Echo cancelling device with filtration into frequency sub-bands.

p0001The present invention relates to echo cancellers devices to be used in signal transmission facilities to the transmission in both directions, frequently referred to by the English term "full duplex". It finds a particularly important application in telephony to solve problems of teleconferencing facilities and called telephones "hands free", or at least amplifée listening. These problems consist of the risk of feedback and the existence of an echo acoustic origin.

p0002The origin of these phenomena appears in Figure 1 which shows, in solid lines, the diagram of a teleconferencing terminal. The signals x entering through a receiving line LR and from a remote terminal 8 are amplified by 10 and broadcast in a listening room amplifier with a speaker 12. In the listening room sound waves coming from on the one hand a speaker whose words are to be transmitted, the other speaker by acoustic coupling 12 (shown schematically by a channel 13) are picked up by a microphone 14 or more. The microphone 14 is connected to an amplifier 16 and the output signal y of the latter is transmitted on the FL transmission line to the remote terminal. A speaker-listener placed at the remote terminal 8 hear accordingly not only the words addressed to him, but also an echo of his own voice with a delay proportional to the longeuer of LE and RS lines, distorted by the channel transfer function acoustic echo 13. This is all the more annoying that its level is high and the gap is large. In a satellite transmission delay up to 600 ms and greatly hinder intelligibility. Acoustic feedback occurs when the coupling consisting of the channel 13 is high enough that the gain in the loop consisting of the two terminals and lines exceeds 1.

p0003Various devices are already being used to fight the acoustic echo and howling, including automatic drives of the amplification gain or echo cancellation devices in each terminal station or "free hand".

p0004Devices gain changers act by introducing attenuation before the loudspeaker or after the microphone, depending on the detected direction of communication. These devices have the drawback of producing a subjective impression of the speech cut when the inserted attenuation is high. Or constantly high attenuation is required in communication systems where the delay is important and where the echo must be strongly attenuated to stay tolerable.

p0005The invention relates to echo cancellation devices designed to directly fight the echo compensating the signal with an equivalent and opposite. For this purpose, an echo canceller device comprises an adaptive filter device connected to the incoming signal arrival line and for providing an estimate of the echo of the signal entering the subtractive input of a subtractor that receives its additive input signal assigned to the useful echo.

p0006The principle of such a device is shown with broken lines in Figure 1. The echo canceller comprises an adaptive filter FA device that receives the incoming signal x and whose output drives the subtractive input of a subtractor 18 whose additive input connected to the output of the amplifier 16. the filter FA coefficients are automatically adjusted to the e transmitted line signal, equal to the difference between the signal y echo-affected of the amplifier 16 and the estimated echo. The FA filter coefficient adaptation algorithm must be such that the filter reproduces the characteristics of the acoustic coupling channel 13, consisting of the beginning of the impulse response of that channel.

p0007With no word uttered by a speaker placed in the room, the outgoing signal on the line LE is reduced to an echo residue. When a speaker speaks into the room in front of the microphone 14, the speech is transmitted to the line LE without being reduced nor altered. Often a device (not shown) is provided to detect the presence of a speech signal from a speaker (eg level detection) and then temporarily block the adaptation of filter FA to avoid any disturbance of the filter by the words of the local speaker.

p0008This solution has the big advantage of the change auto gain, enable bidirectional operation without attenuation of the useful signals, but the currently existing devices that apply are not entirely satisfactory, because their implementation faces two main challenges.<ul><li>In the case of telephone systems, the acoustic echo cancellers comprise transversal adaptive digital filters operating at a sampling frequency of 8 to 16 kHz. Or the impulse response of channel 13 is often very long and it corresponds to several thousand coefficients at this sampling frequency. The calculations to be performed at each sampling time (convolution filtering and adaptation of the filter coefficients) require a huge volume of transactions. To perform them very quickly, numerous and expensive electronics are essential.</li><li>The characteristics of the acoustic coupling channel 13 vary over time, for example when a person moves into the listening room. Theoretical reasons, related to the spectrum of the signal received on the RS line to the required length of the adaptive filter, limit the ability of pursuing these changes (which should be addressed immediately to prevent the untimely reappearance of the echo) and the initial speed of convergence.</li></ul>

p0009It has been proposed to solve these problems, an echo canceller device interposed between the incoming signal reception line and the output signal transmission line to cancel the echo, comprising several processing channels in parallel affected to successive adjacent sub-bands of the spectral band of the outgoing signal, each channel having:<ul><li>a first analysis band-pass filter receiving the signal to be transmitted, multiplied by echo, whose output is connected to the additive input of a subtractor;</li><li>a second bandpass filter analysis, identical to the first filter, receiving the incoming signal and feeding an adaptive filter for providing an estimated echo in the sub-band to the subtractive input of the subtractor; and</li><li>a synthesis filter, symmetrical analysis filters and whose output feeds the transmission line.</li></ul>

p0010A device of this type are described for example in the article "Kompensation akustischer Echos in Frequenzteilbänden, Walter Kellermann, Frequenz, Vol. 39 (1985) No. 7-8, 209-215 pages. The block diagram of such a device is shown in Figure 2. the signal x from the line LR is cut by a bank of M analysis filters BFAR M frequency sub-bands, usually all having the same width. y the amplified signal from the microphone is cut into M sub-bands 1, ..., k, ..., M BFAM by an analysis filter bank BFAR identical to the bench. in each order subband k, an adaptive filter FA_k is fed by the incoming signal x_k, And its output is subtracted by the subtractor S_k of the signal y_k BFAM from the analysis bank. As in the conventional canceller of Figure 1, the adaptive filter AF_k is adjusted to minimize the power of the signal e_k the output of the subtractor S_k.

p0011The adjustment coefficients, shown schematically in Figure 2 by a forward arrow, is performed by a particular circuit using a conventional algorithm which is generally the algorithm of the gradient well as we can in some cases adopt a simpler algorithm, for example the sign algorithm or other adaptive algorithm, such as least squares algorithm.

p0012M e₁ signals, ..., e_k, ..., E_M, Fueling a BFS synthesis filter bank which restores the full-band signal, sent as outgoing signal on the transmission line LE.

p0013This device has several advantages over a conventional echo canceller device, whose band is not divided.<ul><li>The volume calculation is performed per unit of time is significantly reduced because you can sub-sample the signals in each sub-band; if Fe is the sampling frequency considered necessary for full band signals, each of the M subband spectral width Fe / 2M can theoretically be sub-sampled at its frequency critical decimation, Fe / M. The volume of calculations for the same impulse response length is theoretically divided by M, since the computation load in the banks of analysis and synthesis filters is negligible compared to the mass of calculations to be performed in the adaptive filters.</li><li>In each subband k is the adaptation gain of the algorithm used can be optimized according to the power of the signal x_k in this band, which increases the convergence rate and the capacity of further variations of the acoustic channel.</li></ul>

p0014But these theoretical gains are not left completely in practice. A comprehensive study of such a device (A. Gilloire, Experiments with sub-band acoustic echo CANCELLER for teleconferencing, Proc ICASSP. - 87 April 1987 Dallas, pages 2141-2144) showed that in reality it is not not possible to adopt the critical decimation frequency if the sub-bands are directly adjacent. Indeed, the sub-sampling frequency of critical decimation necessary to minimize the volume of calculations can not be used without showing, at the boundaries between sub-bands, aliasing which are not offset by adaptive filters. This leads either to use disjoint sub-bands (without cover), but with the disadvantage of introducing gaps in the spectrum of the reconstructed signal at the output of the BFS synthesis benches (which night was the speech quality when the number of sub-bands is high), or sub-sampling sub-bands at a higher frequency is the frequency of critical decimation for household guard bands avoiding folds, increasing the speed calculations necessary in filters adaptive.

p0015The present invention aims to provide an echo canceller device answering better than those previously known practical requirements, particularly as it eliminates a large extent the problem of aliasing. For this purpose, the invention proposes a device of the above defined type according to the characterizing part of claim 1.

p0016Practically, this result is generally obtained by forming said means by at least one cross adaptive filter fed by the output of at least one adjacent track. If the sub-band analysis filter is selective enough, it will often be sufficient to provide the path of order k of two crossed filters each using as input adjacent channel output signals of order k-1 and k + 1 (or one for channels 1 and M). If instead this selectivity is insufficient, or may need to affect more than two crossed filters to each channel.

p0017The Crusaders filters can often be factored into a fixed part, depending only on the characteristics of schools of analysis and synthesis filters and an adaptive part. The increased volume of calculations is low and the overall reduction in the volume compared to conventional full band canceller is maintained in a report that can reach almost the number of sub-bands.

p0018The invention will be better understood from reading the following description of embodiments given for explanatory and non-limiting. The description refers to the accompanying drawings, in which:<ul><li>Figure 1, already described, is a block diagram of a fully equipped conference facility of a conventional echo canceller device, shown in broken lines;</li><li>Figure 2, already described, is a block diagram of an echo canceller device in the sub-frequency bands;</li><li>Figure 3 shows the general diagram of a device constituting an example of implementation of the invention, a single channel being shown complete way;</li><li>4 illustrates a variant of the invention, wherein only the Crossed filters from webs directly adjacent to a given channel are stored;</li><li>5 shows a possible form of the subbands in the case of the variant described in Figure 4;</li><li>6 illustrates an embodiment of the invention, wherein the cross-filters are factorized into a fixed part and an adaptive part.</li></ul>

p0019The device shown schematically in Figure 3, wherein the members corresponding to those of FIG 2 are designated by the same reference, further comprises a bench BFAM of M analysis filters receiving the output signal y of the amplifier 16, multiplied by echo, and an analysis BFAR bench of the incoming signal x arriving by LR receiving line. The two filterbanks are identical. The output y_k BFAM attack of the additive input of a subtractor S_k which outputs the signal e_k applied to corresponding filters of the synthesis filter bank BFS feeding the transmission line LE. The signals in each sub-band are decimated (sub-sampled) in the M report within BFAM analysis benches and BFAR.

p0020The output of the filter FA_k this time is not applied directly to the subtractive input of the subtractor S_k. It is first added to the output of Crusaders FC filters_{1, k}, ..., CF_{k-1, k}, CF_{k + 1, k}, ..., CF_{M k}. The device comprises therefore, for each channel k, an adder A_k receiving the output of filter FA_k and the corresponding associated crossed filters. The coefficients of all the cross-filters contributing to the formation of the output signal of the adder A_k are adjusted by circuits whose input signal is formed by the output e_k e soutracteur_k. For Crusaders FC filters_2.1, ..., CF_{k, 1}, ..., CF_{M, 1} the coefficients will be adjusted in response to e₁ component of the output signal, etc ...

p0021The device shown in Figure 3 requires that each channel comprises, in addition to the adaptive filter AF, M-1 crossed filters FC. But in practice the aliasing bands at low borders and a high order of subband k often do not extend to the extreme order sub-bands 1 and M, which reduces the number Crusaders filters. This reduction may be particularly important that the filters of the BFAR and BFAM benches are selective. When using so-called quadrature mirror filters (often referred to as Anglo-Saxon abbreviation QMF) or pseudo-QMF, we can often limit the folding of bands corresponding to a channel in an area that does not extend beyond two adjacent channels. 5 shows, for example, the gain variation depending on the frequency possible to obtain QMF filter gain becomes almost zero for the filter of order k of the bench, well before the center frequency adjacent subbands. QMF filters and pseudo-QMF are indeed the property to ensure a gain of 3 dB cutoff frequency and an overall flat response filter bank, that is to say without changing the frequency spectrum.

p0022In this case, the device can be very simplified. As shown in Figure 4, it suffices to provide two crossed filters FC_{k-1, k} and CF_{k + 1, k} for channel k (and even one FC2, or FC 1_{M-1, M}For extreme paths 1 or M). We see in Figure 4 that the output of the filter corresponding to the order subband k BFAR analysis bank feeds the input of the main adaptive filter FA_k and the inputs of two crossed filters only FC_{k, k-1} and CF_{k, k + 1}. The adder A_k three-input receives the output of filter FA main_k and the outputs of filters crossed FC_{k-1, k} and CF_{k + 1, k} ; the output A_k is subtracted from the signal of the subband k from BFAM filter bank sub-S tractor_k.

p0023Adaptive filters FA and FC Crusaders filters may have any one of numerous known constitutions. Use may in particular filters implementing the algorithm commonly used in the transverse gradient adaptive filters. A filter embodying the least squares search algorithm gives further improved results at the cost of greater computational complexity. In some cases too, just use the sign algorithm. For a description of these filters, they work and their advantages, reference may be made to various documents, for example the article by O. MACCHI and other "Update on the transverse adaptive filtering" GRETSI 11th symposium, Nice, June 1987 pages 1G-14G and US patent 4,564,934 (Macchi).

p0024An advantageous solution is to adopt for the cross filters, a special structure, shown in Figure 6 in the particular case of a division into two sub-bands. Each of the cross filter is factored into a fixed part and an adaptive part FO FC '. TF portion is equal to the convolution of the two filters constituting each analysis bank decimated in a report 2. In this case two sub-bands, the filters of BFAR and BFAM benches are a low-pass filter, the other a high pass filter.

p0025The constitution of the Crusaders filters could be transposed to the embodiments of Figure 3 and Figure 4.

p0026Splitting into sub-bands will be made in each case taking into account the characteristics of the transmission facility.

p0027The simplest solution for use in telephony, is to use two sub-bands ranging from 0-4000 Hz respectively and 4000-8000 Hz. A slightly more complex solution, even using equal widths of sub-bands, is to split the total band into four sub-bands of 2000 Hz or each eight subbands 1,000 Hz each.

p0028In some cases however, it will be advantageous to use unequal width subbands. But in this case the correction terms must be adapted to the correct sampling frequency for each sub-band.

p0029In each case, the number of crossed filters used for each subband is selected based on the number of subbands for which an overlap occurs. In the case for example of a structure with eight sub-bands equal to 1.000 Hz wide, overflows a subband adjacent to the bands must not exceed 1000 Hz. This result can be easily obtained using effective filterbanks whose complexity remains limited, for example to 64 or 96 coefficients.

p0030will now be given, for example, features that can be given to a representative device for a sampling rate of 16 kHz, required for speech transmission over a broadband (up to 7 kHz ). To realize a device operating at 16 kHz and for identifying the first 62 milliseconds of the impulse response of the acoustic coupling channel 13, an arrangement according to the diagram of Figure 6 is used. Each sub-band is 8000 Hz. Each sub-sampled main filter FA1 or FA2 has a length of 500 coefficients. Each crossover filter has a fixed part FO length 32 (for filter banks constructed from a prototype lowpass coefficients 32) and an adaptive part FC '160 coefficients, or about a third of the length of the filters main, which is sufficient in practice.

p0031The adaptive filters are adjusted means of a stochastic gradient algorithm, adaptation representative in this case as many calculations as convolution. The volume of calculations to be performed for each pair of input samples is then: 2 * (2 * 500 + 2 * 160) + 2 * 32 + 3 * 32 = 2 800 multiplications-accumulations.

p0032This compares to 4,000 multiplications-accumulations that must perform a classic canceller to treat a pair of samples.

p0033In a second application which is also representative, where the device is still operating at the sampling frequency of 16 kHz but must identify the first 125 milliseconds of the impulse response of the acoustic coupling channel, it uses a device of the kind described in Figure 3. The device comprises M = 8 subbands. Each sub-band is sub-sampled to 2000 Hz. Each FA main filter has a length of 250 coefficients and each cross filter FC has a fixed part 12 and an adaptive coefficients of 84 coefficients. The benches of analysis and synthesis filters are constructed from a low pass prototype 96 coefficients. The band corresponding to higher frequencies (between 7 000 and 8000 Hz) is not treated, because it contains negligible power signals. This is the transition band of the anti-aliasing input filters, not shown in the figures.

p0034The volume of calculations to be performed for each block of 8 samples is: 2 * (7 * 250 + 12 * 84) + 7 * 12 + 3 * (96 + 56) = 5 856 multiplications-accumulations This compares to 32,000 multiplications-accumulations that must perform a classic canceller to treat same block of 8 samples.

p0035The invention is not limited to the particular embodiments that have been shown and described; these applications are not limited to telephony and acoustic echoes. The echo canceller device can also be used for data transmission facilities to cancel echoes are electrical line.