JPH0563609A - Echo canceler system - Google Patents

Abstract

PURPOSE:To reduce the circuit scale with respect to the echo canceller system used for, e.g. a television conference system. CONSTITUTION:The system is provided with a 1st short time spectrum analysis means 3 which applies overlap processing to each frame of an input signal, applies discrete Fourier transformation to each frame to extract a frequency component, an adaptive digital filter means 5 applying filter calculation to the extracted frequency component, an echo cancel means 6 which applies inverse discrete Fourier transformation to an output of the adaptive digital filter means 5, generates a pseudo echo signal to extract a residual echo signal, and a 2nd short time spectrum analysis means 4 which applies overlap processing to each frame of the residual echo signal, applies discrete Fourier transformation to each frame to extract a frequency component and which sends the extracted component to the adaptive filter means 5 as a control signal and the adaptive filter means 5 minimizes the frequency component of the residual echo signal by using the control signal.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an echo canceller system used in a video conference system, for example.

Usually, in a video conference system, an output from a speaker is reflected by a ceiling or a wall to generate an echo that goes around to a microphone. The echo has a length of several 100 ms, which adversely affects the communication quality. Therefore, an echo canceller that suppresses this echo is required.

FIG. 4 is an explanatory view of the principle of the echo canceller.
(a) is a block diagram, (b) is a block diagram of the transversal filter in (a). The operation of FIG. 4 will be described below assuming that the echo length is 256 ms.
The signal input to 11 and the subtractor 12 is a sampled pulse train.

First, assuming that the signal band used in the TV conference system is 7 KHz and the sampling frequency is 16 KHz, 1 ms.
Since there are 16 sampling pulses per cycle, in the case of 256 ms, there will be 4096 sampling pulses.

On the other hand, the sampling pulse train input to the echo canceller shown in FIG. 4 (a) is added to the transversal filter 11, but a part of the output of the speaker (SP) passes through a microphone (MIC) as an echo signal. It is added to the subtractor 12.

The transversal filter 11 is, for example,
It has a non-recursive circuit configuration as shown in Fig. 4 (b).
Is a unit delay element, h _{0 to} h ₄ are coefficient multipliers, 111 is an adder, but when the number of sampling pulses is 4096, the number of coefficient multipliers and unit delay elements is as large as 4096 as described above. Only the configuration is the same.

The sampling pulse applied to the circuit shown in FIG. 4 (b) is multiplied by the coefficient by the coefficient multiplier on the input / output side of each unit delay element and then added by the adder 111 to generate a pseudo echo. As a signal, it is applied to the subtractor 12 shown in FIG.

In the subtractor, the difference between the echo signal and the pseudo echo signal is taken and the residual echo signal is taken out. Therefore, the coefficient of the coefficient multiplier is controlled so as to minimize this signal. However, as the number of coefficient multipliers increases as described above, the amount of calculation also increases, and thus there has been a problem of reducing the amount of calculation, that is, reducing the circuit scale.

[0009]

2. Description of the Related Art FIG. 5 is a configuration diagram of a conventional example, and FIG. 6 is an operation explanatory diagram of a band division filter section and the like in FIG. Here, the symbols on the left side of FIG. 6 indicate the waveforms of the portions having the same symbols in FIG. The operation of FIG. 5 will be described below with reference to FIG.

First, as described above, the pulse train of 16 sampling pulses / ms obtained by sampling the signal of the TV conference system of 7 KHz band at 16 KHz is sequentially input to the band division filter unit 26 (see FIG. 6). (b) -see).

The band division filter unit 26 is composed of 32 band division filter parts which divide the 7 KHz transmission band into 32 parts, and each of the division filter parts has its own H ₁ (Z), H ₂ (Z).
..Suppose you have a transfer function of H ₃₂ (Z) (Fig. 6 (a)
See).

[0012] Now, dividing filter segment 26 having a transfer function H ₁ (Z) retrieves the pulse train corresponding to the transfer function H ₁ (Z) from a pulse train input. At this time, since the transmission band is divided into 32 as described above, the band of one divided filter portion is 1/32 of the transmission band.

Therefore, the number of pulses to be taken out can be reduced to 1/32 by the thinning processing unit 26 '(this is called thinning processing), and one sampling pulse is oversampled for every 2 ms by 1: 2 in order to prevent overlapping with adjacent filters. Added to Part 27.

Since the divided filter portion having another transfer function H ₂ (Z) ··· H ₃₂ (Z) has a different target band from the divided filter portion having the transfer function H ₁ (Z), they are different from each other. The pulse train is applied to the 1: 2 oversampling section.

Since the 1: 2 oversampling section 27 oversamples the input signal at 1 KHz, sampling pulse trains are obtained at 1 KHz intervals, and the corresponding complex coefficient adaptive digital filters (C-ADF) 241,
Add to 242. The oversampling is for suppressing the increase of residual echo caused by the overlap of the band division filters.

Complex coefficient adaptive digital filter 241, 2
42 is composed of the transversal filter described above, and multiplies the applied sampling pulse train at 2 KHz intervals by a coefficient by a coefficient multiplier (not shown) to generate a pseudo echo signal for each component. Add to the subtracters 231, 232. The multiplication at this time is a complex number multiplication.

On the other hand, the echo signal from the microphone (MIC) is
Similar to the above, the band division filter unit 21, the thinning processing unit 21 '
And the 1: 2 oversampling unit 22 decomposes each frequency component and adds it to the subtracters 231 and 232.

Therefore, the subtractor obtains the difference between the echo signal for each frequency component and the pseudo echo signal, that is, the residual echo signal for each frequency component, and again outputs the corresponding complex coefficient adaptive filters 241, 242. Added.

Each complex coefficient type adaptive digital filter is
Since the coefficient value is controlled so that the residual echo signal is minimized, the band synthesizing filter unit 25 synthesizes the outputs of the subtracters 231 and 232 and sends the synthesized signal with the minimum residual echo signal to the other side. .. Here, the calculation amounts of the echo canceller shown in FIG. 4 and the echo canceller shown in FIG. 5 are compared.

As described above, the former must be calculated using 4096 coefficients for each sampling. Considering the case of performing coefficient control by the learning identification method, the required number of multiplications is about 4096 × 2 = 8192.

On the other hand, since the latter has oversampling 1: 2, it must be calculated using 4096 × 2 = 8192 coefficients, but since the operating speed is 1 KHz, the former operating speed Considering that the multiplication becomes 1/16, and the multiplication becomes 4 times in complex multiplication, the total calculation amount is reduced to about 1/2.

The band division filter group, thinning processing,
And the 1: 2 overlap processor can be ignored because it can significantly reduce the amount of calculation by combining the polyphase filter and DFT. The same applies to the composite file portion. Generally, the same argument holds when the number of band divisions is N.

[0023]

As described above, the required calculation amount is considerably smaller than that of the direct FIR type (configuration of FIG. 4), but it is still 10 to suppress the echo of 0.5 to 2 sec. Requires a digital signal processor (DSP) on the order of ~ 20 chips. Therefore, there is a problem that the amount of calculation must be further reduced to reduce the circuit scale.

An object of the present invention is to reduce the circuit scale.

[0025]

FIG. 1 is a block diagram showing the principle of the present invention. In the figure, 3 is divided into 2n input signals into one frame, and the n signals in each frame are overlapped with each other, and each frame is subjected to discrete Fourier transform to determine the frequency component of the input signal. It is the first short-time spectrum analysis means for taking out and outputting.

Reference numeral 5 denotes a coefficient multiplier, adaptive digital filter means for performing a filter operation on the output of the first short-time spectrum analysis means, and 6 an operation result of the adaptive digital filter means. An echo canceling means for extracting the residual echo signal by taking the difference from the input echo signal by extracting the above n signal parts after performing the inverse discrete Fourier transform.

The reference numeral 4 divides the input residual echo signal into 2n units to generate one frame, performs n-overlapping processing on n signals in each frame, and performs discrete Fourier transform on each frame. It is a second short-time spectrum analysis means for extracting the frequency component of the residual echo signal and transmitting it as a control signal.

Then, the adaptive filter means uses the applied control signal to control the coefficient value of the coefficient multiplier so that the frequency component of the residual echo signal is minimized.

[0029]

The present invention is a method for obtaining the frequency component of a signal on a frame-by-frame basis. The short-term spectrum analysis is used to obtain the frequency components in the signal, and these frequency components are added to the adaptive filter means to reduce the calculation amount. I tried to reduce it.

That is, in FIG. 1, the first short-time spectrum analyzing unit sequentially divides the input signal into frames, and the latter half of the signal divided into the first frame and 2
The first half of the signal delimited by the second frame is overlapped (this is called overlap processing).

Then, the discrete Fourier transform is performed for each frame.
By performing (DFT), the frequency component of the signal delimited by the frame is obtained. And, for example, an adaptive digital filter (C-AD composed of a transversal filter).
F) By means, the pseudo echo component for each frequency component is obtained and added to the echo canceling means. The adaptive digital filter is for a complex signal as in the conventional example.

The echo canceling means performs the inverse discrete Fourier transform (IDFT) on the outputs of the respective C-ADF means at once, and then takes out the signal in the latter half part of the frame in order to eliminate the overlap, and outputs the pseudo echo signal. Then, the residual echo signal is obtained by taking the difference from the input echo signal.

The second short-time spectrum analysis means obtains each frequency component of the residual echo signal by the same processing as described above, and controls the coefficient in the adaptive digital filter so that this frequency component is minimized.

After all, the present invention corresponds to performing a convolution operation as a series of processing on the received signal,
If each coefficient value of the adaptive digital filter can be selected well, the block LMS algorithm (algorithm that controls the coefficient value in the direction of reducing the square sum of errors in the processing frame) is used in the normal FIR (finite length impulse response) type configuration.
A pseudo echo signal can be obtained by the same principle as in the case of applying.

The portion of the half overlap and DFT processing for the input signal and the residual echo signal is short-time spectrum analysis, which is equivalent to performing error minimization processing for each band. By doing so, the amount of calculation is reduced and the circuit scale can be reduced.

[0036]

2 is a block diagram of an embodiment of the present invention, and FIG.
FIG. 8 is an operation explanatory diagram of FIG. Here, the 32-sample overlap units 31 and 41 and the 64-point fast Fourier transform units 32 and 42 are the first and second
Of the short-time spectrum analysis means 3 and 4, the transversal filters 501 to 564 are the constituent portions of the adaptive digital filter means 5, the 64-point inverse fast Fourier transform unit 61, the latter half 32 sample unit 62, and the subtraction unit 63 are echo cancellers. Means 6
Is a constituent part of.

The operation of FIG. 2 will be described below with reference to FIG. 3 with 2n = 64. First, the 32 sample overlap section 31 divides the frame into frames F ₁ , F ₂ , F ₃ ··· every time 64 sampling pulses are input, and then overlaps by 32 sampling pulses as shown in Fig. 3-. After
Each frame is added to the 64-point fast Fourier transform unit 32.

Here, in order to overlap the input sampling pulse, this sampling pulse is stored in the buffer memory in the 32 sample overlap section as shown in FIG.
Store as shown in.

That is, first from address 0 to address (2n −1)
Wait until 2n sampling pulses are stored. 2n
2n point FFT when the sampling pulses are stored
Although the following processing is started, the next input sampling pulse is also stored in order from this time, starting from address 2n, but the series of processing is terminated before the sampling pulse is input up to address (3n −1).

Next, at the time when the sampling pulse is stored up to the address (3n −1), from the address n to the address (3n −1)
A series of processing is performed for 2n sampling pulses. Hereinafter, the same processing as the above is performed.

Here, in order to effectively use the buffer memory,
After the sampling pulse is stored in the address (3n −1),
The sampling pulse is stored again from the address 0.
Therefore, the target of processing in the next frame is from address 2n to (3
It is a 2n sampling pulse that combines the addresses (n-1) and 0 to (n-1).

The 64-point fast Fourier transform unit 32 is a well-known portion for efficiently calculating the discrete Fourier transform (DFT). The fast Fourier transform (FFT) is carried out for each frame, and the fast Fourier transform (FFT) is carried out for each frame. The frequency component of is extracted.

Incidentally, as shown in FIG. 3, 32 sampling pulse 32 the first half of the sampling pulse and the frame F ₂ in the latter half of the frame F _1, the first half of the second half and the frame F ₃ of the frame F ₂ · · Overlap, but this portion corresponds to the oversampling described in the conventional example and corresponds to sampling at twice the frequency.

Now, the extracted frequency components are respectively
It is added to the corresponding FIR type adaptive digital filter part, but as an adaptive digital filter part, for example, 6
Use four transversal filters 501-564.

For example, transversal filter 501
As described above, the coefficient multiplier h is applied to the applied frequency component on the input / output side of each unit delay element T.
After multiplying the coefficient by using ₀ ···, add it by the adder S ₀₁ to generate a pseudo echo signal for this frequency component.
It is applied to the 64-point inverse discrete Fourier transform unit (IFFT) 61.

The other transversal filters also operate in the same manner as described above to generate pseudo echo signals for the respective input frequency components and to generate a 64-point inverse discrete Fourier transform unit (I
FFT) (see Figure 3-).

In FIG. 3-, the first sampling pulse in the frame is the transversal filter 501, 64.
The th sampling pulse is shown to be applied to transversal 564.

The 64-point inverse discrete Fourier transform unit 61 performs an inverse discrete Fourier transform on the applied pseudo echo signal for each frequency component, and then extracts the latter half to remove the overlap and subtracts it as a pseudo echo signal. (See Fig. 3-).

Since the echo signal from the microphone MIC is also applied to the subtractor 63, the difference is taken here to extract the residual echo signal, which is applied to the 32-sample overlap section 41.

Similarly to the above, the 32-sample overlap section 41 divides the residual echo signal into one frame each time 64 residual echo signals are input, overlaps each by 32, and then performs 64-point fast Fourier transform on each frame. Add to the conversion unit 42.

Therefore, the 64-point fast Fourier transform unit 32 performs a fast Fourier transform (FFT) for each frame to extract the residual echo signal components for each frame, and the corresponding transformers are arranged so that these components are minimized. The coefficient of the coefficient multiplier in the Versal filter is updated and controlled. For the update control of the coefficient, for example, a known learning identification method may be used.

Next, a comparison will be made of the amount of calculation between the present invention and the conventional example. The length of the echo signal to be suppressed is N samples,
If the number of samples per processing frame in the present invention or the number of band divisions in the conventional example is K, the amount of calculation required in the present invention is

[0053]

[Equation 1]

On the other hand, in the band division type configuration of FIG.

[0055]

[Equation 2]

The calculation of is required. However, the group of band-divided filters is FFT and polyphase filter (multiphase filter).
It is assumed to be configured with?). Further, α is a calculation amount required for the polyphase filter.

For example, if N = 4096 and K = 64, the amount of calculation required in the present invention is 188 / sample from the equation (1),
It is about 1/3 less than the value (572 + α) / sample in the conventional example obtained from equation (2).

Here, the equation (1) is obtained as follows. That is, in FIG. 2, three K-point FFTs or K-point IFFTs are required, but the actual number of multiplications required for one actual input FFT is 2K.
[Log ₂ (K / 2) +1].

Therefore, with 3 pieces, 2K [Log ₂ (K / 2) +
1] × 3, but if it is corrected per sample, it will be thinned out to K / 2

[0060]

[Equation 3]

It becomes On the other hand, since the total number of chips is 2N, the number of multiplications per frame is 2N × 2 × 4 = 16N when the learning identification method for complex signals is processed.
Considering that if this is corrected per sample, K / 2 is thinned out

[0062]

[Equation 4]

It becomes Therefore, from equation (3) + equation (4),
The formula is obtained.

[0064]

As described in detail above, according to the present invention, there is an effect that the circuit scale can be reduced.

[Brief description of drawings]

FIG. 1 is a principle configuration diagram of the present invention.

FIG. 2 is a configuration diagram of an embodiment of the present invention.

FIG. 3 is an operation explanatory diagram of FIG. 2;

4A and 4B are diagrams illustrating the principle of an echo canceller, in which FIG. 4A is a configuration diagram and FIG. 4B is a configuration diagram of a transversal filter in FIG. 4A.

FIG. 5 is a configuration diagram of a conventional example.

FIG. 6 is an operation explanatory diagram of FIG. 5;

[Explanation of symbols]

 3 First Short-Time Spectrum Analysis Means 4 Second Short-Time Spectrum Analysis Means 5 Adaptive Digital Filter Means 6 Echo Cancellation Means

Claims (1)

[Claims] 1. An input signal is divided into 2n pieces (n is a positive integer) to form one frame, and n pieces of signals in each frame are mutually overlapped and each frame is subjected to discrete Fourier transform, A first short-time spectrum analyzing means (3) for extracting and outputting the frequency component of the input signal and a coefficient multiplier are provided, and a filter operation is performed for each output of the first short-time spectrum analyzing means. The adaptive digital filter means (5) to be performed, and after performing an inverse discrete Fourier transform on the calculation result of the adaptive digital filter means, the n signal portions are extracted to generate a pseudo echo signal, and the input The echo canceling means (6) for extracting the residual echo signal by taking the difference from the echo signal and the residual echo signal input are divided into 2n units to generate one frame, and each frame is generated. And a second short-time spectrum analysis means (4) for mutually overlapping processing the n signals of the above and performing discrete Fourier transform of each frame to extract the frequency component of the residual echo signal and transmitting it as a control signal. And the adaptive filter means uses the applied control signal to
An echo canceller system, wherein the coefficient value of the coefficient multiplier is controlled so that the frequency component of the residual echo signal is minimized.