JP2000252884A - Adaptive filter learning system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prevent deterioration in echo cancellation in units of frequency by independently discriminating propriety of learning for each frequency, in a frequency region type echo canceller provided with an adaptive filter for each frequency. SOLUTION: The echo canceller of a frequency region type is provided with a double talk detection section 17. This double talk detection section 17 applies learning processing to a frame of an adaptive filter 12, on the basis of a reception signal, a transmission signal. a pseudo echo signal and a residual error signal with respect to the transmission signal. independently decides propriety of learning especially for each frequency and uses a group average for part of reference values at the decision. Thus, opportunities for learning by frequencies are increased and errors in learning is reduced to enhance performance of double talk detection by frequencies, thereby preventing deterioration in an echo cancellation amount in units of frequency.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an echo canceller for removing an echo component contained in a transmission signal in a voice communication device such as a digital car telephone, and more particularly to an echo canceller provided for each frequency. The present invention relates to an adaptive filter learning method used for a frequency-domain echo canceller.

[0002]

2. Description of the Related Art In recent years, in the field of computers and communications,
Digital signal processing (DSP) has attracted attention and has been applied to various fields. This digital signal processing can easily realize complicated processing such as arbitrary change of characteristic constants and adaptive processing which were difficult in analog processing.
In particular, it is used as a general-purpose technique in the field of audio processing and image processing.

For example, in a hands-free type telephone used in a car telephone or the like, during a hands-free call,
In some cases, the voice received from the speaker wraps around the microphone and is sent to the other party to generate an acoustic echo. An echo canceller generally called an echo canceller is used to cancel the acoustic echo caused by the wraparound of the received voice and maintain the communication quality.

Here, there is a problem in voice communication.
The echo, which is a feedback signal from the speaker to the microphone, can be canceled by using the correlation between the reception signal (sound supplied to the speaker) and the transmission signal (sound input from the microphone). Specifically, the regression path between the speaker and the microphone is connected to one FIR filter (FIR: finite impulse res
ponse, finite impulse response), and a method of estimating the coefficient of this filter has been put to practical use.

A practical filter coefficient estimating method uses a time-sequential adaptive filter ("Basics of Digital Signal Processing", published by Shigeo Tsujii, IEICE). However, in this time-domain echo canceller, since the amount of computation for learning the adaptive filter is large, in recent years, the frequency-domain echo canceller separates the signal in frame units and learns the adaptive filter for each frame. (“Unconstrained Frequency”
-Domain Adaptive Filter ",
IEEE trans. Vol. ASSP-30, N
o. 5, Oct. 1982).

The outline of the frequency-domain echo canceller system is as follows. First, an FFT (Fast Fourier Transform) is performed by dividing a reception signal and a transmission signal into frames, and an individual complex first-order adaptation is performed for each frequency component. A filter is prepared and learned once per frame, and echo is canceled for each frequency component. If the residual component is inversely transformed, a signal after echo removal can be obtained.

Learning of the adaptive filter cannot be performed while a signal other than the echo, that is, the voice of the speaker on the transmitting side is mixed in the transmitting signal. The function of determining whether learning is possible is called double talk detection. The determination of double talk is performed by comparing the power of the received signal, the transmitted signal, and the power of the echo residual component with a threshold. In general, double talk detection is performed for each sample in the case of a time domain type echo canceller, and is performed for each frame in the case of a frequency domain type echo canceller.

[0008]

Conventionally, in a frequency-domain echo canceller, whether or not learning is possible is determined on a frame-by-frame basis. However, an adaptive filter used in the frequency domain type is individual for each frequency such as a low band, a middle band, and a high band, so that learning is not always possible at all frequencies. For example, if learning is performed when the power of the frequency component of the received signal is small, the echo canceling amount may deteriorate only at that frequency. For this reason, it is desirable to determine whether or not learning is possible for each frequency.

The present invention has been made in view of the above points, and in a frequency-domain echo canceller provided with an adaptive filter for each frequency, the determination as to whether or not learning is possible is made independently for each frequency. It is another object of the present invention to provide an adaptive filter learning method for preventing deterioration of an echo canceling amount in frequency units.

[0010]

According to the adaptive filter learning method of the present invention, a received speech signal is frequency-converted in frame units, and a pseudo echo signal is generated from the received speech signal using a different adaptive filter for each frequency. An adaptive filter learning method used in a frequency-domain echo canceller that removes an echo component included in the transmission signal by subjecting the pseudo echo signal to inverse frequency conversion and subtracting from the transmission signal, wherein the reception signal includes: The adaptive filter performs learning processing on the frame based on the residual error signal between the transmission signal and the pseudo echo signal and the transmission signal, and determines whether or not to perform the learning independently for each frequency. It was done.

In this case, first, it is determined whether or not the frame is suitable for learning based on the power of the entire frame of the reception signal, the transmission signal, and the residual error signal, and then the frame is suitable for learning. In this case, it is also possible to determine whether or not the frequency can be learned from the frequency component power of the received signal, the transmitted signal, and the residual error signal for each frequency.

Further, the frequency component of the frame is divided into several groups, the power of the power of the frequency component of the residual error signal is averaged for each group, a threshold value set for each group, and the frequency of the reception signal are set. Whether the adaptive filter can learn the frame or not may be determined using the power of the component.

When the number of frequencies determined to be learnable is larger than a certain value, learning is performed for all the frequencies of the frame, and when the number of frequencies determined to be learnable is smaller than a certain value, the learning is performed. Learning may not be performed for all frequencies of the frame.

Further, the frequency components of the frame may be divided into several groups, and the determination result of the possibility of learning in each group may be the same.

As described above, by independently determining whether or not learning is possible for each frequency, it is possible to perform optimal learning processing according to each frequency, and to prevent deterioration of the echo cancellation amount in frequency units. .

When it is determined that learning is possible at most of the frequencies, learning is performed for each frame in order to increase the speed of learning as a whole. Conversely, when the number of frequencies determined to be learning is small, Since the possibility of erroneous learning is high, learning cannot be performed for each frame, so that the learning processing of the entire frame can be performed in consideration of the correlation between the frequencies.

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a frequency-domain echo canceller according to an embodiment of the present invention.

The echo canceller includes a frame generator 10 for a received signal, a Fourier transformer (FFT) 11, an adaptive filter (AF) 12, and an inverse Fourier transformer (IFF).
T) 13, a transmission signal frame creation unit 14, a frame synthesis unit 15, a Fourier transform unit 16, and a double talk detection unit (DTD) 17.

A feature of the present invention is a double talk detecting unit 17 for determining whether or not learning of an adaptive filter 12 (filter coefficient) provided for each frequency is performed, and other components of the echo canceller (10 to 10). For 16), an existing method may be used, and a detailed description will be omitted. In the present embodiment, an example will be described using FFT-EC of a document cited as a conventional method.

The reception signal, which is the voice signal of the communication partner, is
First, the frame creation unit 10 divides the frame into 2N-length frames xk (0),..., Xk (2N-1). k is the frame number, 2N is the transform order of the FFT, and the received signal is overlapped by N samples to create a frame.

Next, this frame is converted to a Fourier transform unit 11.
, And frequency component Xk of the received signal.
(0),..., Xk (N-1) are obtained. Where Xk
(I) is a complex number, but the real part of Xk (0) is a DC component, and the imaginary part is a high-frequency component having a period of two samples. The filter component is applied to the frequency component Xk (i) of the received signal in the adaptive filter unit 12. As a result of multiplication of the Xk (i) and the filter coefficient, a pseudo echo component X'k (i)
Is generated. This pseudo echo component X'k (i) is inversely transformed in the inverse Fourier transform unit 13, and the pseudo echo x '
k (0),..., x′k (2N−1) are generated.

On the other hand, a transmission signal, which is a voice signal on the communication side, is sent to a frame forming unit 14 by a frame yk of length N.
, Yk (N-1). Then, in the frame synthesizing unit 15, the first N samples are 0, and the second N samples are the second half x'k of the pseudo echo from the transmission signal.
(N),..., X′k (2N−1), and a frame in which the value is subtracted. The frame is transformed by the FFT unit 16, and the residual error frequency components Ek (0),. −
Obtain 1).

Where Ek (i) is a complex number and Ek (0)
Is a DC component, and the imaginary part is a high-frequency component having two cycles.

Finally, for each frequency component i, the frequency component Ek (i) of the residual error and the frequency component Xk
The i-th filter coefficient of the adaptive filter 12 is updated using (i). In this case, the update of the filter coefficient is enabled by the flag mode (i) indicating whether or not learning is possible for the frequency i (=
This is performed only in the case of 1). The algorithm for updating the filter coefficients is well known in literatures and the like, and therefore, the description thereof is omitted here.

Flag mode (i) (i = 0,..., N)
Is determined by the double talk detecting unit 17. As described above, the double talk detecting unit 17 performs frequency conversion (FFT) on the received signal and the transmitted signal in the echo cancellation to generate a pseudo echo from the received signal and subtract the pseudo echo from the transmitted signal. When a pseudo echo component is generated using a different adaptive filter (filter coefficient), it is characterized in that whether or not the adaptive filter of each frequency can learn the frame is independently determined for each frequency.

Hereinafter, the processing operation of the double talk detecting unit 17 will be described.

FIGS. 2 to 4 are flowcharts showing the algorithm of the double talk detecting unit 17.

Here, for convenience, the flag for the DC component (real part of frequency 0) is mode (0), and the flag for the frequency component of two samples in the period (imaginary part of frequency 0) is m.
mode (N) is used. Other frequencies are frequency Nos. Corresponding to

In FIGS. 2 to 4, the definition of each symbol is as follows.

K; frame number N; frame length of transmission signal and 1/2 of FFT order (for example, 128) Xk (i); frequency component i of reception signal frame k (complex number) Ek (i); residual error Frequency component i of frame k of the signal
(Complex number) Note that when i = 0,..., N−1, i = 0, the real part of the frequency component is a DC component, and the imaginary part is the highest frequency. yk (i); simple value (real number) of frame k of transmission signal i = 0,..., N-1 pe; power of entire frame of residual error signal px; power of entire frame of reception signal py; Power of the entire frame tmp; S / N ratio t; variable threshold value of the frame flag; status flag of the frame c; number of frames V, a, b; constant (for example, V = 1024 * N, a = 1,
b = 1) α, β; constant (for example, α = 0.95, β = 0.25) L; update time (for example, 100) Pe (i); average power of received signal frequency components by group Px (i) Average power B, BN for each group of frequency components of the residual error signal; number of frequency groups, number of frequencies included in each group For example, B = 4, BN = N / 4, and the number of groups is 4
When the number of frequencies is 128, the number of frequencies included in each group is 128/4 = 32 Ca; the number of learnable frequencies M1, M0; a constant (M1> M0) T (i); Variable threshold value for each group, i = 0,.
B-1 mode (i); learning enable / disable flag for each frequency group,
i = 0,..., N A0, A1, A2, VB; constant (for example, A0 = 2, A1
= 1, A2 = 1) Note that R () and I () represent a real part and an imaginary part, respectively.

In the figure, section 1 is a process for determining whether or not learning is possible for a frame, section 2 is a frame learning disabling process when the adaptive filter is abnormal,
section3 is a process of determining whether or not learning is possible for all frequencies of the frame based on the determination result of the learning possibility of each frequency, section4 is a process of updating a variable threshold value for each frequency group, and section5 is a possibility of learning of each frequency. Is shown.

As shown in the flow charts of FIGS. 2 to 4, the double talk detecting unit 17 first executes step S1.
Then, each parameter required for the processing is cleared to 0, and in steps S2 to S16, a learning possibility determination process (section 1) for the frame is executed as follows.

That is, in step S2, the power pe of the entire frame of the residual error signal, the power px of the entire frame of the reception signal, and the power py of the entire frame of the transmission signal are calculated in accordance with a predetermined formula. The S / N ratio tmp between the received signal and the received signal is calculated as follows.

Tmp = log10 (pe / px) Here, the state in which learning is not possible is at the time of double talk and at the time of the talker talking. The state in which learning is possible is when the receiver (the other party) is talking.

Therefore, in step S2,
The power pe of the entire frame of the residual error signal, the power px of the entire frame of the received signal, the power py of the entire frame of the transmitted signal, and the S / S of the residual error signal and the received signal.
After obtaining the N ratio tmp, in the next step S3, it is determined whether or not the receiver (the other party) is in a talking state by comparing the power px of the entire reception signal frame with the constant V. As a result, the power p of the entire received signal frame
If the value of x is equal to or less than the value of the constant V (px ≦ V),
It is determined that the receiver (the other party) is not in a talking state, and "0" is set in the learning flag of the frame in step S4. Note that flag = 0 indicates that learning of the frame is not possible.

On the other hand, when the value of the power px is higher than the value of the constant V (px> V), that is, when it is determined that the receiver (the other party) is in a talking state, the transmission is performed in the next step S5. The power py of the whole frame of the speech signal and the constant V /
By comparing this with 2, it is determined whether or not the talker is in a talking state at that time. As a result, if the value of the power py of the entire frame of the transmission signal is higher than the value of the constant V / 2 (py> V / 2), it is determined that the transmitter is in a talking state. Note that the constant at this time was set to V / 2,
This is based on the assumption that the echo attenuation is 6 dB. For example, if the attenuation is 12 dB, it can be changed to 3 V / 4.

Here, if the talker is also in a talking state, it is a double talk and learning is not possible.
At 6, the S / N ratio tmp between the residual error signal and the reception signal
And the S / N ratio between the residual error signal and the transmission signal = log10
(Pe / py), and both are constants b
Is exceeded, that is, tmp> b & l
If og10 (pe / py)> b, it is determined that the adaptive filter 12 is not operating normally and is in an abnormal state (malfunction). When the adaptive filter 12 is in an abnormal state, not only the echo cannot be canceled but also the echo may be amplified. Therefore, when it is determined that the adaptive filter 12 is in the abnormal state, the learning flag of the frame is determined in step S7.
Is set to "-1". In addition, flag = -1
Indicates that learning is not possible due to a filter abnormality.

In step S6, the S / N ratio tmp between the residual error signal and the received signal and the S / N ratio between the residual error signal and the transmitted signal = log10 (pe / pe
py) If there is no problem in the S / N ratio, it is determined that double talk (a state in which both the transmitting side and the receiving side are talking at the same time). In the case of double talk, in step S4, "0" is set to the learning flag of the frame, and
Show that learning is not possible.

On the other hand, in step S3, the value of the power px of the entire frame of the received signal is higher than the value of the constant V (px> V), and in step S5,
The value of the power py of the entire frame of the transmission signal is a constant V / 2
Is smaller than or equal to (py ≦ V / 2), only the receiver (the other party) is determined to be in a talking state, and learning of the frame is enabled.

At this time, the S / N ratio tmp between the residual error signal and the received signal obtained in step S2 is checked, and if the S / N ratio tmp is equal to or larger than the value obtained by adding the constant a to the variable threshold value t of the frame. If there is (tmp ≧ t + a), step S9
Sets "2" to the learning flag of the frame. Also, the S / N ratio tmp is equal to the variable threshold value t of the frame.
Is smaller than the value obtained by adding the constant a to (tmp <t +
a) In step S10, the learning fla of the frame
Set “1” to g. flag = 1 or flag
= 2 indicates a state in which learning is possible. In addition, flag =
In the case of 2, the learning of the entire frame is dangerous, but it can be considered that learning is possible if only a part of the frequencies is possible.

Subsequently, in steps S11 to S16, a process of updating the variable threshold value t of the frame is performed.

In the process of updating the variable threshold value t, the variable threshold value t which is currently set is compared with the value of the S / N ratio tmp of the received error signal and the residual error signal obtained in step S2 to obtain a variable value. If the threshold value t is lower than the S / N ratio tmp (tmp <t), the variable threshold value t is updated in step S12 according to a predetermined calculation formula. Also, the S / N ratio tmp
Is smaller than 0 (tmp <0), step S
As shown in 14 to S16, when the update time L has elapsed, the variable threshold value t is updated according to a predetermined calculation formula.

Thus, the power px of the received signal and the power pe of the residual error (= power sum of frequency components),
When it is determined whether or not the frame k is suitable for learning from the power py of the transmission signal, a process of determining whether or not learning is possible for each frequency is performed.

That is, first, in step S17, while the value of n indicating the current frequency to be learned is cleared to 0, n is sequentially incremented from 1 to B (B; the number of frequency groups). In S10, the average power Pe (n) of the frequency components of the received signal for each group
And the average power Px (n) for each group of the frequency components of the residual error signal.

When the average powers Pe (n) and Px (n) of the frequency components of the received signal and the error signal are obtained for each group, fl is set in the frame unit in the next step S20.
Check ag. As a result, when flag is a value lower than “1” (flag <1), that is,
If g = 0 or flag = -1, learning is not possible in the frequency group n, and therefore, step S21 is performed.
Sets the learning enable / disable flag mode (n) for each frequency to "0".
To

Further, at step S22, flag =-
When it is confirmed that the adaptive filter 12 is in an abnormal state, it is determined that the adaptive filter 12 is in an abnormal state. Clear to 0 to deal with the abnormal state of the adaptive filter 12 (s
section2).

On the other hand, in step S20, when the flag in the frame unit is "1" or more (f
lag ≧ 1), that is, flag = 1 or flag =
If it is 2, a determination process of whether or not learning is possible for each frequency is performed in the next steps S24 to S36 (section 5).
Note that if safety of learning is ensured, the processing of section 5 may be performed only when flag = 1.

In this case, for example, when a 256-order FFT operation is performed, 256 values appear, the first of which is a DC component, which is not a complex number but a single value. . Therefore, in the present embodiment, as described above, for convenience, the DC component and the frequency component having a period of two samples are divided, the DC component is processed in steps S24 to S27, and the normal frequency component is processed in steps S28 to S32. , And cycle 2 in steps S33 to S36.
The process of determining whether or not learning is possible for the frequency component of the sample is performed. In the case of FFT, since the frequency components of two samples in the period are not complex numbers but real numbers, the processing shown in steps S33 to S36 is required.

The process of determining whether or not learning is possible for each frequency is basically the same as the process of determining whether or not learning is possible on a frame basis.

That is, based on the average power Pe (n) of the frequency components of the received signal per group and the average power Px (n) of the frequency components of the residual error signal per group.
The value of the N ratio tmp is obtained according to a predetermined formula (steps S24, S28, S33). If the S / N ratio tmp is higher than a predetermined threshold, it is determined that the learning is not suitable, and
The learning enable / disable flag mode (n) for each frequency is set to "0" (steps S26, S30, S35).

If the S / N ratio tmp is lower than a predetermined threshold value, it is determined that the learning is suitable for learning, and the learning enable / disable flag mode (n) for each frequency is set to "1" (steps S27 and S36). . At this time, in step S31, the learning enable / disable flag mode (n) for each frequency is set to "1".
, The number Ca of frequencies at which learning is permitted is counted.

When the determination as to whether or not learning is possible for each frequency is made in this way, it is determined in steps S37 to S39 whether or not learning is possible for all frequencies of the frame based on the number Ca of frequencies for which learning is possible (section 3). ).

That is, in step S37, the number Ca of frequencies for which learning is possible is compared with a constant M1. As a result, if the number Ca of learnable frequencies is larger than the constant M1 (Ca> M1), it is determined in step S38 that learning is performed on all frequencies of the frame in order to increase the speed of the entire learning. Mode of learning by frequency mode
(I) = 1.

On the other hand, if the number Ca of frequencies that can be learned is less than or equal to the constant M1 (Ca ≦ M1), the number Ca of frequencies that can be learned is compared with the constant M0 in step S39. Therefore, if the number Ca of the learnable frequencies is smaller than the constant M0 (Ca <M0), the possibility of erroneous learning is high, so that it is determined in step S40 that learning is not performed for all the frequencies of the frame. It is assumed that another learning availability flag mode (i) = 0.

Here, there is a relationship of M1> M0. For example, when there are 128 frequencies, M1 = 100 and M0 =
If the number of learnable frequencies is set to 10 or more and the number of learnable frequencies is more than 100, the learning for all the frequencies of the frame is turned ON. Turn off learning. If the frequency at which learning is possible is between 100 and 10, the previous learning result is left as it is.

As described above, when the number of frequencies determined to be learned for the frame is larger than a certain value, learning is performed for all the frequencies, and the number of frequencies determined to be learned is smaller than the certain value. If the number is small, the learning is not performed for all the frequencies, so that the learning processing of the entire frame in consideration of the correlation between the frequencies can be performed. Steps S37 to S3 here
The process 9 (section 3) is provided for an exceptional case in which the learning determination result for a specific frequency is different from the learning determination results for the preceding and following frequencies, and is not always necessary.

Finally, the threshold value T (n) for each frequency group
Update processing.

That is, in step S41, the value of n indicating the frequency to be learned is cleared to 0, and in step S42
, And the average power Pe of the frequency components of the received signal for each group
Based on (n) and the average power Px (n) of the frequency components of the residual error signal for each group, the value of the S / N ratio tmp is obtained according to the following formula.

Tmp = log10 (Pe (n) / Px (n)) Then, the average power Pe (n) for each group of the frequency components of the received signal obtained as described above, and the group of the frequency components of the residual error signal S / with other average power Px (n)
If the value of the N ratio tmp is lower than the currently set threshold value T (n) and the value of the group-wise average power Px (n) of the frequency component of the residual error signal is higher than the constant VB, (tmp <T (n) & Px (n)> VB), the variable threshold value (n) is updated in step S44 according to a predetermined calculation formula.

Subsequently, in steps S45 to S48, an update process (section 4) of the threshold value T (n) is performed for each frequency group.

First, in step S45, the average power Pe (n) for each group of the frequency components of the received signal is described.
Then, the S / N ratio tmp of the average power Px (n) of the frequency component of the residual error signal with respect to each group and the state of the learning flag of the frame are checked. As a result, if the S / N ratio tmp is lower than 0 and the flag is “1” (tmp <0 & flag = 1), step S46
As shown in S48 to S48, when the update time L has elapsed, the variable threshold T (n) is updated according to a predetermined calculation formula.

Thereafter, in step S49, the updating process is repeated for a predetermined number B of groups while updating the number n of frequencies to be processed. Then, in step S50,
When it is confirmed that n = B, the process returns to step S2, and the same processing as above is performed again.

When the above processing is simply obtained, the double talk detecting unit 17 first determines the power px of the entire frame of the received signal, the power pe of the entire frame of the residual error, and the power py of the entire frame of the transmitted signal. Frame k
Is determined to be suitable for learning (steps S1 to S1).
6).

Next, the frequency components of the frame are divided into a plurality of groups, and the average value of the power of the frequency components of the residual error signal is calculated for each group (step S17).
To S19), for the frequency component n, it is determined whether the learning of the frequency n is possible or not from the power of the group average residual error signal and the power of the n component of the received signal (steps S20 to S3).
6).

Then, according to the number of frequencies determined to be learnable, learning of all frequencies is allowed if the value is equal to or more than a certain value, and learning of all frequencies is disabled if the value is less than the certain value (steps S37 to S40). ). Finally, the variable threshold is updated (steps S41 to S50).

As described above, by independently determining whether or not learning is possible for each frequency and using the group average as a part of the reference value at the time of determination, it is possible to increase learning opportunities for each frequency and reduce erroneous learning. In addition, the performance of frequency-dependent double-talk detection can be improved.

It should be noted that a portion (s
The option 1) may be omitted. In that case, subsequent processing is performed irrespective of the value of the flag for frame learning.

When flag = −1, that is, when the filter is abnormal, the variable threshold value and the adaptive filter coefficient are cleared to 0, but this processing (section 2) is not always necessary.

Further, mo is determined by the number of frequencies that can be learned.
The process of changing de (section 3) may also be deleted. The process of updating the variable threshold based on the minimum value of the power ratio between the last received signal and the residual error (section 4) is not always necessary.

In the above section 5, t
mp = log10 (Pe (n / BN) / Px (n / B
N)). In this case, the availability flag (mode (n)) in the same group has the same value. That is, when the frequency components of the frame are divided into several groups, it is assumed that the adaptive filter learning of each frequency is the same in the group. Depending on the number of groups, this may be less likely to cause erroneous learning.

Although the case where the FFT order is twice the length of the transmission frame has been described in accordance with the literature, the present invention can be similarly applied to other lengths. For example, the FFT degree may be 256 and the transmission frame length may be 80 samples. At this time, the same applies if the overlap length of the reception frame is 176.

[0073]

As described above, according to the present invention, in a frequency-domain echo canceller, a frame of an adaptive filter is determined based on a received error signal, a transmission signal, a residual error signal between a pseudo echo signal and a transmission signal. In the case of performing the learning process for, the possibility of the learning is determined independently for each frequency, and the group average is used as a part of the reference value at the time of determination, so that the learning opportunity for each frequency is increased and the erroneous learning is performed. Thus, the performance of double-talk detection for each frequency can be improved, and deterioration of the echo cancellation amount in frequency units can be prevented.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a frequency-domain echo canceller according to an embodiment of the present invention.

FIG. 2 is a flowchart (part 1) illustrating an algorithm of a double talk detecting unit used in the frequency-domain echo canceller.

FIG. 3 is a flowchart (part 2) illustrating an algorithm of a double talk detecting unit used in the frequency domain echo canceller.

FIG. 4 is a flowchart (part 3) illustrating an algorithm of a double talk detection unit used in the frequency-domain echo canceller.

[Description of Signs] 10: Frame creation unit 11: Fourier transform unit (FFT) 12: Adaptive filter (AF) 13: Inverse Fourier transform unit (IFFT) 14: Frame creation unit 15: Frame synthesis unit 16: Fourier transform unit ( FFT) 17: Double talk detector (DTD) x: Received signal y: Transmitted signal z: Residual error signal px: Power of the entire frame of the received signal py: Power of the entire frame of the transmitted signal pz: Residual error signal Power of the whole frame

Claims (5)

[Claims] 1. A reception signal is frequency-converted in frame units, a pseudo echo signal is generated from the reception signal by using a different adaptive filter for each frequency, and the pseudo echo signal is frequency-inverted and converted from a transmission signal. An adaptive filter learning method used in a frequency domain echo canceller for removing an echo component included in the transmission signal by subtraction, wherein the reception signal, the transmission signal, the pseudo echo signal, and the transmission A learning process for the frame of the adaptive filter based on a residual error signal with respect to the signal, and determining whether or not the learning is possible is performed independently for each frequency. 2. After determining whether or not the frame is suitable for learning based on the power of the entire frame of the reception signal, the transmission signal, and the residual error signal, if the frame is suitable for learning, 2. The adaptive filter learning method according to claim 1, further comprising determining whether or not the frequency can be learned from the frequency component power of the received signal, the transmitted signal, and the residual error signal for each frequency. 3. The frequency component of the frame is divided into several groups, a value obtained by averaging the power of the frequency component of the residual error signal for each group, a threshold value set for each group, and a frequency of the received signal. The adaptive filter learning method according to claim 1, wherein whether the adaptive filter can learn the frame is determined using the power of the component. 4. When the number of frequencies determined to be learnable is larger than a certain value, learning is performed for all frequencies of the frame. When the number of frequencies determined to be learnable is smaller than a certain value, The adaptive filter learning method according to claim 1, wherein learning is not performed for all frequencies of the frame. 5. The adaptive filter learning method according to claim 1, wherein the frequency components of the frame are divided into several groups, and the determination result as to whether or not learning is possible is the same in each group.