JP5466581B2 - Echo canceling method, echo canceling apparatus, and echo canceling program - Google Patents

Description

  The present invention relates to an echo canceling technique for suppressing an echo component caused by a received signal reproduced by a speaker from a collected signal collected by a microphone by multiplying a gain for each frequency.

  Some echo cancellation apparatuses have a two-stage configuration of linear echo cancellation by an adaptive filter and nonlinear echo suppression by amplitude spectrum control. An echo canceling device 10 described in Non-Patent Document 1 is known as a prior art of a two-stage echo canceling device. An outline of the echo canceling apparatus 10 will be described with reference to FIG.

  The received signal x (n) reproduced by the speaker 2 goes around the microphone 3 via the echo path 5. The echo canceling apparatus 10 suppresses echo components caused by the received signal x (n) reproduced by the speaker 2 from the collected signal y (n) collected by the microphone 3. Note that n is an integer representing time.

In this configuration, the adaptive filter unit 11 uses the received signal x (n) input from the receiving end 1 to eliminate the echo component from the collected sound signal y (n) by linear processing, and the residual echo signal d ₁ ( n). Further, in the frequency domain transform unit 13, the residual echo signal d ₁ (n) is converted from the current time n to L _{1 of} d ₁ (n), d ₁ (n−1),..., D ₁ (n−L + 1). One frame is converted into a frequency domain signal D ₁ (f, k). D ₁ (f, k) is a Fourier transform of the residual echo signal d ₁ (n), f is a discrete angular frequency, k is a frame time, and f is 1 when the Fourier transform length is F. It is an integer of F.

The noise suppression unit 15 suppresses a noise component included in the residual echo signal D ₁ (f, k), and obtains a noise removal signal D ₂ (f, k). The frequency domain converter 17 converts the received signal x (n) into a frequency domain signal X (f, k). Further, the residual echo suppression unit 18 uses the signal X (f, k) to suppress the residual echo component included in the noise removal signal D ₂ (f, k), and transmits the transmission signal D ₃ (f, k). ) In the time domain conversion unit 19, the transmission signal D ₃ (f, k) is converted into a transmission signal d ₃ (n) in the time domain and output to the transmission end 4.

Here, attention is focused on the part of echo suppression processing in the residual echo suppression unit 18. The residual echo suppression unit 18 obtains an echo suppression gain G (f, k) and multiplies D ₂ (f, k), which is an input signal of the residual echo suppression unit 18, by G (f, k) in the frequency domain. This suppresses the echo. Specifically, the echo suppression gain G (f, k) is set to G (f, k) = (| D ₂ (f, k) | ² − | Y ^ (f, k) | ² ) / | D ₂ (f , K) | ² (1)
Calculate as | · | Represents taking an absolute value. Further, the transmission signal D ₃ (f, k) is changed to D ₃ (f, k) = G (f, k) D ₂ (f, k) (2)
Calculate as
Y ^ (f, k) in equation (1) is a pseudo-residual echo.
E [| Y ^ (f, k) | ² ] = E [| H (f, k) | ² ] | X (f, k) | ² + βE [| Y ^ (f, k-1) | ² ] (3)
Asking. H (f, k) represents a pseudo residual echo path, and is obtained using the minimum value of the ratio of E [| X (f, k) | ² ] and E [| D ₂ (f, k) | ² ]. . E [•] represents taking a set average. β is a forgetting constant and is set to a value that matches the reverberation time.

  The amplitude spectrum control in the residual echo suppression unit 18 can eliminate the residual echo component that remains when the adaptive filter unit 11 cannot cancel the echo. However, unlike the adaptive filter unit 11, the transmitted voice that is not related to the echo is partially suppressed according to the echo suppression amount. As a result, there is a problem that the transmitted voice is distorted and difficult to hear.

Therefore, Non-Patent Document 1 proposes a method of setting the original sound addition rate 1-α as a method of reducing audio distortion. That is, instead of the expression (2), the transmission signal is represented by D ₃ (f, k) = (1−α) D ₃ (f, k) + αG (f, k) D ₂ (f, k) (4)
To reduce the influence of the echo suppression gain G (f, k). Here, the original sound addition rate α is a real number from 0 to 1.

Hanai Seiyu, Haneda Yoichi, Tanaka Masafumi, Sasaki Junko, Kataoka Akitoshi, “Acoustic Echo Canceller with Noise Suppression and Echo Suppression Functions”, IEICE Transactions A, 2004, Vol. J-87- A, No.4, pp.448-457

  If the original sound addition rate is increased and the echo suppression gain is decreased, the distortion of the sound is reduced, but the echo cancellation performance is deteriorated accordingly, and the two are in a trade-off relationship. The optimal original sound addition rate varies depending on the signal to be suppressed, but the original sound addition rate of the prior art is fixed, and it is not always possible to set a value according to the situation, and the optimal original sound addition rate cannot be set. There is.

  In the echo canceller, when the optimal original sound addition rate is set for the vowel part signal, the consonant part signal has a small original amplitude, and the frequency spectrum characteristics change due to suppression. It seems that there will be a negative effect of making mistakes. Hereinafter, a description will be given with reference to FIG. When the transmitted voice is a vowel, the original spectrum and the outline are not so different even when the transmitted voice is lost due to the residual echo suppression process in the signal (see FIG. 2A) in which the residual echo is superimposed on the transmitted voice. No (see FIG. 2B). When the transmission voice is a consonant with the same original sound addition rate, if the transmission voice is lost due to the residual echo suppression process in the signal in which the residual echo is superimposed on the transmission voice (see FIG. 2C), the original amplitude is small. In addition, since the characteristics of the frequency spectrum change due to the suppression (see FIG. 2D), the frequency spectrum is greatly different from the original spectrum, causing problems such as misunderstanding of another consonant.

  On the other hand, when the original sound addition rate is optimal for the signal of the consonant part, there arises a problem that sufficient echo canceling performance cannot be obtained in the vowel part.

In order to solve the above-described problem, the echo canceling technique according to the present invention converts a signal d (n) and a received signal x (n) obtained based on a collected sound signal into a frequency domain signal D () for each frame. f, k) and X (f, k), and using the signals D (f, k) and X (f, k), the echo suppression gain Gb ^ (f, k) is obtained and the signal D (f , K) using the signal D ′ (f, k) obtained by removing the echo component, it is determined whether the signal to be suppressed is a vowel or a consonant, and the signal to be suppressed is determined to be a vowel. In this case, γ ₂ is the relaxation coefficient β (k), and in other cases, γ ₁ is the relaxation coefficient β (k), and the signal D (f, k) and the echo suppression gain Gb ^ (f, k) The result obtained by subtracting the product of the signal D (f, k) and the relaxation coefficient β (k) from the product of the relaxation coefficient β (k) and adding it to D (f, k) is obtained. It performs processing as a second residual echo suppressed signal _D 3 (f, k) the calculated, converted second residual echo suppressed signal _D 3 (f, k) to the signal _d 3 (n) of the time domain . Here, n represents time, f = 1, 2,..., F represents discrete angular frequency, k represents frame time, and γ ₁ <γ ₂ .

  The present invention produces an effect of simultaneously reducing sound distortion while changing the magnitude of the echo suppression gain according to the situation and sufficiently suppressing the echo.

The block diagram for demonstrating the conventional echo cancellation apparatus 10. FIG. 2A shows a signal in which the residual echo is superimposed on the transmission voice when the transmission voice is a vowel, FIG. 2B shows a signal after the residual echo suppression processing is performed on the signal in FIG. 2A, and FIG. 2C shows the transmission voice. FIG. 2D is a diagram illustrating a signal obtained by performing a residual echo suppression process on the signal illustrated in FIG. 2C. 1 is a block diagram for explaining an echo cancellation apparatus 100 according to a first embodiment. The figure for demonstrating the processing flow of the echo cancellation apparatus 100 of Example 1. FIG. FIG. 3 is a block diagram for explaining an adaptive filter unit 11 of the echo canceling apparatus 100 according to the first embodiment. FIG. 3 is a block diagram for explaining a noise suppression unit 15 of the echo canceling apparatus 100 according to the first embodiment. FIG. 3 is a block diagram for explaining a first residual echo suppression unit 130, a vowel consonant determination unit 140, a relaxation coefficient determination unit 150, and a second echo suppression unit 160 of the echo cancellation apparatus 100 according to the first embodiment. The figure for demonstrating the processing flow of the 1st residual echo suppression part 130 of the echo cancellation apparatus 100 of Example 1. FIG. The figure for demonstrating the processing flow of the vowel consonant determination part 140 of the echo cancellation apparatus 100 of Example 1, the relaxation coefficient determination part 150, and the 2nd echo suppression part 160. FIG. FIG. 3 is a block diagram for explaining a relaxation coefficient determination unit 150 of the echo cancellation apparatus 100 according to the first embodiment. FIG. 11A shows a second residual echo suppressor 160a for calculating the formula D ₃ (f, k) = {1-β (k) (1-Gb ^ (f, k))} D ₂ (f, k). 11A is used to calculate the equation D ₃ (f, k) = (1−β (k)) D ₂ (f, k) + β (k) D ′ ₃ (f, k). The block diagram for demonstrating the 2nd residual echo suppression part 160b. The block diagram for demonstrating the relaxation coefficient determination part 250 of the echo cancellation apparatus 200 of Example 2. FIG. The figure for demonstrating the processing flow of the relaxation coefficient determination part 250 of the echo cancellation apparatus 200 of Example 2. FIG. FIG. 10 is a diagram for explaining a relaxation coefficient determination unit 250 of the echo canceling apparatus 200 according to the second embodiment. FIG. 10 is a block diagram for explaining a relaxation coefficient determining unit 350 of the echo canceling apparatus 300 according to the third embodiment. The figure for demonstrating the processing flow of the relaxation coefficient determination part 350 of the echo cancellation apparatus 300 of Example 3. FIG.

Hereinafter, embodiments of the present invention will be described in detail.

<Echo canceling apparatus 100>
The echo canceling apparatus 100 suppresses the echo component caused by the received signal x (n) reproduced by the speaker 2 from the collected signal y (n) collected by the microphone 3 by multiplying the echo suppression gain for each frequency. .

For example, as shown in FIG. 3, the echo cancellation apparatus 100 includes an adaptive filter unit 11, frequency domain conversion units 13 and 17, a noise suppression unit 15, a time domain conversion unit 19, a first residual echo suppression unit 130, and a vowel consonant determination. Unit 140, relaxation coefficient determination unit 150, and second residual echo suppression unit 160. The echo canceling apparatus 100 according to the first embodiment will be described with reference to FIGS. 3 and 4. In FIG. 3, parts corresponding to those in FIG. The same applies to the following figures.
<Adaptive filter unit 11>
The adaptive filter unit 11 uses the received signal x (n) input from the receiving end 1 to cancel the echo component from the collected sound signal y (n) input from the microphone 3 by linear processing, and the residual echo signal d ₁ (n) is obtained (s11) and output to the frequency domain transform unit 13. For example, as illustrated in FIG. 5, the adaptive filter unit 11 includes an echo prediction unit 11a, a subtraction unit 11b, and an echo path estimation unit 11c.

  The echo prediction unit 11a receives the filter coefficient vector H ′ (n) and the received signal x (n), convolves them as in the following equation, obtains a pseudo echo signal y ′ (n), and subtracts this from the subtraction unit 11b. Send to.

y ′ (n) = H ′ ^T (n) X (n)
However,
H ′ (n) = [h ′ (n, 0)... H ′ (n, L−1)] ^T
X (n) = [x (n)... X (n−L + 1)] ^T
[] ^T represents vector transposition, L represents filter length, and h ′ (n, l) represents each filter coefficient.

The subtractor 11b receives the collected sound signal y (n) and the pseudo echo signal y ′ (n), subtracts the pseudo echo signal y ′ (n) from the collected sound signal y (n), and obtains a residual echo signal d ₁ (n ) (= Y (n) −y ′ (n)) is obtained and sent to the frequency domain transform unit 13 and the echo path estimation unit 11c.

The echo path estimation unit 11c receives the residual echo signal d ₁ (n) and the received signal x (n), and based on this, the error between the collected sound signal y (n) and the pseudo echo signal y ′ (n) is small. The filter coefficient vector H ′ (n) of the echo prediction unit 11a is updated so as to be sent to the echo prediction unit 11a. For example, the filter coefficient h ′ (n + 1) is updated using the NLMS (Normalized Least Mean Square) algorithm as shown in the following equation.

H ′ (n + 1) = H ′ (n) + (μd ₁ (n) X (n)) / (X ^T (n) X (n))
However, μ is a step size set to stabilize the estimation.
<Frequency domain conversion units 13 and 17>
The frequency domain transform unit 13 receives, for example, the residual echo signal d ₁ (n), and starts from the current time n to L 1 of d ₁ (n), d ₁ (n−1),..., D ₁ (n−L + 1). The minute is defined as one frame, and is converted into a frequency domain signal D ₂ (f, k) for each frame (s 13) and sent to the noise suppression unit 15. When the adaptive filter unit 11 is not provided in the echo canceling apparatus 100, the frequency domain conversion unit 13 may be configured to receive the collected sound signal y (n). For L, the number of samples corresponding to 10 ms or 20 ms is usually used.

The frequency domain converter 17 receives the received signal x (n), converts it into a frequency domain signal X (f, k) for each frame (s17), and sends it to the first echo suppressor 130. Examples of the conversion method include discrete Fourier transform (DFT) and short-time Fourier transform (STFT).
<Noise Suppression Unit 15>
The noise suppressing unit 15 receives the residual echo signal D ₁ (f, k) in the frequency domain, suppresses the noise component N (f, k) included in the signal D ₁ (f, k), and denoises the signal D. ₂ (f, k) is obtained (s15) and sent to the first residual echo suppressor 130 and the second residual echo suppressor 160. For example, as shown in FIG. 6, the noise suppression unit 15 includes a noise level estimation unit 15a, a noise suppression gain calculation unit 15b, and a multiplication unit 15c.

The noise level estimation unit 15a receives the signal D ₁ (f, k), and obtains a set average E [| N (f, k) | ² ] from the input signal D ₁ (f, k) in a section where no speech exists. . However, N (f, k) is a noise component included in the residual echo signal D ₁ (f, k).

The noise suppression gain calculation unit 15b receives the signal D ₁ (f, k) and the set average E [| N (f, k) | ² ], and calculates the noise suppression gain Ga ^ (f, k) by the following equation. Ask.

The multiplier 15c multiplies the residual echo signal D ₁ (f, k) by a noise suppression gain Ga ^ (f, k) to obtain a noise removal signal D ₂ (f, k). At that time, as shown in the following equation, the residual echo signal D ₁ (f, k) (original sound) is added to the noise removal signal D ₂ (f, k) at an appropriate ratio 1-α to mask the audio distortion. Thus, a configuration that suppresses deterioration in the audibility of the noise removal signal D ₂ (f, k) may be employed.
_{D 2 (f, k) =} (1-α) D 1 (f, k) + αGa ^ (f, k) D 1 (f, k)

<First Residual Echo Suppression Unit 130>
The first residual echo suppression unit 130 receives the noise removal signal D ₂ (f, k) and the received signal X (f, k), and uses this to determine the echo suppression gain Gb ^ (f, k), The signal D ₂ (f, k) is multiplied to obtain a first residual echo suppression signal D ′ ₃ (f, k) (s130). The first residual echo suppression unit 130 sends the first residual echo suppression signal D ′ ₃ (f, k) to the vowel consonant determination unit 140 and the echo suppression gain Gb ^ (f, k) to the second residual echo suppression unit 160. Send to.

  The first residual echo suppression unit 130 includes an echo suppression gain calculation unit 131 and a multiplication unit 135, for example, as shown in FIG. Furthermore, the echo suppression gain calculation unit 131 includes an acoustic coupling amount estimation unit 132, an echo level estimation unit 133, and a gain calculation unit 134. The processing of each unit will be described with reference to FIGS.

The acoustic coupling amount estimation unit 132 receives the noise removal signal D ₂ (f, k) and the reception signal X (f, k). The acoustic coupling amount estimation unit 132 sets the collective averages E [| D ₂ (f, k) | ² ], E [| X (f) of the noise removal signal D ₂ (f, k) and the received signal X (f, k). , k) | seeking ^2], _{respectively, E [| D 2 (f} , k) | 2], E [| X (f, k) | by updating the minimum value of the ratio of ^2, an acoustic coupling amount Frequency characteristic E [| H (f, k) | ² ] is obtained (s132) and sent to the echo level estimation unit 133.

The echo level estimation unit 133 receives the frequency characteristic E [| H (f, k) | ² ] of the acoustic coupling amount and the received signal X (f, k), and uses the pseudo residual echo Y ^ (f , K), the set average E [| Y ^ (f, k) | ² ] is obtained (s133) and sent to the gain calculation unit 134.

E [| Y ^ (f, k) | ² ] = E [| H (f, k) | ² ] | X (f, k) | ² + βE [| Y ^ (f, k-1) | ² ] (3)
Gain calculator 134, the pseudo residual echo Y ^ (f, k) and the noise cancellation signal _D 2 (f, k) receive, by the equation (1), the echo suppression gain Gb ^ (f, k) the calculated (s131 , S134), and sent to the multiplier 135 and the second residual echo suppressor 135.

G (f, k) = (| D ₂ (f, k) | ² − | Y ^ (f, k) | ² ) / | D ₂ (f, k) | ² (1)
The multiplication unit 135 multiplies the noise removal signal D ₂ (f, k) by the echo suppression gain Gb ^ (f, k) according to the equation (2) to obtain the first residual echo suppression signal D ′ ₃ (f, k). (S135) and sent to the vowel consonant determination unit 140.
D ′ ₃ (f, k) = G (f, k) D ₂ (f, k) (2)

<Vowel Consonant Determination Unit 140>
The vowel consonant determination unit 140 receives the first residual echo suppression signal D ′ ₃ (f, k) and uses it to determine whether the signal D ₂ (f, k) to be suppressed is a vowel or a consonant. Determine (s140). The vowel consonant determination unit 140 includes, for example, a determination evaluation value calculation unit 141 and a determination unit 143 as shown in FIG. The processing of each unit will be described with reference to FIGS.

The evaluation value calculator for determination 141 receives the first residual echo suppression signal D ′ ₃ (f, k), and uses the following formula to determine the sparsity of the spectrum of the first residual echo suppression signal D ′ ₃ (f, k). A value S (D ′ ₃ (k)) indicating is obtained (s141) and sent to the determination unit 143.

However, D ′ ₃ (k) is a vector notation of D ′ ₃ (f, k), and D ′ ₃ (k) = {D ′ ₃ (0, k), D ′ ₃ (1, k),. , D ′ ₃ (F, k)}, f _h represents the highest frequency considered, and f _l represents the lowest frequency considered. For example, 300 Hz to 3 kHz and audible range 20 Hz to 20 kHz used in voice call communication are set as the lowest frequency and the highest frequency. In this formula (5),

1 when the value of | D ′ ₃ (f, k) | at f _l ≦ f ≦ f _h is the most sparse (when only one frequency component has a value and the other frequency component is 0). √ (f _h −f _l +1) is taken when it is least sparse (when all frequency components have the same value). Therefore, 0 ≦ S (D ′ ₃ (k)) ≦ 1, and when D ′ ₃ (f, k) is a vowel spectrum, S (D ′ ₃ (k)) is a value close to 1 (FIG. 2B). In the case of a consonant, S (D ′ ₃ (k)) is a value close to 0 (see FIG. 2D).

Therefore, the determination unit 143 receives a value S (D ′ ₃ (k)) indicating sparsity, determines whether S (D ′ ₃ (k)) is equal to or greater than a predetermined threshold T, and is equal to or greater than the threshold T. Is determined as a vowel, and when it is less than the threshold T, it is determined as a consonant (s143). The determination unit 143 sends the determination result j (k) to the relaxation coefficient determination unit 150. The threshold T is 0 ≦ T ≦ 1, and is determined so that a vowel consonant can be determined in advance by experiments or the like (for example, T = 0.5). Further, in the determination result j (k), for example, 0 may be set as information indicating that it is a consonant, and 1 may be set as information indicating that it is a vowel.

Note that the first residual echo suppression signal D ′ ₃ (f, k) is used for the vowel consonant determination when the echo component derived from the received signal remains in the signal used for the determination. This is because of erroneous determination. Therefore, any signal from which the echo component is removed can be used for vowel consonant determination. The signal from which the echo component has been removed is, for example, a signal in which the adaptive filter unit 11 has eliminated the echo component by linear processing, or has been subjected to nonlinear echo suppression by the first residual echo suppression unit 130, or has undergone at least one processing If it is. Therefore, as shown by the long broken line in FIG. 7, the noise removal signal D ₂ (f, k) may be sent to the vowel consonant determination unit. However, since the residual echo component is included, the accuracy of the determination is lowered.

<Relaxation coefficient determination unit 150>
The relaxation coefficient determination unit 150 sets 1 as a relaxation coefficient β (k) when it is determined that the signal to be suppressed is a vowel, and otherwise sets γ as a relaxation coefficient β (k) (s150). ). However, γ is set to 0 ≦ γ <1, and an appropriate value is obtained in advance through an experiment or the like and determined in advance.
For example, the relaxation coefficient determining unit 150 includes storage units 151 and 153 and a switching unit 155 as illustrated in FIG. The processing of each unit will be described with reference to FIGS. 9 and 10. The relaxation coefficient determination unit 150 receives the determination result j (k). When j (k) is information indicating that it is a vowel, the switching unit 155 is connected to the storage unit 151. The relaxation coefficient determination unit 150 extracts 1 from the storage unit 151, determines β (k) = 1, and outputs the relaxation coefficient β (k) (s150, s151). When j (k) is information indicating that it is a consonant, the switching unit 155 is connected to the storage unit 153. The relaxation coefficient determining unit 150 extracts γ from the storage unit 153, determines β (k) = γ, and outputs the relaxation coefficient β (k) (s150, s153).

  In addition, the process of the determination part 143 of the vowel consonant determination part 140 and the relaxation coefficient determination part 150 can be represented with the following formula | equation.

<Second Residual Echo Suppression Unit 160>
For example, the second residual echo suppressor 160 receives D ₂ (f, k), Gb ^ (f, k), and β (k), and receives the second residual echo by the following equation. An echo suppression signal D ₃ (f, k) is obtained (s160) and sent to the time domain conversion unit 19.

D ₃ (f, k) = {1-β (k) (1-Gb ^ (f, k))} D ₂ (f, k) (7)
A configuration example of the second residual echo suppressing unit 160 at this time is shown in FIG. 11A. The process will be briefly described below. The subtraction unit 162a subtracts the received echo suppression gain Gb ^ (f, k) from the value 1 extracted from the storage unit 161a to obtain (1-Gb ^ (f, k)). The multiplication unit 163a multiplies this value by the relaxation coefficient β (k) to obtain β (k) (1−Gb ^ (f, k). The subtraction unit 165a calculates β (k from the value 1 taken out from the storage unit 164a. ) (1-Gb ^ (f, k)) is subtracted to obtain {1-β (k) (1-Gb ^ (f, k))}, which is multiplied by the noise removal signal D ₂ ( The second residual echo suppression signal D ₃ (f, k) is obtained by multiplying by f, k) and output.
With such a configuration, when the transmitted voice is determined to be a consonant, it is possible to reduce the loss of the frequency component of the consonant of the transmitted voice by weakening the echo suppression gain.

<Time domain conversion unit 19>
The time domain conversion unit 19 receives the second residual echo suppression signal D ₃ (f, k), converts it into a time domain signal d ₃ (n) (s 19), and sends it to the transmitting end 4. The conversion method may be an inverse Fourier transform or the like corresponding to the conversion method of the frequency domain conversion units 13 and 17.
[Program and recording medium]
The echo canceling apparatus described above can also be operated by a computer. In this case, each process of a program for causing a computer to function as a target device (a device having the functional configuration shown in the drawings in various embodiments) or a processing procedure (shown in each embodiment) is processed by the computer. A program to be executed by the computer may be downloaded from a recording medium such as a CD-ROM, a magnetic disk, or a semiconductor storage device or via a communication line into the computer, and the program may be executed.
<Effect>
With such a configuration, the relaxation coefficient (original sound addition rate) can be changed according to the situation, and there is an effect that the sound distortion is simultaneously reduced while sufficiently suppressing the echo. Therefore, it becomes easier to hear the voice as compared with the prior art.

  Since it is determined whether the signal to be suppressed is a consonant or a vowel, and the relaxation coefficient (original sound addition rate) is changed according to the determination result, if the signal to be suppressed is a consonant, the echo suppression gain is reduced to a small value, The distortion of the voice is reduced and the occurrence of listening errors is prevented. When the signal to be suppressed is a vowel, the echo suppression gain can be increased and sufficient echo cancellation performance can be obtained.

  That is, in the present embodiment, an appropriate echo suppression gain can be set for each time according to the nature of the voice, and the echo cancellation amount and the ease of listening to the voice can be balanced. As a result, it becomes easier to hear the voice in a hands-free call or the like.

  This relaxation of the echo suppression gain is effective for nonlinear suppression processing, and even if it is introduced to the adaptive filter unit 11 side, there is no speech distortion and only the echo cancellation amount is reduced. It is an effect. Although it is possible to introduce noise suppression, noise often has a broad spectrum close to the consonant of the speech, so that the noise is determined as a consonant and results in a reduction in noise suppression performance. Can't get.

[Modification]
When the input signal and the collected sound signal input to the echo canceling apparatus 100 are analog signals, the echo canceling apparatus 100 may include an A / D conversion unit (not shown) that converts the analog signal into a digital signal. When outputting an analog signal to the transmitting end 4, the echo canceling apparatus 100 may include a D / A conversion unit (not shown) that converts a digital signal into an analog signal.

  The adaptive filter unit 11 may eliminate the echo component using the frequency domain received signal X (f, k) and the collected sound signal Y (f, k). In that case, the frequency domain transform unit 13 is provided in the preceding stage of the adaptive filter unit 11. The adaptive filter unit 11 receives the output signals X (f, k) and Y (f, k) from the frequency domain transform units 13 and 17.

The second residual echo suppression unit 160 receives D ′ ₃ (f, k) instead of Gb ^ (f, k) as shown by a long broken line in FIG. The echo suppression signal D ₃ (f, k) may be obtained.

D ₃ (f, k) = (1-β (k)) D ₂ (f, k) + β (k) D ' ₃ (f, k) (8)
Incidentally, the equation (2), a _{D '3 (f, k)} = Gb ^ (f, k) D 2 (f, k). FIG. 11B shows the configuration of the second residual echo suppression unit 160 in this case. The multiplication unit 162b subtracts the received relaxation coefficient β (k) from the value 1 extracted from the storage unit 161b to obtain (1−β (k)). The multiplier 163 multiplies the received noise removal signal D ₂ (f, k) by this value (1-β (k)) to obtain (1-β (k)) D ₂ (f, k). The multiplier 164b multiplies the received D ′ ₃ (f, k) by the relaxation coefficient β (k) to obtain β (k) D ′ ₃ (f, k). The adder 165b adds (1−β (k)) D ₂ (f, k) and β (k) D ′ ₃ (f, k) to obtain the second residual echo suppression signal D ₃ (f, k). Is output.

Note that the configuration of the second residual echo suppression unit 160 is not limited to the configuration of FIGS. 11A and 11B, and the noise removal signal D ₂ (f, k), the echo suppression gain Gb ^ (f, k), and the relaxation. A product obtained by subtracting the product of the noise removal signal D ₂ (f, k) and the relaxation coefficient β (k) from the product of the coefficient β (k) and adding the subtraction result to D ₂ (f, k) is obtained. It suffices if the second residual echo suppression signal D ₃ (f, k) can be obtained by performing such processing.

The point of the present invention is that the vowel consonant determination unit 140 determines whether the signal to be suppressed is a vowel or a consonant, and changes the relaxation coefficient β (k) using the determination result. Therefore, as shown by a broken line in FIG. 4, the linear echo cancellation process (s11) in the adaptive filter unit 11 and the noise suppression process (s15) of the noise suppression unit 15 do not necessarily have to be performed, and corresponding units are not provided. Also good. Further, when a signal other than the first residual echo suppression signal D ′ ₃ (f, k) is sent to the vowel consonant determination unit 140, the first residual echo suppression process (s 130) in the first residual echo suppression unit 130 At least the echo suppression gain calculation unit 131 only needs to obtain the echo suppression gain (s131). As indicated by the broken line in FIG. 8, the multiplication processing by the multiplication unit 135 (s135) may not be performed, and the multiplication unit 135 is provided. Not necessary. Note that the processes in the adaptive filter unit 11, the noise suppression unit 15, the first residual echo suppression unit 130, and the vowel consonant determination unit 140 are examples, and other conventional techniques may be used.

For example, the determination evaluation value calculation unit 141 of the vowel consonant determination unit 140 may determine the sparsity of the spectrum of the first residual echo suppression signal D ′ ₃ (f, k) by the method described in Reference Document 1.
[Reference 1] Akiko Araki, Tomohiro Nakatani, Hiroshi Sawada, "Sound source number estimation and sound source separation based on speech sparsity using Dirichlet prior distribution", Acoustical Society of Japan 2009 Fall Meeting, 2009
In Reference Document 1, when the value of φ is smaller than 1, the Dirichlet distribution becomes larger as the vector α becomes sparse.

Further, the vowel consonant determination unit 140 does not use a value indicating the sparsity of the spectrum, for example, by the method described in Reference 2 or 3, the signal D ₂ (f, k) to be suppressed is a vowel or a consonant. You may determine whether there is.
[Reference 2] Hideyuki Sawada and Satoshi Okado, "Sensing of a specific speaker for voice interface construction under noisy environment", IEEJ Transactions, 2006, Vol.126, No.11, pp.1446- 1453
[Reference 3] Katsuyuki Niyada and Masakatsu Hoshimi, “Consonant recognition method for unspecified speakers using time series of band power and LPC cepstrum coefficient”, IEICE Transactions D, 1986, Vol. -D, No.6, pp.949-957
In this case, in Reference Document 2, a vowel consonant is determined based on the time average magnitude of the absolute value of the waveform, and in Reference Document 3, a power dip (consonant part) is extracted by looking at power fluctuations.

When the adaptive filter unit 11 or the like is not provided, the signal received by the frequency domain conversion unit 13 is a signal obtained based on the sound collection signal y (n) other than the residual echo signal d ₁ (n) (for example, collection). Sound signal y (n) itself, etc.).

The signals received by the first residual echo suppression unit 130 and the second residual echo suppression unit 160 are signals Y (f, k) and D ₁ (f) in the frequency domain other than the noise removal signal D ₂ (f, k). , K), and may be changed as appropriate according to the configuration of the echo canceller.

In the relaxation coefficient determination unit 150, β (k) = 1 or γ is set, but the present invention is not limited to this, and β (k) = γ ₁ (= αγ) or γ ₂ (= α) (provided that 0 The relaxation coefficient may be multiplied by a constant α as <α <1). The values of α and αγ are determined in advance as appropriate relaxation coefficients for vowels and appropriate relaxation coefficients for consonants by experiments or the like (for example, α = 0.5, γ = 0.5, γ ₁ = 0.25, γ ₂ = 0.5 etc.).

Further, γ ₁ , γ ₂ , and the relaxation coefficient β (k) may have different values for each frequency. At this _{time, γ 1 = {γ 1 (} 0), γ 1 (1), ..., γ 1 (F)}, γ 2 = {γ 2 (0), γ 2 (1), ..., γ 2 (F )}, Β (k) = {β (0, k), β (1, k),..., Β (F, k)}, and γ ₁ (f) ≦ γ ₂ (f), at least It is sufficient that γ ₁ (f ′) <γ ₂ (f ′) at some discrete angular frequencies f ′. With such a configuration, an appropriate relaxation coefficient can be set for each frequency. For example, since the consonant part increases as the frequency increases, a configuration in which the relaxation coefficient is set to be small can be considered.

<Echo canceling device 200>
Only a portion different from the first embodiment will be described with respect to the echo canceling apparatus 200 according to the second embodiment with reference to FIGS. The configuration and processing contents of the relaxation coefficient determination unit 250 are different from those in the first embodiment.

In addition to the determination result j (k), the vowel consonant determination unit 140 has a value S (D ′ ₃ (k) indicating the sparsity obtained by the evaluation value calculation unit 141 for determination, as indicated by a dashed line in FIG. ) Is also output to the relaxation coefficient determination unit 250.
<Relaxation coefficient determination unit 250>
The relaxation coefficient determination unit 250 receives the determination result j (k) and the value S (D ′ ₃ (k)) indicating sparsity. When j (k) is information indicating that it is a vowel, the switching unit 258 is connected to the storage unit 251. The relaxation coefficient determination unit 250 extracts 1 from the storage unit 251, determines β (k) = 1, and outputs the relaxation coefficient β (k) (s250, s251).

When j (k) is information indicating that it is a consonant, the switching unit 258 is connected to the adding unit 257. The relaxation coefficient determination unit 250 receives γ ₁ (k) = 1−κ (TS (D ′ (k))) from the addition unit 257, sets β (k) = γ ₁ and sets the relaxation coefficient β (k). The output is determined and output (s250, s257), where 0 ≦ κ ≦ 1 / T, and the relationship between S (D ′ (k)) and β (k) is shown in FIG.

Note that the subtraction unit 254 subtracts S (D ′ (k)) received from the threshold T extracted from the storage unit 254 to obtain (TS−D (k ′)). The multiplication unit 256 multiplies the value κ extracted from the storage unit 255 by (TS (D ′ (k))) to obtain κ (TS (D ′ (k)). The addition unit 257 stores the value. Κ (TS (D ′ (k))) is subtracted from the value 1 taken out from the unit 251, and γ ₁ (k) is obtained and stored.

  In addition, the process of the determination part 143 of the vowel consonant determination part 140 and the relaxation coefficient determination part 250 can be represented with the following formula | equation.

<Effect>
By adopting such a configuration, the same effects as those of the first embodiment are obtained. Furthermore, even within the range where S (D ′ (k)) <T, a signal having a very low sparsity is set to a small suppression, and a signal having a certain degree of sparsity is set to a large suppression. Setting is possible.

[Modification]
In the second embodiment, the relaxation coefficient β (k) is obtained for each case according to the relationship between the threshold value T and S (D ′ (k)). However, the relaxation coefficient β (k) is determined as S ( It may be a value that monotonously increases as D ′ (k)) increases.

As described above, since 0 ≦ S (D ′ ₃ (k)) ≦ 1, such a configuration can be realized by setting the threshold T = 1. Furthermore, the processing of the vowel consonant determination unit can be omitted and simplified. That is, in FIG. 7, the vowel consonant determination unit 140 includes only the evaluation value calculation unit 141 for determination, and outputs only S (D ′ (k)). In FIG. 12, the storage unit 251 and the switching unit 258 are not provided, and the relaxation coefficient determination unit 250 calculates and outputs β (k) = 1−κ (TS (D ′ (k))) for each frame. Even in such a configuration, the magnitude of the echo suppression gain can be changed according to the situation, and a signal with very low sparsity is small, and for signals with a certain degree of sparsity, A flexible setting such as setting suppression to a large value is possible.

Note that κ may have a different value for each frequency. At this time, κ = {κ (0), κ (1),..., Κ (F)}, and at least a part of the discrete angular frequency f ′ is 1−κ (f ′) (TS−D (D It suffices if “(k)) <γ ₂ (f ′). With such a configuration, β (k) is set to a different value for each frequency, and a finer relaxation coefficient can be set.

<Echo canceling device 300>
Only the parts different from the first embodiment of the echo canceling apparatus 300 according to the third embodiment will be described with reference to FIGS. The configuration and processing contents of the relaxation coefficient determination unit 350 are different from those in the first embodiment.
<Relaxation coefficient determination unit 350>
The relaxation coefficient determination unit 350 receives the determination result j (k), the received signal X (k), and the first residual echo suppression signal D ′ ₃ (k). When j (k) is information indicating that it is a vowel, the switching unit 356 is connected to the storage unit 354.

  The transmitted voice detection unit 351 and the determination unit 352 each receive the determination result j (k), and when j (k) is information indicating that it is a consonant, the following processing is performed.

First, in the transmitted voice detection unit 351, || D ′ ₃ (k) || / || (X (k) || is obtained. ||. || represents that the norm is taken, and X (k) = {X (0, k), X (1, k),..., X (F, k)}.

The determination unit 352 receives this value || D ′ ₃ (k) || / || (k) ||, determines whether or not it is smaller than the threshold value _Tr, and switches the determination result j ₂ (k). Output to the unit 356. In the case of information indicating that j ₂ (k) is smaller than the threshold value _Tr , the switching unit 356 connects to the storage unit 354 regardless of the value of the determination result j (k). The relaxation coefficient determining unit 350 extracts 1 from the storage unit 354, determines β (k) = 1, and outputs the relaxation coefficient β (k) (s350, s354). However, _Tr is a predetermined positive real number, and is a value for adjusting the relaxation coefficient to be 1 when the consonant part of the transmitted voice is sufficiently smaller than the received signal, such as an experiment. To obtain an appropriate value in advance. T _r is a value larger than 0, for example, T _r = 0.01.

Except for the above case, switching unit 356 receives information indicating that determination result j (k) is a consonant, and indicates that determination result j ₂ (k) is greater than threshold value _Tr. Is connected to the storage unit 355. The relaxation coefficient determination unit 350 takes out γ ₁ (0 ≦ γ ₁ <1) from the storage unit 355, determines the relaxation coefficient β (k) as β (k) = γ ₁ and outputs it (s350, s355). .

  In addition, the process of the determination part 143 of the vowel consonant determination part 140 and the relaxation coefficient determination part 350 can be represented with the following formula | equation.

<Effect>
By adopting such a configuration, the same effect as in the first embodiment can be obtained. Furthermore, when the transmission voice does not exist or the transmission voice is very small, the relaxation coefficient is set to 1 even if the first echo suppression signal D ′ ₃ (f, k) has sparsity. Sufficient echo cancellation can be performed without reducing the suppression gain. As described above, by reducing the gain using both the sparsity determination and the call state determination, it is possible to sufficiently suppress the echo in a section where there is no transmission voice that does not require suppression.

  The echo canceling method of the present invention can be used for hands-free calling, hands-free speech recognition, and the like.

100, 200, 300 Echo canceling device 11 Adaptive filter unit 13, 17 Frequency domain conversion unit 15 Noise suppression unit 19 Time domain conversion unit 130 First residual echo suppression unit 140 Vowel consonant determination unit 150, 250, 350 Relaxation coefficient determination unit 160 Second residual echo suppressor

Claims (9)

n represents time, f = 1, 2,..., F represents discrete angular frequency, k represents frame time, and γ ₁ <γ ₂ .
Frequency domain conversion step of converting the signal d (n) and the received signal x (n) obtained based on the collected sound signal into frequency domain signals D (f, k) and X (f, k) for each frame, respectively. When,
An echo suppression gain calculation step for obtaining an echo suppression gain Gb ^ (f, k) using the signals D (f, k) and X (f, k);
A vowel consonant determination step for determining whether a signal to be suppressed is a vowel or a consonant using a signal D ′ (f, k) obtained by removing an echo component from the signal D (f, k);
In the vowel consonant determination step, when it is determined that the signal to be suppressed is a vowel, γ ₂ is set as a relaxation coefficient β (k), and in other cases, γ ₁ is set as a relaxation coefficient β (k ) And a relaxation coefficient determination step
From the product of the signal D (f, k), the echo suppression gain Gb ^ (f, k), and the relaxation coefficient β (k), the signal D (f, k) and the relaxation coefficient β (k) A second residual echo suppression step for obtaining a second residual echo suppression signal D ₃ (f, k) by performing a process so as to obtain a result obtained by subtracting the product and adding the result to D (f, k);
A time domain conversion step of converting the second residual echo suppression signal D ₃ (f, k) into a time domain signal d ₃ (n);
An echo canceling method. The echo cancellation method according to claim 1 ,
In the vowel consonant determination step, a value S (D ′ (k)) indicating the sparsity of the spectrum of the signal using the signal D ′ (f, k) obtained by removing the echo component from the signal D (f, k). ), And when the value S (D ′ (k)) is greater than or equal to the threshold T, it is determined as a vowel, and when it is less than the threshold T, it is determined as a consonant.
An echo canceling method characterized by the above. The echo cancellation method according to claim 1 or 2,
In the relaxation coefficient determination step, when it is determined that the signal to be suppressed is a vowel, the γ ₂ = 1 is set as the relaxation coefficient β (k), and in other cases, the 0 ≦ γ ₁ <1 Is the relaxation coefficient β (k),
An echo canceling method characterized by the above. 3. The echo canceling method according to claim 2, wherein 0 ≦ S (D ′ (k)) ≦ 1, 0 ≦ T ≦ 1, 0 ≦ κ ≦ 1 / T,
In the relaxation coefficient determining step, when it is determined that the signal to be suppressed is a vowel, the γ ₂ = 1 is set as the relaxation coefficient β (k), and in other cases, the γ ₁ (k) = 1−κ (TS (D ′ (k)) is a relaxation coefficient β (f),
An echo canceling method characterized by the above. The echo cancellation method according to claim 1 or 2,
T _r is a predetermined positive real number,
When it is determined in the relaxation coefficient determination step that the signal to be suppressed is a vowel or when (|| D ′ (k) |||| (X (k) ||) < _Tr Γ ₂ = 1 is a relaxation coefficient β (k), and in other cases, 0 ≦ γ ₁ <1 is a relaxation coefficient β (k).
An echo canceling method characterized by the above. n represents time, f = 1, 2,..., F represents discrete angular frequency, k represents frame time,
Frequency domain conversion step of converting the signal d (n) and the received signal x (n) obtained based on the collected sound signal into frequency domain signals D (f, k) and X (f, k) for each frame, respectively. When,
An echo suppression gain calculation step for obtaining an echo suppression gain Gb ^ (f, k) using the signals D (f, k) and X (f, k);
A value S (D ′ (k)) indicating the sparsity of the spectrum of the signal D (f, k) using the signal D ′ (f, k) obtained by removing an echo component from the signal D (f, k). An evaluation value calculation step for vowel consonant determination to obtain
A relaxation coefficient determining step for increasing the relaxation coefficient β (k) as the value of S (D ′ (k)) increases;
From the product of the signal D (f, k), the echo suppression gain Gb ^ (f, k), and the relaxation coefficient β (k), the signal D (f, k) and the relaxation coefficient β (k) A second residual echo suppression step for obtaining a second residual echo suppression signal D ₃ (f, k) by performing a process so as to obtain a result obtained by subtracting the product and adding the result to D (f, k);
A time domain conversion step of converting the second residual echo suppression signal D ₃ (f, k) into a time domain signal d ₃ (n);
An echo canceling method. The echo canceling method according to any one of claims 1 to 6,
The γ ₁ , γ ₂ , β (k) can take different values for each frequency, and γ ₁ = {γ ₁ (0), γ ₁ (1),..., Γ ₁ (F)}, γ ₂ = {Γ ₂ (0), γ ₂ (1),..., Γ ₂ (F)}, β (k) = {β (0, k), β (1, k),. )}, And γ ₁ (f ′) <γ ₂ (f ′) at least at some of the discrete angular frequencies f ′.
An echo canceling method characterized by the above. n represents time, f = 1, 2,..., F represents discrete angular frequency, k represents frame time, and γ ₁ <γ ₂ .
Frequency domain converter for converting signal d (n) and received signal x (n) obtained based on the collected sound signal into frequency domain signals D (f, k) and X (f, k) for each frame, respectively. When,
An echo suppression gain calculator for obtaining an echo suppression gain Gb ^ (f, k) using the signals D (f, k) and X (f, k);
A vowel consonant determination unit that determines whether a signal to be suppressed is a vowel or a consonant using a signal D ′ (f, k) obtained by removing an echo component from the signal D (f, k);
When the vowel consonant determination unit determines that the signal to be suppressed is a vowel, the γ ₂ is set as a relaxation coefficient β (k); otherwise, γ ₁ is set as a relaxation coefficient β (k ) And a relaxation coefficient determination unit
From the product of the signal D (f, k), the echo suppression gain Gb ^ (f, k), and the relaxation coefficient β (k), the signal D (f, k) and the relaxation coefficient β (k) A second residual echo suppression unit that obtains a second residual echo suppression signal D ₃ (f, k) by performing a process such that a product is subtracted and added to D (f, k);
A time domain converter that converts the second residual echo suppression signal D ₃ (f, k) into a time domain signal d ₃ (n);
An echo canceling device. An echo canceling program for causing a computer to execute each step constituting the echo canceling method according to any one of claims 1 to 7.

Images