequency division unit 161 each include a first signal obtained by frequency division and The second signal is sent to the correction unit 166m corresponding to each frequency, and each correction unit 166m receives the voice detection result of the voice detection unit 5 input from the input terminal 167, the first signal in the frequency component, and the second signal. An estimated value of echo leakage is calculated from the above signal, and the first signal is corrected based on this estimated value and output to the frequency synthesizer 164. The frequency synthesizer 1 4 outputs the signal output from the correction unit 166m in frequency synthesis to the output terminal 165.

Here, the correction unit 166m uses the amplitude or power of the first signal with respect to the amount corresponding to the amplitude or power of the second signal in the period in which the near-end speech is not detected as the estimated value of the echo leakage. A ratio of quantities or a value obtained by smoothing this ratio is used. An amount corresponding to the amplitude or power of the first and second signals may be smoothed. The smoothing time constant of the amount corresponding to the amplitude or power of the first and second signals is controlled to be smaller when the first and second signals increase than when the first and second signals decrease. You can do it. The time constant for smoothing the ratio may be long or infinite when near-end speech is detected, and may be shortened otherwise, compared to when no near-end speech is detected. It is also possible to make the direction when it is detected much larger, and when the near-end speech is not detected, the time when the ratio increases is smaller than when the ratio decreases.

The correction unit 166m estimates the amount of echo included in the first signal from the estimated value of the echo leakage and the second signal, and subtracts the estimated amount of echo from the first signal. Therefore, the first signal may be corrected, or the ratio of the near-end signal included in the first signal is estimated from the estimated value of the echo leakage and the first and second signals. The first signal may be corrected by multiplying the first signal by the estimated ratio.

The frequency division in the frequency division units 160 and 161 can use any linear transformation such as Fourier transform, cosine transformation, and subband analysis filter bank, and the frequency synthesis in the frequency synthesis unit 164 is the inverse Fourier transform corresponding to them. Any processing such as inverse cosine transform, subband synthesis filter bank, etc. can be used.

Next, embodiments of the present invention will be described with reference to the drawings. First, an embodiment using a spectrum subtraction unit as the conversion unit 100 of FIG. 1 will be described.

FIG. 3 is a block diagram of the first embodiment of the present invention. 26 differs from the prior art in FIG. 26 in that a spectral subtraction unit 6 is inserted between the subtracter 4 and the near-end signal output terminal 9 in FIG. It is that the output signal of the canceller 3 is received and the sound detection result output from the sound detection unit 5 is received.

The spectral subtraction unit 6 develops the output signal of the subtractor 4 and the output signal of the linear echo canceller 3 in the frequency domain, and removes the echo for each frequency. The configuration example and operation will be described with reference to FIG.

FIG. 4 is a block diagram illustrating a configuration example of the spectrum subtraction unit 6. The output signal of the subtracter 4 in FIG. 3 is input from the input terminal 62. The Fourier transformer 60 receives the signal input from the input terminal 62, calculates its M-point Fourier transform, and uses the calculated result as the first Fourier coefficient, the Fourier coefficient subtractor 66m (corresponding to each frequency). m = 1 to M). On the other hand, an output signal of the linear echo canceller 3 in FIG. 3, that is, an echo replica signal is input from the input terminal 63. The Fourier transformer 61 receives the signal input from the input terminal 63, calculates its M-point Fourier transform, and uses the calculated result as the second Fourier coefficient to the Fourier coefficient subtractor 66m corresponding to each frequency. send.

In the Fourier coefficient subtractor 66m, the first Fourier coefficient output from the Fourier transformer 60, the second Fourier coefficient output from the Fourier transformer 61, and the voice detection result (terminal) output from the voice detector 5 in FIG. 67), the Fourier coefficient from which the echo component is removed is calculated, and the calculation result is sent to the inverse Fourier transformer 64. The inverse Fourier transformer 64 receives the Fourier coefficient group output from the Fourier coefficient subtraction units 661 to 66M, calculates the inverse Fourier transform, and outputs the real part of the calculation result from the output terminal 65.

Next, each configuration example and operation of the Fourier coefficient subtractor 66m (m = 1 to M) will be described with reference to FIG. FIG. 5 is a block diagram showing a first configuration example of the Fourier coefficient subtractor 66m (m = 1 to M). The first Fourier coefficient for each frequency output from the Fourier transformer 60 in FIG. 4 is sent to the absolute value calculator 701 and the subtractor 706 through the terminal 700. The second Fourier coefficient output from the Fourier transformer 61 in FIG. 4 is sent to the absolute value calculator 704 and the multiplier 707 through the terminal 703. The absolute value calculation unit 701 receives the first Fourier coefficient, calculates the absolute value, and sends the calculation result to the divider 702. The absolute value calculation unit 704 receives the second Fourier coefficient, calculates the absolute value, and sends the calculation result to the divider 702. The divider 702 receives the calculation result of the absolute value calculation unit 701 and the calculation result of the absolute value calculation unit 704, and calculates a value obtained by dividing the calculation result of the absolute value calculation unit 701 by the calculation result of the absolute value calculation unit 704. The calculation result is sent to the smoothing unit 705.

Here, the absolute value of the Fourier coefficient, that is, the absolute value of the amplitude of the frequency component is obtained by the absolute value calculation units 701 and 704, but the square root of the root mean square of the amplitude and the absolute value of the amplitude are also obtained. Any other amount can be used as long as the amount is proportional to the amplitude, such as the Q-th root of the Q-power average (Q is an arbitrary positive real number). Also, an amount proportional to power, for example, an even power such as a square can be used. The same applies to the other embodiments described below.

Smoothing section 705 smoothes the calculation result received from divider 702 and sends the result to multiplier 707. The smoothing time constant in the smoothing unit 705 is controlled by the sound detection result input from the terminal 67. When the voice detection result indicates that there is voice (near-end voice), the smoothing time constant is set to be long or infinite. When the speech detection result indicates that there is no speech (near-end speech), the smoothing time constant is shortened. The configuration of the smoothing unit 705 will be described later.

Multiplier 707 multiplies the value output from smoothing unit 705 by the second Fourier coefficient received from Fourier transformer 61 through terminal 703, and sends the result to subtractor 706. The subtractor 706 subtracts the value output from the multiplier 707 from the first Fourier coefficient received from the Fourier transformer 60 through the terminal 700 and outputs the calculation result from the terminal 799. The calculation result output through the terminal 799 is sent to the inverse Fourier transformer 64 in FIG.

Next, a configuration example and operation of the smoothing unit 705 will be described with reference to FIG. The input signal (the output of the divider 702) is sent to the subtracter 801 through the terminal 800. The subtractor 801 receives the input signal received through the terminal 800 and the output of the delay unit 804 for one sample (the output of the smoothing unit itself), and delays for one sample from the input signal received through the terminal 800. A signal obtained by subtracting the output of 804 is output and sent to the multiplier 802. Multiplier 802 receives the output signal of subtractor 801 and the output of smoothing coefficient determination unit 808, and sends the result of multiplying them to adder 803. The adder 803 receives the output of the multiplier 802 and the output of the delay unit 804, and sends the result of adding the two to the limiter 807. The limiter 807 receives the output of the adder 803, limits the upper limit and the lower limit so that the value does not exceed a certain range, and sends the limited result to the output terminal 899 and the delay unit 804. The delay unit 804 receives the output of the limiter 807, delays it by one sample, and sends the delay result to the subtracter 801 and the adder 803.

The smoothing unit 705 constitutes a so-called leak integrator or a so-called primary IIR type low-pass filter. However, the coefficient for determining the time constant is not a constant, but is supplied as a time-varying coefficient from the smoothing coefficient determination unit 808. Note that the smoothing coefficient and the smoothing time constant have an inversely proportional relationship.

The smoothing coefficient determination unit 808 supplies different smoothing coefficients depending on the sound detection result supplied from the terminal 67. If there is speech (near-end speech), a very small non-negative coefficient, for example, 0.0 is supplied. When there is no voice (near-end voice), a relatively large positive coefficient, for example, 0.005 is supplied. When there is speech (near-end speech) due to these time-varying smoothing coefficients, the smoothing coefficient one sample before is maintained or the smoothing coefficient hardly changes. Further, when there is no voice (near-end voice), the output smoothing coefficient changes.

By this control, the output of the smoothing unit 705 becomes an estimated value of echo leakage at that frequency. The reason will be described. Since the input of the smoothing unit 705 is a quotient obtained by dividing the Fourier coefficient of the near-end signal by the Fourier coefficient of the far-end signal, the Fourier coefficient of the near-end signal is obtained by multiplying this quotient by the Fourier coefficient of the linear echo canceller output signal. The same value as can be generated. That is, it is a value representing the degree of echo leakage. This value is a value that changes drastically, and when there is a voice (near-end voice), it is greatly disturbed and cannot be trusted as it is. However, smoothing provides a stable and reliable estimate. Further, by controlling the smoothing constant so that the change is slow or does not change when there is a voice (near-end voice), the influence of disturbance when there is a voice (near-end voice) can be reduced. As a result, the output of the smoothing unit 705 is an estimated value that is better when it is controlled by the sound detection result as an estimated value of echo leakage at that frequency than when it is not controlled.

The multiplier 707 multiplies the Fourier coefficient of the linear echo canceller output signal by the value output from the smoothing unit 705 obtained as described above, whereby the estimated value of the Fourier coefficient of the echo signal remaining in the echo replica signal is obtained. can get. By subtracting the estimated value of the Fourier coefficient of the echo signal from the Fourier coefficient of the near-end signal in the subtractor 706, an estimated value of the Fourier coefficient of the near-end signal with the echo component suppressed is obtained. The estimated value is sent to the inverse Fourier transformer 64 in FIG. 4, synthesized into a near-end signal, and output from the output terminal 65. As a result, echo is suppressed in this synthesized near-end signal.

The operation of the above Fourier coefficient subtractor 66m will be described using equations. Let S be the Fourier coefficient of the near-end signal, A of the near-end speech component is E, E is the echo component, and N is the noise component. here,
S = A + E + N (Formula 1)
There is a relationship. Also, let R be the Fourier coefficient of the far-end signal. Since there is a time axis, the phases of E and R should be almost the same. When there is no A, that is, when there is no near-end speech, the near-end signal is E + N, and all signals can be erased. Assume that E + N is estimated from R using the signal in this case, and E + N is subtracted from the near-end signal when there is near-end speech. The result P1 (corresponding to the output of the smoothing unit 705) obtained by smoothing the S / R only when there is no near-end speech using the speech detection result is:
P1 = Av [S / R] = Av [(E + N) / R] (Formula 2)
It becomes. Here, Av [•] represents smoothing. This P1 approximates how much of the far-end signal R leaks as an echo, and can be said to be an echo gain in the echo path.

Therefore, the result P2 (corresponding to the output of the multiplier 707) obtained by multiplying P1 by R is an estimated value of the echo component and the noise component.
P2 = P1 x R
= R × Av [(E + N) / R]
= Ex [E + N] (Formula 3)
Here, Ex [•] represents an estimated value.

The result P3 obtained by subtracting P2 from S (corresponding to the output of the subtractor 706) is:
P3 = S-P2
= S− (R × Av [(E + N) / R])
= (A + E + N) −Ex [E + N]
= Ex [A] (Formula 4)
It becomes. In P3, an estimated value of the Fourier coefficient component A of the near-end speech from which the echo component E and the noise component N are removed is obtained.

Returning to FIG. 3, how the embodiment of the present invention operates when the speaker in the echo path is distorted will be described. In the case where there is distortion, in the embodiment of the present invention, the component due to distortion in the echo is removed by nonlinear calculation in the frequency domain of the spectral subtraction unit 6. The linear echo canceller 3 adjusts the temporal change of the signal component, which is important in the frequency domain nonlinear calculation, thereby effectively removing the component due to distortion in the echo. The output signal of the microphone 1 includes a distortion echo of the far end signal in addition to the far end signal itself. This distortion echo can be considered as a harmonic of the frequency component of the far-end signal echo. To simplify the explanation, consider a case where the echo component E is only a harmonic due to distortion. As can be seen from Equations 2 and 3, the spectral subtraction unit can theoretically remove the echo component E as long as the Fourier transform component R of the far-end signal is not zero. Here, what is important for removing the echo component E is the estimation accuracy of the echo gain P1 in the echo path. Since the amount of harmonics due to distortion varies greatly with time due to the nature of the far-end signal, such as amplitude, in order to obtain high accuracy in P1, the denominator on the right side of Equation 1, that is, the Fourier transform of the far-end signal. It is desirable that the time change of the component R is as similar as possible to the time change of the echo component E in the molecule. If these temporal changes are greatly different, high accuracy cannot be obtained as P1, and distortion echoes cannot be largely removed. In the embodiment of the present invention, since the value based on the output signal of the linear echo canceller 3 is used as R, the timings of temporal changes in E and R are substantially aligned. That is, it is close to a desirable state for obtaining high estimation accuracy in P1. Therefore, the spectral subtraction unit 6 can greatly suppress harmonics generated by distortion.

The first embodiment of the present invention also has an effect of removing residual echoes when the echo path estimation is incorrect in the linear echo canceller 3 of FIG. In the description of the previous paragraph, the case where the echo component E is only a harmonic due to distortion is considered for the sake of simplicity, but the echo component of the far-end signal that does not depend on distortion, that is, an echo component that is not a harmonic exists. The discussion is also the same, and it is possible to suppress echo components that are not harmonics. For example, the estimation of the echo path may be wrong, and the subtracter 4 in FIG. 3 may add the reverse without removing the echo. However, even in that case, since the component of the far-end signal is removed in the spectrum subtraction unit 6, the echo can be suppressed.

Further, when this effect of the two-input spectrum subtraction unit 6 is used, the amount of calculation can be reduced by reducing the number of taps of the linear echo canceller 3. As shown in FIG. 26, when only the linear echo canceller is used, the echo removal amount is reduced by reducing the number of taps. However, in the first embodiment of the present invention shown in FIG. 3, even if the number of taps is reduced, the decrease in the echo removal amount is small and a practical echo removal amount can be obtained.

As described above, in the first embodiment of the present invention, the combination of the linear echo canceller 3 and the nonlinear calculation in the frequency domain of the spectrum subtraction unit 6 compensates for weak points and improves performance. It has gained. That is, even if the echo cannot be sufficiently suppressed by the linear echo canceller 3 alone, such as when the echo path is distorted or the linear echo canceller 3 erroneously estimates the echo path due to the presence of the spectral subtraction unit 6, It can be greatly suppressed. In addition, by using the output of the linear echo canceller, it is possible to suppress harmonics caused by distortion by simple estimation of only the amplitude without worrying about the time lag that cannot be handled by the frequency domain calculation of the spectral subtraction unit. it can.

FIG. 7 is a second configuration example of the Fourier coefficient subtractor 66m (m = 1 to M) in the spectral subtraction unit 6 of FIG. The difference from the first configuration example of the Fourier coefficient subtracter shown in FIG. 5 is that the smoothing unit 710 is inserted in the signal path from the absolute value calculation unit 701 to the divider 702 and the absolute value calculation. The only difference is that a smoothing unit 711 is inserted in the signal path from the unit 704 to the divider 702.

A configuration example of the smoothing unit 710 and the smoothing unit 711 is shown in FIG. FIG. 8 is a so-called leak integrator or first-order IIR type low-pass filter. Since the operation is very similar to the configuration example of the smoothing unit 705 shown in FIG. 6, the operation will be described while comparing FIG. 8 with FIG. The only difference between FIG. 8 and FIG. 6 is that the smoothing coefficient supplied to the multiplier 802 is supplied as a smoothing constant at the terminal 806. In the configuration example of FIG. 6, the input signal is smoothed without being affected by the voice detection result, whereas the configuration example of FIG. 8 is influenced by the voice detection result.

In the second configuration example of the Fourier coefficient subtractor 66m (m = 1 to M) illustrated in FIG. 7, the smoothing unit 710 and the smoothing unit 711 smooth the two inputs of the divider 702. The quotient input to the smoothing unit 705 as the output of the device 702 is also smoothed. As a result, the estimated value that is the output of the smoothing unit 705 is more stable in the second configuration example in FIG. 7 than in the first configuration example in FIG. However, in both the first configuration example and the second configuration example, the function of obtaining an estimated value of echo leakage at the output of the smoothing unit 705 remains the same. Therefore, even when the second configuration example shown in FIG. 7 is used as the Fourier coefficient subtractor 66m (m = 1 to M), the effect of the present invention is the case where the first configuration example shown in FIG. 5 is used. Is obtained in the same way.

FIG. 9 shows another configuration example of the smoothing unit 710 and the smoothing unit 711 in FIG. The difference between FIG. 8 and FIG. 9 is that the smoothing coefficient supplied to the multiplier 802 is supplied from the smoothing coefficient determining unit 810 that receives the output signal of the subtractor 801. The smoothing coefficient determination unit 810 supplies a relatively large coefficient, for example, 0.01 when the output signal of the subtractor 801 is positive, that is, when the output increases, and the output value of the subtractor 801 is negative. If the input is smaller than the output and the output decreases, a relatively small coefficient, for example 0.001, is supplied. By these time-varying smoothing coefficients, the speed at which the value of the smoothing unit output terminal 899 increases, that is, the rising speed is high, and the decreasing speed, that is, the speed at which the output falls is low. The actual amplitude change of voice or music, that is, the envelope, often rises quickly and falls slowly. In the configuration example of the smoothing unit shown in FIG. 9, it is possible to accurately follow such an envelope, and estimation of an estimated value of echo leakage in the configuration example of the Fourier coefficient subtractor 66m shown in FIG. Accuracy can be improved.

FIG. 10 shows another configuration example of the smoothing unit 705 in FIGS. The difference between FIG. 6 and FIG. 10 is that the smoothing coefficient determining unit 809 that supplies the smoothing coefficient to the multiplier 802 receives the voice detection result that has passed through the terminal 67 and the output signal of the subtractor 801 as inputs. It is replaced by the conversion factor determination unit 809. The smoothing coefficient determination unit 809 supplies different smoothing coefficients according to the sound detection result supplied from the terminal 67 and the sign of the output signal of the subtractor 801. If there is speech, supply a very small non-negative coefficient, eg 0.0. When there is no voice (near-end voice) and the output signal of the subtractor 801 is positive, a relatively large coefficient, for example, 0.01 is supplied. When there is no sound and the output value of the subtracter 801 is negative, that is, when the input is smaller than the output, a relatively small coefficient, for example, 0.001 is supplied. By these time-varying smoothing coefficients, the speed at which the value of the smoothing unit output terminal 899 increases, that is, the rising speed is high, and the decreasing speed, that is, the speed at which the output falls can be reduced. It is possible to accurately follow the envelope in the actual signal, and the estimation accuracy of the estimated value of echo leakage in the configuration example of the Fourier coefficient subtractor 66m shown in FIGS. 5 and 7 can be improved. .

FIG. 11 is a block diagram of the second embodiment of the present invention. The difference from the first embodiment shown in FIG. 3 is that the output of the subtracter 4 is used in the first embodiment as a signal input to the spectrum subtraction unit 6 whereas the second embodiment uses the output. That is, the output signal of the microphone 1 is used. In the output of the subtractor 4, the main component of the echo is removed by the linear echo canceller 3, whereas in the output signal of the microphone 1, the echo is not removed. This difference is only the difference between whether the main component of the echo is removed by the linear echo canceller 3 and the subtractor 4 or by the spectral subtraction unit 6. The effect on distortion is the same as that of the first embodiment. Is exactly the same. Accordingly, even in the second embodiment, even when the echo is not sufficiently suppressed only by the linear echo canceller 3 such as when the acoustic system is distorted or when the linear echo canceller 3 erroneously estimates the echo path, the echo is increased. Can be suppressed.

In addition, as a structure of the spectrum subtraction part 6 in the 1st and 2nd Example, in addition to the structure example demonstrated so far, the structure of the spectral subtraction (Nonpatent literature 2), or in the nonpatent literature 3 It is also possible to use a configuration of spectral subtraction.

Next, an embodiment using a spectrum suppression unit as the conversion unit 100 of FIG. 1 will be described.

FIG. 12 is a block diagram of the third embodiment of the present invention. The only difference from the first embodiment in FIG. 3 is that the spectral subtraction unit 6 is replaced by a spectral suppression unit 7. The spectrum suppression unit 7 will be described with reference to the drawings.

FIG. 13 is a block diagram showing the configuration of the spectrum suppression unit 7. The output signal of the subtracter 4 in FIG. The Fourier transformer 70 receives the signal input from the input terminal 72, calculates its M-point Fourier transform, and uses the calculated result as the first Fourier coefficient, the Fourier coefficient multiplier 76m (corresponding to each frequency). m = 1 to M). On the other hand, the output signal of the linear echo canceller 3 in FIG. The Fourier transformer 71 receives the linear echo canceller output signal input from the input terminal 73, calculates its M-point Fourier transform, and uses the calculated result as a second Fourier coefficient, which is a Fourier coefficient corresponding to each frequency. Send to multiplier 76m. In the Fourier coefficient multiplier 76m, the first Fourier coefficient output from the Fourier transformer 70, the second Fourier coefficient output from the Fourier transformer 71, and the voice detection result output from the voice detection unit 5 in FIG. The Fourier coefficient with the echo component reduced is calculated, and the calculation result is sent to the inverse Fourier transformer 74. The inverse Fourier transformer 74 receives the Fourier coefficient group output from the Fourier coefficient multiplier 76m (m = 1 to M), calculates the inverse Fourier transform, and outputs the real part of the calculation result from the output terminal 75. The Fourier coefficient multiplier 76m (m = 1 to M) provides a signal with a reduced echo component at the output terminal 75.

Next, each configuration example and operation of the Fourier coefficient multiplier 76m (m = 1 to M) will be described with reference to FIG. FIG. 14 is a diagram illustrating a first configuration example of the Fourier coefficient multiplier 76m (m = 1 to M). The first Fourier coefficient for each frequency output from the Fourier transformer 70 in FIG. 13 is sent to the absolute value calculator 731 and the multiplier 737 through the terminal 730. The second Fourier coefficient output from the Fourier transformer 71 in FIG. 13 is sent to the absolute value calculation unit 734 through the terminal 733. The absolute value calculator 731 receives the first Fourier coefficient, calculates the absolute value, and sends the absolute value to the divider 742 and the divider 745. The absolute value calculation unit 734 receives the second Fourier coefficient, calculates the absolute value, and sends the calculation result to the divider 742 and the divider 745. The divider 742 receives the calculation result of the absolute value calculation unit 731 and the calculation result of the absolute value calculation unit 734 and divides the calculation result of the absolute value calculation unit 731 by the calculation result of the absolute value calculation unit 734. And the calculation result is sent to the smoothing unit 743.

Smoothing unit 743 smoothes the calculation result received from divider 742 and sends the result to multiplier 746. The smoothing time constant in the smoothing unit 743 is controlled by the sound detection result input from the terminal 67. When the voice detection result indicates that there is voice (near-end voice), the smoothing time constant is set to be long or infinite. When the speech detection result indicates that there is no speech (near-end speech), the smoothing time constant is shortened. The configuration example of the smoothing unit 743 can be represented in FIG. 6 or FIG. 10 described above, and the operation is also the same. The output value of the smoothing unit 743 is an estimated value of the ratio between the first Fourier coefficient and the second Fourier coefficient, and is an estimated value of the degree of echo leakage at the frequency.

The divider 745 receives the calculation result of the absolute value calculation unit 731 and the calculation result of the absolute value calculation unit 734, and divides the calculation result of the absolute value calculation unit 734 by the calculation result of the absolute value calculation unit 731. And the calculation result is sent to the multiplier 746.

Multiplier 746 receives the output of smoothing unit 743 and the output of divider 745, calculates a value obtained by multiplying them, and sends the calculation result to smoothing unit 747. The smoothing unit 747 receives the output of the multiplier 746, smoothes the output, and sends it to the subtracter 744. The subtractor 744 subtracts the calculation result of the smoothing unit 747 from 1.0 and sends the result to the multiplier 737. Multiplier 737 multiplies the value output from subtractor 744 by the value of the first Fourier coefficient received from Fourier transformer 70 through terminal 730 and outputs the calculation result from terminal 789. The calculation result output through the terminal 789 is sent to the inverse Fourier transformer 74 in FIG.

Here, the value output from the subtractor 744 will be described using equations. The result P4 obtained by dividing the entire second row of Equation 4 used in the description of the Fourier coefficient subtraction unit by S and smoothing is expressed as Equation 5. The right side of Equation 5 is nothing but the value output from the subtracter 744 in FIG.
P4 = Av [P3 / S]
= Av [1-{(R / S) × Av [(E + N) / R]}]
= 1 − Av [{(R / S) × Av [(E + N) / R]}]
(Formula 5)
P4 is the result of smoothing by dividing the entire third row of Equation 4 by S.
P4 = Av [{(A + E + N) −Ex [E + N]} / S]
= Av [Ex [A] / S]
= Ex [A / S]
(Formula 6)
It appears like Comparing Equation 6 with Equation 5, it can be seen that the output P4 of the subtractor 744 is an estimated value of the ratio of the near-end speech in the near-end signal.

By multiplying the output value of the subtractor 744 by the Fourier coefficient of the output signal of the subtractor 4 in the multiplier 737, that is, the signal whose echo is reduced by the linear echo canceller, a signal other than the echo signal in the near-end signal, That is, an estimated value of the Fourier coefficient of the near-end speech with the echo suppressed is obtained. The estimated value is sent to the inverse Fourier transformer 74 in FIG. 13, synthesized into a near-end signal, and output from the output terminal 75. As a result, echo is suppressed in this synthesized near-end signal.

A description will be given of how the third embodiment of the present invention operates when a speaker or the like in the echo path is distorted. As described using Equation 5 and Equation 6, P4 is an estimated value of the ratio of near-end speech in the near-end signal. In calculating this P4, P3 used in the first embodiment of the present invention is used. As already described in the first embodiment of the present invention, P3 is an estimated value of the Fourier coefficient component of the near-end speech, and includes not only the echo component and noise component but also the harmonic echo component generated by distortion. It has been removed. Accordingly, the proportion of harmonic echo components generated by distortion is also removed in P4, and the distortion echo component is suppressed in the Fourier coefficient obtained by multiplying P4.

As described above, also in the third embodiment shown in FIG. 12, the echo is sufficiently suppressed by the linear echo canceller 3 alone, such as when the echo path is distorted or when the linear echo canceller 3 erroneously estimates the echo path. Even if it is not possible, echo can be greatly suppressed.

FIG. 15 is a second configuration example of the Fourier coefficient multiplier 76m (m = 1 to M) in FIG. The difference from the first configuration example of the Fourier coefficient multiplier 76m shown in FIG. 14 is that the smoothing unit 740 is inserted in the signal path from the absolute value calculation unit 731 to the divider 742, and the absolute value. The only difference is that the smoothing unit 741 is inserted in the signal path from the calculation unit 734 to the divider 742. This difference is the same as the difference between the first configuration example of the Fourier coefficient subtracter shown in FIG. 5 and the second configuration example of the Fourier coefficient subtracter shown in FIG. 7, and the effect is also the same. Therefore, when the second configuration example shown in FIG. 15 is used as the Fourier coefficient multiplier 76m (m = 1 to M) in FIG.

FIG. 16 is a block diagram of the fourth embodiment of the present invention. The difference from the third embodiment shown in FIG. 12 is that the output of the subtracter 4 is used as a signal input to the spectrum suppression unit 7 in the third embodiment, whereas the fourth embodiment uses the output. That is, the output signal of the microphone 1 is used. This difference is the same as the difference between the first embodiment and the second embodiment, and the effect is the same as that of the third embodiment.

Next, an embodiment will be described in which a signal in which distortion of the output of the linear echo canceller 3 is emphasized is input as the second signal to the conversion unit 100 in FIG.

FIG. 17 is a block diagram of the fifth embodiment of the present invention. The only difference from the first embodiment shown in FIG. 3 is that a distortion generator 8 is inserted between the output of the linear echo canceller 3 and the input of the spectrum subtraction unit 6. The distortion generator 8 actively generates harmonics of the output signal of the linear echo canceller 3, that is, the frequency component of the far-end signal. Since distortion similar to distortion generated in the acoustic system is obtained by the distortion generation unit 8, when the harmonics included in the far-end signal are too small, echo suppression in the spectral subtraction unit 6 is facilitated, and distortion is generated. Higher echo suppression can be obtained than when there is no part 8.

FIG. 18 is a block diagram of the sixth embodiment of the present invention. The only difference from the second embodiment shown in FIG. 11 is that a distortion generator 8 is inserted between the output of the linear echo canceller 3 and the input of the spectrum subtraction unit 6. This difference is the same as the difference between the first embodiment and the fifth embodiment, and the effect is also the same.

FIG. 19 is a block diagram of a seventh embodiment of the present invention. The only difference from the third embodiment shown in FIG. 12 is that a distortion generator 8 is inserted between the output of the linear echo canceller 3 and the input of the spectrum suppression unit 7. This difference is the same as the difference between the first embodiment and the fifth embodiment, and the effect is also the same.

FIG. 20 is a block diagram of the eighth embodiment of the present invention. The only difference from the fourth embodiment shown in FIG. 16 is that a distortion generator 8 is inserted between the output of the linear echo canceller 3 and the input of the spectrum suppression unit 7. This difference is the same as the difference between the first embodiment and the fifth embodiment, and the effect is also the same.

Although the embodiments of the present invention have been described above, the present invention is not limited to the above embodiments, and various other additions and modifications can be made. For example, additions and changes can be made as follows.

In the above description, the case where the Fourier transform is performed for each sample in the spectrum subtraction unit and the spectrum suppression unit has been described. However, the processing may be performed in units of frames at regular intervals, not every sample. Processing with overlapping frames is also possible. At this time, it is possible to reduce the amount of calculation by using a technique such as overlap save or overlap add. Techniques such as overlap saving and overlap add are described in Non-Patent Document 4, for example.

In the above description, the case where the Fourier transform is performed in the spectrum subtraction unit and the spectrum suppression unit has been described. However, in addition to the Fourier transform, a linear transform such as a cosine transform or a filter bank may be used. It is also possible to perform processing after conversion to the subband region. In these cases, the Fourier coefficient subtractor and the Fourier coefficient multiplier may be read corresponding to the linear transformation to be used. For example, when cosine transform is used, a cosine coefficient subtracter and a cosine coefficient multiplier are provided. These operations are the same as in the case of the Fourier transform.

When a transform domain echo canceller is used as the linear echo canceller in the present invention and the transform domain is the same transform domain in the spectrum subtraction unit or the spectrum suppression unit, the amount of computation is reduced and the delay time associated with the computation is reduced. Can be shortened. Here, the transform domain echo canceller represents an echo canceller that performs an echo canceller operation in a transform domain developed by performing linear transform, and re-synthesizes the original region by inverse linear transform.

An example of using a Fourier transform domain echo canceller shown in Non-Patent Document 4 as a linear echo canceller will be described with reference to the drawings. FIG. 21 shows a ninth embodiment of the present invention. In the ninth embodiment of the present invention, the spectral subtraction is performed with the echo canceller in the Fourier transform region. The difference from the first embodiment shown in FIG. 3 is that the linear echo canceller 3 is realized by the echo canceller 13, the spectral subtraction unit 6 is replaced by the spectral subtraction unit 16, and 2 of the input signals to the spectral subtraction unit 16 One of them is that they are replaced with a transform domain signal group 1 and a transform domain signal group 2 output from the echo canceller 13.

FIG. 22 is a block diagram showing a configuration example of the echo canceller 13 in the ninth embodiment of the present invention. The far-end signal input from the terminal 31 is developed in the Fourier transform region by the Fourier transformer 35 and sent to the adaptive filter group 38 for each frequency. The subtraction result input from the subtracter 4 in FIG. 21 via the terminal 33 is developed in the Fourier transform region by the Fourier transformer 37 and sent to the multiplier 39m (m = 1 to M) for each frequency. It is done. Each of the multipliers 39m (m = 1 to M) multiplies the signal received from the Fourier transformer 37 by the voice detection result received from the terminal 34, and sends the result to the adaptive filter group 38. The adaptive filter group 38 is composed of M adaptive filters, and receives a signal group received from the Fourier transformer 35 and a signal group received from the multiplier 39m (m = 1 to M), and outputs a corresponding signal. To perform adaptive filter processing. Each filter result obtained by the adaptive filter processing is sent to the inverse Fourier transformer 36. The inverse Fourier transformer 36 collects the filter results obtained from the adaptive filter group 38, calculates the inverse Fourier transform thereof, and outputs it from the terminal 32. A signal output from the terminal 32 is an output as an echo canceller.

In addition to the output as an echo canceller, the echo canceller 13 outputs the signal group output from the Fourier transformer 37 as a transform domain signal group 1 from the vector type output terminal 41 for spectral subtraction, and the adaptive filter group 38 outputs The filtered result group is output from the vector type output terminal 42 as the transform domain signal group 2. Transform domain signal group 1 and transform domain signal group 2 are sent to spectrum subtraction unit 16 in FIG. The transformation domain signal group 1 is obtained by Fourier transforming the output signal of the subtracter 4 in FIG. Further, the transform area signal group 2 can be interpreted as a signal obtained by Fourier transforming a signal from the echo canceller 13 to the subtracter 4 in FIG.

The configuration and operation of the spectrum subtraction unit 16 will be described with reference to the drawings. FIG. 23 is a block diagram illustrating a configuration example of the spectrum subtraction unit 16. The difference from the configuration example of the spectral subtraction unit shown in FIG. 4 described in the first embodiment of the present invention is that two input signals are replaced by transform domain signal group 1 and transform domain signal group 2; That is, the Fourier transformer 60 and the Fourier transformer 61 in FIG. 4 are deleted. As described in the configuration example of the echo canceller shown in FIG. 22, the transform domain signal group 1 is a Fourier transform of the output signal of the subtractor 4 in FIG. 21, and the transform domain signal group 2 is an echo in FIG. 21. The signal from the canceller 13 to the subtracter 4 is Fourier-transformed. These are exactly the same as the two signals input to the Fourier coefficient subtractor 66m (m = 1 to M) in the spectral subtraction unit shown in FIG. Therefore, the spectral subtraction unit 16 shown in FIG. 23 can output the same signal as the spectral subtraction unit shown in FIG. Therefore, it can be seen that the ninth embodiment of the present invention shown in FIG. 21 has the same effect as that of the first embodiment of the present invention. The two input signals to the spectrum subtraction unit 16 are directly connected as the transform domain signal group 1 and the transform domain signal group 2 from the echo canceller 13, thereby reducing the Fourier transform inside the spectrum subtraction unit 16 and The effects of the invention can be obtained.

It is also possible to implement other embodiments of the present invention in the transformation domain so that the ninth embodiment of the present invention corresponds to the first embodiment of the present invention. In addition to the Fourier transform region, a cosine transform region or the like can be used.

When the subband region echo canceller shown in Non-Patent Document 4 is used as the linear echo canceller and the processing is performed after converting the spectrum subtraction unit or the spectrum suppression unit to the subband region, the connection of each processing and portion It is possible to omit the filter for converting to the subband region in FIG. This example will be described with reference to the drawings. FIG. 24 is a block diagram showing a tenth embodiment of the present invention. In the tenth embodiment of the present invention, the spectral subtraction is performed with the echo canceller in the subband region. First, the signal from the microphone 1 is developed into N bands in the subband analysis filter bank 91, and the far-end signal is developed into N bands in the subband analysis filter bank 92. Each developed band has an echo canceller section 93n, a subtractor 94n, a voice detection section 95n, and a spectral subtraction section 96n (where n = 1 to N). The outputs of the spectral subtraction unit 96n for each band are collected by the subband synthesis filter bank 99, converted back to the original signal region, and output as a near-end signal. The processing of the subtractor 94n, the voice detection unit 95n, and the spectral subtraction unit 96n (where n = 1 to N) in each band is different in the number of echo canceller taps and the scale of the Fourier transformer in the spectral subtraction unit. The rest of the operation is the same as that of the first embodiment of the present invention shown in FIG. In the tenth embodiment of the present invention, all processing is performed after conversion into the subband region, so the synthesis filter bank in the linear echo canceller 3 and the subband analysis filter bank in the spectral subtraction unit are omitted. Can be connected. In this case, the calculation amount corresponding to the omitted subband analysis filter bank and subband synthesis filter bank is reduced, and the delay time corresponding to the calculation is reduced. A configuration in which the spectrum subtraction unit in the tenth embodiment of the present invention is replaced with a spectrum suppression unit is also possible. A configuration in which the transform area is other than the subband area is also possible.

Other embodiments of the present invention may be implemented in the subband region so that the tenth embodiment of the present invention corresponds to the first embodiment of the present invention. In addition to the Fourier transform region, a cosine transform region or the like can be used.

FIG. 25 is a block diagram showing an eleventh embodiment of the present invention. In the eleventh embodiment of the present invention, echo cancellation and spectral subtraction are performed in the Fourier transform domain. First, the signal from the microphone 1 is developed into M bands in the Fourier transformer 191, and the far-end signal is developed into M bands in the Fourier transformer 192. Each developed band has an echo canceller 193m, a subtracter 194m, a voice detector 195m, and a Fourier coefficient subtractor 66m (where m = 1 to M). The output of the Fourier coefficient subtractor 66m for each band is collected by an inverse Fourier transformer 199, inversely transformed into the original signal domain, and output as a near-end signal. The processing of the subtractor 194m and the sound detection unit 195m (where m = 1 to M) in each band is different from the number of taps of the echo canceller, but other than that of the first embodiment of the present invention shown in FIG. Since the operation is similar, detailed description of the processing is omitted. Both the tenth embodiment of the present invention and the eleventh embodiment of the present invention perform processing in the conversion area, but the difference is that the number of bands M is large because the conversion areas are different, and FIG. In FIG. 25, the spectral subtraction unit 24 is replaced with a Fourier coefficient subtractor 66m. That is, the Fourier transformer and the inverse Fourier transformer in the spectrum subtraction unit are not necessary, and only the Fourier coefficient subtractor 66m performs an operation necessary for the spectrum subtraction. This is because in the eleventh embodiment of the present invention, since all processing has already been developed in a large number of bands in the Fourier transform region, it is not necessary to perform Fourier transform again to perform spectral subtraction. In the eleventh embodiment of the present invention, the amount of calculation corresponding to the omitted Fourier transformer and inverse Fourier transformer is reduced. Further, a configuration in which the Fourier coefficient subtracter in the eleventh embodiment of the present invention is replaced with a Fourier coefficient multiplier is also possible.

Other embodiments of the present invention can be implemented in the Fourier transform domain so that the eleventh embodiment of the present invention corresponds to the first embodiment of the present invention. In addition to the Fourier transform region, a cosine transform region or the like can be used.

In the above, the case of using a linear echo canceller has been described. However, in the case of using a nonlinear echo canceller instead of the linear echo canceller, when combined with a spectrum subtraction unit or a spectrum suppression unit, The effect is obtained.

As described above, the present invention has been described using a hands-free telephone as an application example. However, not only hands-free telephones but also sound collection in an environment where music is amplified from a speaker and echoes from a receiver in a handset are problematic. The present invention can also be used for such sound collection.

It is a block diagram which shows embodiment of this invention. It is a block diagram which shows the structural example of a conversion part. It is a block diagram which shows the 1st Example of this invention. It is a block diagram which shows the structural example of a spectrum subtraction part. It is a block diagram which shows the 1st structural example of a Fourier coefficient subtractor. It is a block diagram which shows the structural example of a smoothing part. It is a block diagram which shows the 2nd structural example of a Fourier coefficient subtractor. It is a block diagram which shows the structural example of a smoothing part. It is a block diagram which shows the structural example of a smoothing part. It is a block diagram which shows the structural example of a smoothing part. It is a block diagram which shows the 2nd Example of this invention. It is a block diagram which shows the 3rd Example of this invention. It is a block diagram which shows the structural example of a spectrum suppression part. It is a block diagram which shows the 1st structural example of a Fourier coefficient multiplier. It is a block diagram which shows the 2nd structural example of a Fourier coefficient multiplier. It is a block diagram which shows the 4th Example of this invention. It is a block diagram which shows the 5th Example of this invention. It is a block diagram which shows the 6th Example of this invention. It is a block diagram which shows the 7th Example of this invention. It is a block diagram which shows the 8th Example of this invention. It is a block diagram which shows the 9th Example of this invention. It is a block diagram which shows the structure of an echo canceller. It is a block diagram which shows the structural example of a spectrum subtraction part. It is a block diagram which shows the 10th Example of this invention. It is a block diagram which shows the 11th Example of this invention. It is a block diagram which shows the conventional echo removal method. It is a block diagram which shows the structure of a linear echo canceller.

Explanation of symbols

1: Microphone 2: Speaker 3: Linear echo canceller 4, 94n (n = 1 to N), 194m (m = 1 to M): Subtractor 5, 95n (n = 1 to N), 195m (m = 1 to 1) M): Speech detection unit 6, 96n (n = 1 to N): Spectrum subtraction unit 7: Spectrum suppression unit 30: Adaptive filter 38: Adaptive filter group 35, 37, 60, 61, 70, 71, 191, 192: Fourier transformers 36, 64, 74, 199: Inverse Fourier transformer 66m (m = 1 to M): Fourier coefficient subtractor 91, 92: Subband analysis filter bank 93n (n = 1 to N), 193m (m = 1 to M): Echo canceller unit 99: Subband synthesis filter bank 100: Conversion unit 160, 161: Frequency division unit 166m (m = 1 to M): Correction unit 164: Frequency synthesis 701, 704: Absolute value calculator 702: Divider 705: Smoother 706: Subtractor 707: Multiplier 710, 711: Smoother 76m (m = 1 to M): Fourier coefficient multipliers 731, 734: Absolute value Calculation units 742 and 745: Dividers 740, 741, 743 and 747: Smoothing unit 706: Subtractors 707, 737 and 746: Multiplier 8: Distortion generator

Claims (1)

In the method of suppressing echo due to acoustic coupling between the sound collector and the loudspeaker, either the output signal of the sound collector or the signal obtained by subtracting the output signal of the echo canceller from the output signal of the sound collector is the first. 1 signal and a signal in which distortion of the output signal of the echo canceller is emphasized as the second signal, the first signal and the second signal for each frequency component of the first and second signals. An echo suppression method, comprising: calculating an estimated value of echo leakage from a signal; and correcting the first signal based on the calculated estimated value.

Images