JP3391144B2 - Band split type echo canceller - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a video conference system,
The present invention relates to a sub-band echo canceller used for a voice conference system and the like, and particularly to a sub-band echo canceller having a circuit configuration with a small amount of calculation. 2. Description of the Related Art In a system such as a video conference or a voice conference, a phenomenon is known in which a sound louder from a speaker is reflected on a wall of a room or the like and goes around a microphone, thereby returning an echo sound from a far end. ing. Echo cancellers are installed at both ends of the system to eliminate this echo. The echo canceller learns the acoustic impulse response from the speaker input to the microphone output,
It works by creating an echo replica and subtracting it from the microphone input signal. However, since the acoustic impulse response is long, the number of taps required in a normal FIR type echo canceller becomes enormous and hardware implementation becomes difficult. Become. In order to solve this problem, a sub-band echo canceller has been proposed. The sub-band echo canceller is a system that divides a signal band into a plurality of bands to form a narrow band signal and lowers a sampling speed to facilitate hardware realization. [0004] Now, the echo path length of the acoustic system is set to Techo,
Assuming that the sampling rate determined from the signal band is fo = 1 / To, the length L of the transversal filter, which is required in principle, is given by Equation 4. [0005] Assuming that this transversal filter is realized by a signal processor DSP, the data memory capacity of the DSP requires L words (actually twice as large as that including the coefficient memory). The operation speed performance A required by the DSP is as follows. That is, if an instruction of α cycles is required for each tap of the transversal filter, Equation 5 is obtained because it is necessary to perform L taps during the sampling period To. [0008] As described above, the required memory amount is proportional to the sampling speed and the echo path length, and the operation speed performance is proportional to the square of the sampling speed and the echo path length. [0010] Now, it is assumed that by dividing the signal band into N _B channel. In general terms, let us reduce the sampling rate to fs = 1 / M · fo. In this case, the required memory amount is N
_Equation 6 is given considering the entire _B channel. [0011] Regarding the speed, the number of L taps is processed within the time enlarged by M times. [0013] From these equations (6) and (7), first, M = N
_{If B} , there is no gain in the amount of memory, and 1 /
It can be seen that a gain of M occurs. Japanese Patent Laid-Open No. 4-123606 is known as a conventional example of the sub-band echo canceller. In this official gazette, as a known example,
Techniques For M = N _B is described in Equation 7.
When N bands are divided and M = N _{B is} decimated, the echo canceller inherently has a major drawback. Since the filter of the band division cannot be designed as an ideal filter, the aliasing distortion generated mutually between the adjacent channels at the band edge skirt portion cannot be avoided, and this aliasing distortion component disturbs the convergence of the adaptive digital filter (ADF) and the residual This is to reduce the echo level performance. In order to solve such a problem, in the invention described in the above publication, an oversampling complex signal processing system has been proposed. That is, the aliasing distortion is avoided by oversampling with M = N _B / 2. Then, echo cancellation is performed via the complex ADF on the band-divided signal as a complex signal. According to the above prior art, a path has been opened for realizing a transversal filter having a long tap number by complex band division signal processing and complex ADF. However, such signal processing is generally realized by a digital signal processor (DSP), but even at the present stage of the development of the integration technology, it must be said that the DSP is not good at the complex signal operation. Absent. For example, multiplication of two complex numbers is (a + jb) * (c + jd) = (a * c−b * d) as j = √ (−1).
+ J (a * d + b * c), including four multiplications of real numbers and two additions of real numbers. In DSP, such complex multiplication is performed by 1
The instruction (1 cycle) process cannot be performed, and the calculation of the real number has to be repeated. Naturally, the memory also has a real part and an imaginary part, and it is necessary to read / write while appropriately addressing, so that the overhead becomes considerable. As is well known, the adaptive algorithm of the complex ADF is expressed by the following equations (8), (9), and (10) in which all the real variables are extended to complex variables. [Equation 8] [Equation 9] [Equation 10] Consider, for example, the convolution operation of equation (8). In this method, the product sum of complex numbers is performed L times the number of taps. If the variable is a real number, the current DSP can perform an operation of accessing and multiplying two memory areas simultaneously and adding the result to the accumulator in one cycle by pipeline processing. However, when the variable is a complex number, the efficiency is reduced to 1/6 or less per product-sum because the product alone is expanded into four multiplications and two additions as described above. This further reduces efficiency, since in practice no efficient pipeline processing is possible. The situation is exactly the same for the coefficient update of equation (10). As described above, it has been found that the execution of the complex ADF operation by the DSP is extremely burdensome. It is considered that the load is likely to be one order of magnitude heavier than the real ADF operation (depending on the DSP architecture). It is an object of the present invention to avoid complex ADF operations,
An echo canceller group is realized by a real number ADF, and an efficient and economical echo canceller is realized by using a DSP. An object of the present invention is to provide an adaptive digital filter (A) for real number arithmetic processing configured as follows.
DF) is achieved by a sub-band echo canceller using groups. That is, N input signals (N = 4m, m = 4
..), A second analyzer that also divides a transmission input signal from the microphone into N band signals, and the first and second analyzers. Are output as band-division reception input signals and band-division transmission input signals, respectively. N adaptive digital filter groups for performing an echo canceling operation for each band, and N band signals are received by receiving the outputs of the adaptive digital filter groups. The first and second analyzers have a band [0, 2π].
Is divided into N complex bandpass filter groups, and 11
The filter group performs a thinning operation at a ratio of M: 1 (M = N / 4), and after multiplying the output of the filter group by the complex frequency shift operator number 12, extracts only the real part. The synthesizer outputs a real signal as N signals to be synthesized, multiplies the complex frequency shift operator by the inverse operator number 13, and outputs the same complex signal as the above. Interpolation at a ratio of 1: M (M = N / 4) is performed by the number 11 of band-pass filter groups, and the result of addition and synthesis of all the bands is used as a synthesizer output.
It is a group of echo cancellers that receive a real reception input signal and a real transmission input signal and perform an echo canceling operation by an adaptive algorithm of real variables. (Equation 11) [Mathematical formula-see original document] (Equation 13) The principle of the present invention will be described with reference to FIG. First, the band is divided into a plurality of bands in order to reduce the sampling speed of the ADF, and the oversampling is used as in the conventional case. It is assumed that the signal band [0, 2π] is divided into N, and Equation 14 is set. (Equation 14) Equation 11 is used as a filter group for dividing into N bands.
Get. Here, H (Z) is a prototype filter designed with [-π / N, π / N] as a passband.
The number 15 of N subband signals split by this filter group is shown by a black trapezoid in FIG. (Equation 15) Each signal is a complex signal except for ν = 0 and N / 2. In addition, since a digital filter cannot be designed ideally, amplitude distortion occurs in the band at each band end skirt, and spectrum leakage occurs outside the band. In this state, thinning is performed at a ratio of M: 1 to reduce the sampling rate. The mathematical principle of decimation is given by equation (16). [Equation 16] That is, the amplitude attenuates to 1 / M and
All shifted copy spectra overlap. [Mathematical formula-see original document] When M = N, it is immediately understood that an overlap occurs at the end of each band and a so-called aliasing distortion occurs. If M = N / 2, aliasing is avoided because the frequency is repeated every two frequency slots. But the signal is still a complex number. Therefore, in the above means of the present invention, in order to obtain a real number signal output, M = N / 4 by further oversampling in preparation for inversion folding. Discard the imaginary part of the complex signal,
If only the real part is taken, it is considered that the spectrum is folded back symmetrically with respect to the DC origin and becomes the overlapping number 18. (Equation 18) Y _ν (W) is the spectrum of the obtained real signal, and it is natural that y _ν (−W) = y _ν (W). Referring to FIG. 3, the complex spectrum number 16 decimated at M = N / 4 generates a copy spectrum every four frequency slots.
Therefore, even if the spectrum at the odd slot position (v = 1, 3... Channel) is viewed symmetrically with respect to the DC origin, no "overlap" occurs and the spectrum at the even slot position (v = 0, 2) ,... Channels) generate an “overlap” due to the symmetrical folding of the origin. When "overlap" occurs, the ADF does not converge. Therefore, for at least the channel where ν is even, the frequency is shifted to an odd slot position by complex modulation, and then only the real part is extracted. It is convenient to move to the next slot position, and the complex multiplication of Equation 19 is performed. n is a time index corresponding to M thinning. (Equation 19) The real signal spectrum obtained by the above process is shown at the odd channel position in FIG. The same applies to even-numbered channels. The frequency spectrum is a spectrum which is symmetrical in origin as a whole and has no aliasing distortion. The group of adaptive digital filters ADF is operated using the real signal group obtained in this way as a reception input and a transmission input. The operation rate is given by Equation 2 with respect to the original signal sampling rate fo.
0. [Mathematical formula-see original document] As a convergence algorithm, a well-known learning identification method is used. The principle of band combining the outputs of the ADF group is as follows. First, the spectrum of the corresponding channel is returned to the original position. Multiply Equation 21. (Equation 21) Thereafter, for example, the up-sample interpolation operation and the band addition number 22 may be performed through the same complex BPF group number 11 used on the analyzer side. [Mathematical formula-see original document] Since the signal amplitude is attenuated to 1 / M and 1/2 by the downsampling of M: 1 and taking out of the real part, it goes without saying that these gains are corrected. EXAMPLES The present invention will be described below with reference to examples. FIG. 1 shows the overall configuration of a band division echo canceller according to one embodiment of the present invention. The voice signal from the opposite party is provided to the reception input terminal 10 via the reception transmission path. The received voice signal is output from the reception output terminal 20 as a speaker and is band-divided into a plurality of bands by the reception analyzer 100-1 to become a band-divided reception input signal. On the other hand, the sound output from the speaker is reflected on the wall of the room or the like and appears at the transmission input terminal 30 via the microphone as an echo component. This transmission input signal is band-divided by the transmission analyzer 100-2 in exactly the same manner as the previous reception analyzer, and becomes a plurality of band-divided transmission input signals. The band-divided reception input signal and transmission input signal are input to a corresponding plurality of adaptive digital filter groups ADFi. Each of the ADFis independently performs an echo cancellation operation and outputs a band division transmission output signal. Synthesizer 200 synthesizes the bands of the plurality of band-divided transmission signals, and sends the resultant to transmission output terminal 40 as a transmission output signal. The echo signal is canceled by the above operation. Of course, the voice signal of the near-end speaker is transmitted from the transmission input terminal 30 to the transmission output terminal 40 through the band division / synthesis process only once. Next, the manner of band division will be described again with reference to FIG. First, (a) is a prototype L which is the basis of band division.
This is a characteristic of PF H (Z), and [-π / N, π / N] is a pass band. BPF shifted by Equation 23
The characteristics of Equation 11 are shown by black trapezoids. [Mathematical formula-see original document] Therefore, this characteristic may be considered as the band-divided signal spectrum itself. By the way, among these N spectrum groups, ν = i and ν = N−i (i = 1, 2, 2,
.. N / 2-1) have a complex conjugate relationship. If one of the complex conjugate signals is known, the other can be reproduced, so one of them is deleted as a processing target. And ν =
The spectra corresponding to 0 and ν = N / 2 (these are real signals) are located at the end of the transmission band (near DC and near half the sampling frequency) and are not required to be transmitted practically. Therefore, in the embodiment of FIG.
The echo cancellation processing is performed for the 2-1 channel. (Of course, the channel may be further reduced according to the required transmission band.) As described in the section above, the decimation operation of M = N / 4 is performed, and the processing rate is reduced. Lower. Then, in preparation for realization of the output signal, the slot position where ν is even is complex-modulated to an odd position, and realization is performed. Realized spectra are shown in (c) and the like. According to this figure, the gap between the spectra is sufficiently wide and there is no overlap at all. This means that the out-of-band attenuation of the prototype LPF actually designed is finite,
In addition, it is preferable to match the actual property that the amount of attenuation increases as the distance from the pass band increases. FIG. 2 shows the configuration of an analyzer / synthesizer that performs the above processing. The analyzer includes a data shift memory 110, an M: 1 downsampling circuit 120, a polyphase decomposition sub-filter group 130, and an N-point DFT 14.
0, a complex modulation circuit 150, and the synthesizer has a data addition memory 210, 1: an M up-sampling circuit 22.
0, polyphase decomposition sub-filter group 230, N-point DF
T240 and a complex modulation circuit 250, and the band division / combination circuit in FIG. 2 has a well-known polyphase configuration. Prototype LPF that meets desired specifications
It is assumed that there is H (Z) (of course, FIR type). H
The tap coefficients of (Z) can be grouped for every N taps and decomposed into the number 24 of the sub-filter group 130. That is, Equation 2
[Mathematical formula-see original document] (Equation 25) In order to form a polyphase BPF group, Z in the above equation (25) may be replaced with equation (26). (Equation 26) The N polyphase BPF groups are represented by the following equation (27). That is, the ν-th BPF number 27 is given by the inner product of the constant vector number 28 and the sub-filter vector number 29. (Equation 28) (Equation 29) When the above equation is rewritten for ν = 0, 1, 2,..., N−1, the number 30 of the discrete Fourier transform (DFT) form of the sub-filter transfer function is obtained. [Equation 30] From the equation (30), it is possible to develop the input signal in parallel to the sub-filter and DFT the output to obtain a collective BPF output. In the analyzer of FIG. 2, the downsampling circuit 120 moves to the left end within a range where the thinning process of M = N / 4 is allowed, thereby reducing the amount of calculation. Further, as described above, channel outputs such as ν = 0, N / 2 to N−1 are discarded, and the complex modulation circuit 150 outputs ν = 2, 4,.
For even channels such as those described above, after the modulation shift to odd slot positions, only the real part is extracted and output as an analyzer output. The N-point DFT 140 may use FFT,
Only the necessary channels may be directly subjected to DFT. The operation on the band synthesizer side is as follows. First, the complex modulation circuit 250 performs a modulation shift in the direction opposite to that of the previous channel for the even-numbered channel, returns the spectrum position to the original position, and inputs the result to the BPF group 230 for interpolation noise removal. A signal group in which “0” is inserted and up-sampled M times by the up-sampling circuit 220 is represented by Expression 31, and a BPF group for interpolation noise removal is represented by Expression 32. (Equation 32 may have the same characteristics as those on the analyzer side). ## EQU31 ## ## EQU00002 ## The combined output Y (Z) from the data addition shift memory 210 is given by the following equation (33). [0113] This equation shows that a signal obtained by DFT of the number of input signals 15 may be added through the number of sub-filters 34 while the number of delays is 35. ## EQU34 ## [0093] By the way, the "0" input signal inserted for up-sampling has a zero output even in the DFT, and the DFT and the sub-filter operation are actually thinned out even in the case of 36 sub-filters in the FIR format. The upsampling of M times can be moved to the right end of the synthesizer. (36) It should be noted that a "0" signal is input to the signal channel deleted because it is a complex conjugate signal.
It is necessary to double the gain correction corresponding to this. Therefore, the total gain correction (2 * 2 * M) times is included in the circuit on the analyzer side. Also, the final output to be synthesized is a real number signal whose spectrum is positive / negative symmetric, so there is no need to calculate the imaginary part. Since the number of sub-filters 37 is a real tap coefficient filter, only the real part of the DFT output is extracted.
Thereafter, only the real part needs to be calculated. That is, the complex operation is limited to only the complex modulation section and the DFT section. (37) As described above, since the divided band signals are real numbers, the adaptive filter group 300 (FIG. 1) can use the ADF of the real number algorithm. For example, a learning identification method in which all of Equations 8 to 10 are real variables can be applied. The real variable learning identification method can be efficiently processed by a digital signal processor (DSP) that can execute a product-sum operation in one cycle. The results of a verification simulation of the present invention are shown. fo = 16KHz, signal bandwidth 7KHz, Techo
= 250 ms is configured by dividing an echo canceller into N = 32. First, the transmission frequency characteristics of band division / combination are shown in FIG. This is obtained by evaluating the transmission from the analyzer to the synthesizer using a sine wave signal. The in-band ripple is ± 0.03 db or less, which is satisfactory. FIG. 5 shows a simulation of how the residual signal attenuates from the initial reset state as the echo canceller characteristic. The test signal indicates that the amount of echo suppression reached 41 db after 4 seconds of Gaussian white noise. According to the convergence of the ADF, it can be said that the band division thinning method without aliasing noise has been demonstrated. The designed prototype LPF H
If the design is made so as to further increase the out-of-band attenuation of (Z), the echo suppression amount 41 db can be further improved. Table 1 shows a comparison between the real operation type ADF according to the present invention and the amount of memory and operation using the complex operation type ADF proposed in the past. [Table 1] The evaluation is summarized as follows. 1. The hardware amounts of the analyzer and the synthesizer are almost equal. 2. The same applies to the amount of memory required for the ADF. 3. The calculation amount is 4 which is the number of required cycles α per actual tap.
And the required cycle number β per complex tap. When these three terms are evaluated by, for example, a convolution operation, in the current DSP, α = 1, and therefore, 4α = 4
It is. On the other hand, in complex convolution, β = 4 + 2 + 2 = 8, 4α: β = 4: 8 in consideration of multiplication and the sum of the real part / imaginary part. Although this depends on the architecture of the DSP, the difference may actually be larger in consideration of the address control of the memory and the like, and the merit of being able to perform ADF processing with a real number is great. To summarize the results somewhat qualitatively, the thinning rate of the real ADF is １ ／ less than that of the complex ADF. The amount of computation is the sampling speed 2
In order to work with the power, it requires essentially four times the amount of computation. However, the complex operation type may be derived from the fact that a complex operation 1 operation actually requires four times or more of a real number operation. As an example, Techo = 250 ms, f
A comparison is made between a case where an acoustic echo canceller of o = 16 KHz is directly constructed and a case where the acoustic echo canceller is constructed according to the present invention. In the direct configuration, with α = 4, A ₀ = 256MI
PS In the present invention, if N = 64 (N _B = 32), Equation 3
According to 8, A ₂ = 32 MIPS And a DSP with a processing capability of 32 MIPS.
Is used, the direct configuration requires eight, but according to the present invention, one is required for the band division / synthesis operation, and one is required for the ADF group, so that a total of two can be realized. As described above, according to the present invention,
In realizing an echo canceller requiring a long tap, band division synthesis can be performed so that an ADF that performs real number arithmetic processing can be used, and an acoustic echo canceller can be efficiently realized using this. Further, an acoustic echo canceller can be efficiently and economically realized using a small number of DSPs, and has a large practical value.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is an overall configuration diagram of a band division type echo canceller according to an embodiment of the present invention. FIG. 2 is a detailed configuration diagram of a band division analyzer and a band synthesis synthesizer of the band division type echo canceller according to one embodiment of the present invention. FIG. 3 is a spectrum diagram illustrating an operation principle of an analyzer / synthesizer of a band division type echo canceller according to an embodiment of the present invention. FIG. 4 is a transmission frequency characteristic diagram of an analyzer / synthesizer section of a band division echo canceller according to one embodiment of the present invention. FIG. 5 is an example diagram showing an echo cancellation characteristic of the band division type echo canceller of one embodiment of the present invention. [Explanation of Codes] 100-1, 100-2 ... band division analyzer, 200
... band synthesis synthesizer, 300 ... adaptive digital filter group, 110 ... data shift memory, 210 ... data addition shift memory, 120 ... M: 1 down-sampling circuit, 220 ... 1: M up-sampling circuit, 130, 23
0: polyphase decomposition sub-filter group, 140, 240
... N-point DFT circuit, 150, 250... Complex modulation circuit.

Continuation of the front page (58) Fields investigated (Int. Cl. ⁷ , DB name) H04B 3/00 H03H 21/00 H04R 3/00 INSPEC (DIALOG) JICST file (JOIS)

Claims (1)

(57) [Claims 1] N reception input signals (N = 4m, m = 1, 2, 3) from a line to be a reception output signal to a speaker
..), A second analyzer that divides a transmission input signal from the microphone into the same N band signals, and outputs the first and second analyzers respectively. A group of N adaptive digital filters for performing an echo canceling operation for each band as a band-divided reception input signal and a band-divided transmission input signal, and N band signals are subjected to band synthesis by receiving the outputs of the adaptive digital filters. The first and second analyzers are composed of one complex bandpass filter group that divides a band [0, 2π] into N, and the filter group is M: A thinning-out operation is performed at a ratio of 1 (M = N / 4), and after multiplying the output of the filter group by the complex frequency shift operator number 2, only the real part is extracted to be an analyzer output. Yes The synthesizer is provided with real signals as N signals to be synthesized, multiplies the complex frequency shift operator by the inverse operator number 3, and then obtains 1 by the same complex bandpass filter group number 1 as described above. : M (M =
N / 4), and the result of addition and synthesis of all bands is used as a synthesizer output. The adaptive digital filter group includes a real number of received input signals,
A band-division echo canceller, which is a group of echo cancellers that receives a real number of transmission input signals and performs an echo canceling operation by an adaptive algorithm of real variables. (Equation 1) (Equation 2) (Equation 3)