GB2439988A

GB2439988A - Subband coefficient adaptor for adaptive filter

Info

Publication number: GB2439988A
Application number: GB0511162A
Authority: GB
Inventors: Hamid Sepehr; Amir M Karshenas
Original assignee: TECTEON PLC
Current assignee: TECTEON PLC
Priority date: 2005-06-01
Filing date: 2005-06-01
Publication date: 2008-01-16
Also published as: GB0511162D0

Abstract

An adaptive filter has a set of subband filters for decomposing the incoming signal and the error signal into a number of subband signals, and generates subband coefficient weights from the subband signals. The subband filters have a polyphase structure having a modulation of sinusoidal form or combined sine and cosine form, to create double sideband filters having real outputs. Coefficient weights for adapting the filter are derived from the subband coefficient weights according to a frequency mapping which corresponds to the modulation. Such modulations enable real outputs to be used in a delayless adaptive filter without undue distortion or without an unduly small decimation factor. Having real outputs enables the processing load in some stages to be reduced by a factor of two. Keeping the calculation load low is important in applications such as telecommunications signal processing such as echo cancellation.

Description

1 2439988

TECTEON PATENT APPLICATION

SUBBAND COEFFICIENT ADAPTOR FOR ADAPTIVE FILTER

Field of the Invention

The invention relates to adaptive filters, to echo cancellers, adaptive controllers, to central office equipment, to methods of producing enhanced signals using such filters, to methods of offenng telecommunications services using the enhanced signals, to methods of using subscriber equipment to use the enhanced signals, and to corresponding software.

Background

Adaptive filters have been widely used in communications systems, control systems and various other systems in which the statistical characteristics of the signals to be filtered are either unknown a priori or, in some cases, slowly time-variant (nonstationary signals). Some applications of adaptive filters include echo cancellation, noise cancellation, adaptive antenna systems in which adaptive filters are used for beam steering and for providing null in the beam pattern to remove undesired interference, digital communication receivers in which adaptive filters are used to provide equalization of intersymbol interference and for channel identification, and system modeling in which an adaptive filter is used as a model to estimate the characteristics of an unknown system.

Although both infinite impulse response (IIR) and finite impulse response (FIR) filters have been considered for adaptive filtering, the FIR filter is by far the most practical and widely used. The reason for this is that the FIR filter has only adjustable zeroes and hence it is free of stability problems associated with adaptive hR filters that have adjustable poles as well as zeroes. Nevertheless, the stability of even FIR filters depends critically upon the algorithm used for adjusting its coefficients.

An important consideration in the use of an adaptive filter is the criterion for adapting and optimizing the adjustable filter parameters.

For accuracy, the filters may have many, perhaps hundreds, of taps (termed long adaptive filters'). This leads to long convergence times and heavy computational burden. To address this, subband adaptive filters were developed. These involve using bandpass filters to split the inputs into bands which are then processed separately. The number of taps and the weight update rate can be decimated in each band. This enables faster convergence and much reduces computational load.

A disadvantage of the subband adaptive filter is the delay introduced by the bandpass filters. This is critical for some applications. For example, for noise cancellation the delay limits the bandwidth over which good cancellation can be achieved. Echo cancellation is similarly limited by such delay. Such delay is propagated into the network. Uncancelled echo has an annoying psycho acoustic effect which prevents natural full duplex conversation. Performance of existing echo cancellers is still not satisfactory under all conditions. Accordingly, a delayless subband architecture has attracted much attention. It is shown in IEEE transactions on signal processing vol 43 no.8 august 1995 "A delayless subband filter architecture" by Morgan et a!.

In this architecture (shown in figure 1), adaptive weights are computed in subbands and then transformed to an equivalent wideband filter. The inputs, usually an incoming signal (termed a far-end signal in echo cancellation applications) x(t) and an error signal, e(t), are decomposed into sets of subband signals using single-sideband bandpass filters. In each subband, the signals are decimated by a factor D. The adaptive weights in each subband are transformed into the frequency domain, stacked, and inverse transformed, to obtain N wideband filter coefficients for the wideband filter W. The prospect of reducing the amount of calculation by using double-sideband subband filters was mentioned in the same paper, but no one has succeeded in achieving this tantalising possibility. The reason it could reduce the amount of calculation is that it produces real values for weights, rather than the complex values proauced by the single sideband subband filters. As the subsequent step of transformation into the frequency domain involves a multiplication, this involves a 2 x 2 scalar complex multiplication for complex numbers (=4 multiply steps) compared to just one multiply step for real numbers.

Numerous papers have been published on the issue of a real subband adaptive filter.

Two papers (K Nishikawa et a! "Structure of delayless subband adaptive filter using Hadamard transform", IEJCE Trans. Fundamentals, Vol E81-A, No 6 June 1998 and N Hiraama et al "Delayless subband adaptive filtering using Hadamard transform", IEEE Trans. On Signal Processing, Vol 47, No 6 June 1999) propose a tree structure subband adaptive filter for real subband decomposition and the Hadamard transform for the wideband adaptive filter construction. This method however cannot use the efficient polyphase structure with fast transforms. Another paper (R Merched and A H Sayed "An embedding approach to frequency-domain subband adaptive filtering", IEEE Trans on Signal Processing Vol 48, No 9, September 2000) has proposed general trigonometric transforms such as DCT (Discrete Cosine Transform), DST (Discrete Sine Transform) and DHT (Discrete Hartley Transform) for the adaptive structure which then involves only real arithmetic. This method however uses transform domain modulated bandpass filters which are known to have poor frequency characteristics.

As can be seen, there have been many attempts to fulfil the prospect of a real delayless subband adaptive filter, but none have been successful, and problem remains.

SUMMARY OF THE INVENTION

It is an object of the present invention to provide improved methods and apparatus.

A first aspect of the invention provides an adaptive filter for filtering an incoming signal using an error signal to adapt the filter, and having: a set of subband filters having a modulation of sinusoidal form or combined sine and cosine form, to create double sideband filters having real outputs for decomposing the incoming signal and the error signal into a number of real subband signals, a coefficient adaptor for generating subband coefficient weights from the subband signals, and a combiner for deriving coefficient weights for the filter from the subband coefficient weights according to a frequency mapping which corresponds to the modulation of the set of subband filters.

An advantage of such modulations and the corresponding mapping is that they enable real outputs of the filter to be used in a delayless adaptive filter such as that shown by Morgan, mentioned above, without undue distortion or without an unduly small decimation factor. This will now be explained further. Decimation factor is the proportion of the samples dropped or ignored, to the samples selected to be taken forward and used in a next stage. A higher decimation factor is desirable, as more dropped samples leads to a reduced calculation load. Too much decimation can cause loss of too much frequency information, leading to more distortion. Too much distortion can slow convergence of the filter, or cause the adaptive filter to diverge rather than converge, thus giving poor filtering performance. The modulations of the invention can maintain more frequency information for a given decimation factor, and thus enable or provide a better balance or ratio between filtering p&formance and calculation load.

Keeping the decimation factor high and therefore keeping the calculation load low is particularly important in applications such as telecommunications signal processing where the throughput of calls or data traffic is limited by the processing capacity (usually measured in MIPS, that is millions of instructions per second) of existing installed hardware. In such cases, reducing the calculation load (preferably by a software upgrade, without changing the hardware), is commercially significant because it can translate very directly into increased revenue for the operating company.

Decomposing into subbands and recombining inevitably also produces some distortion, but this is normally outweighed by the calculation reduction benefits of subbands, as set out in the background section above. The number of subbands can be chosen to suit the required performance or calculation load or frequency range.

These modulations can also allow a wide frequency range to be spanned without gaps. This is also a significant factor or constraint in some applications.

Preferably the set of subband filters has a polyphase structure. This is a more efficient way of implementing the filters, and can make use of fast and efficient transforms such as FFT, FDCT (Fast Discrete Cosine transform), FDST (Fast Discrete Sine transform) or FDHT (Fast Discrete Hartley Transform).

Preferably the subband filters use an FDCT algorithm. This is a particularly efficient algorithm, involving approximately O.5nlog2n operations for n taps, and one which can produce real outputs, rather than complex outputs, thus reducing the amount of calculation in subsequent steps.

Preferably the modulation is cos (ir(m+O.5)nIM) where M is the number of subbands, and m=O,1,2,M-1. This can give an optimum high level of decimation factor for conditions of wideband frequency coverage and negligible frequency distortion in each subband. This can minimise the calculation load as explained above, while maintaining convergence. The (m+O.5) term is concerned with eliminating a DC factor in the signal. The 1/M factor is concerned with the cut-off frequency, which is bounded by the number of subbands M. Preferably this modulation is modulation of a prototype filter of a polyphase structure.

Preferably the filter also has a transformer for transforming the weights into the frequency domain, the combiner being arranged for combining the transformed weights into a wideband set of weights, and an inverse transformer, for returning the wideband set to the time domain, for use as updated coefficients in the adaptive filter.

Preferably the frequency mapping involves selecting a proportion of the subband coefficient weights according to an amount of frequency information held in each weight. This can reduce or minimise the distortion.

Preferably the frequency mapping involves selecting successive quarters of bins of successive subbands, for use in sequence as wideband frequency bins. This corresponds closely to the modulation specified above, and so can minimise the amount of distortion.

Preferably the filter is in the form of software. This recognises the value of software as a separately tradeable entity.

Preferably the filter is in the form of a system or apparatus.

Another aspect of the invention provides an echo canceller having the above adaptive filter. Another aspect provides an adaptive controller having the above adaptive filter. Another aspect provides a corresponding method of adaptive filtering. Another aspect provides a method of using the output of the adaptive filter in a voice or data network. Another aspect provides a method of offering a voice or data transmission service using signals enhanced by the adaptive filter set out above.

Other aspects provide for methods or software corresponding to any of the apparatus or system aspects, or combinations or components of the above aspects.

Other advantages than those set out above may be apparent to those skilled in the art, particularly over other prior art of which the inventor is not yet aware. The features of dependent claims within each aspect can be combined with each other or with other aspects of the invention as would be apparent to those skilled in the art.

Brief description of the drawings

Embodiments of the invention will now be described with reference to the figures as follows: Figure 1 shows a prior art delayless subband adaptive filter structure having a polyphase FFT structure, Figure 2 sl'ows the polyphase FET structure of Figure 1 in more detail, Figure 3 shows a delayless subband adaptive filter according to a first embodiment of the invention, Figure 4 shows a polyphase FDCT structure for use in a filter of another embodiment of the invention, Figure 5 shows a polyphase FDHT structure using a combination cosine and sine modulation, Figure 6 shows an application of an adaptive filter in an echo canceller in a telecommunications network, and Figure 7 shows a typical DSP implementation of elements of a central office including an adaptive filter.

Detailed Description

Figures 1,2 prior art echo canceller having delayless subband adaptive filter with polyphase FFT structure Description of embodiments of the invention will be facilitated by reference to and comparison with the above mentioned Morgan paper, and so an example from this paper will now be described in more detail with reference to figure 1. Figure 1 shows a closed loop implementation of a delayless subband echo canceller using a polyphase FF1 structure. The closed loop version implies the output of the subtractor of the canceller is fed back as an error signal to adapt the filter weights or coefficient weights, rather than using an input to the subtractor, as an error signal to adapt the filter weights. This latter arrangement is an example of open loop control.

The invention is applicable to either. Also, for the sake of clarity, it is worth noting that this is a subband filter in the sense that the weights are calculated for each subband. As shown, the weights are recombined into a single set for the filtering to create an echo cancelling signal, implying this filter is a single wideband filter. It is also known to have subband filters each using their own set of weights for carrying out the filtering in separate subbands.

As shown in the figure, a far-end signal x(t) is fed to a loudspeakerl90 or other echo creating element (such as a 2/4 wire hybrid in a conventional telephone circuit). The microphone 200 generates an outgoing signal from which it is desired to cancel any echo. The value H represents the impulse response relating the echo to the original far end signal. The relationship will include a variable amount of delay, and a variable frequency response. The adaptive filter tries to produce an identical echo model by learning the impulse response H, by adapting the filter weights, and applying that response to the far-end signal. The resulting echo model can then be subtracted from the outgoing signal to cancel the echo using subtractor 195.

The far-end signal x(t) and the output error or residual echo are fed to respective polyphase FFT structures 210, 220 to generate two sets of subband signals y and e.

These polyphase structures are described in more detail with reference to figure 2 below. A coefficient adaptation algorithm such as the well known LMS (Least Mean Squares) algorithm 230 is used to calculate updated weights for each subband. These weights wO-wM/2 each compnse a set of 2N/M weights. N is the length of the wideband echo model filter, suggested as 512 points in Morgan. M is the number of subbands, suggested as 32 in Morgan. Next each set of weights for each band is converted to the frequency domain by FFT processes 240. This enables them to be combined in the frequency domain by the frequency stacking and conjugate complement element 250. The resulting N weights can be reconverted to the time domain by the FFT-1 process 260, and used in the wideband echo model filter 270. An 8kHz sampling rate is used, and the coefficients are updated every N/4=128 samples.

Figure 2 shows the Polyphase EFT structure in more detail. Polyphase FFT is an efficient implementation of uniform single sideband filter bank where subband filters are related by: hk(n)=h(n)e2'M k=0,1,*,M-1 where M is the number of subbands and h(n) is the prototype filter. The filter bank structure shown in the figure has a series of delays z1, 280 which generate a series of parallel data streams each delayed by one sample from a preceding one.

Decimation by a decimation factor M is carried out on each stream by decimation elements 290. Each stream is then filtered by filters 300 Em(Z) m0,1,..., M-1. The filters are different for each stream, and derived from a prototype filter. How they vary is termed the modulation of the subband filters. In this case they are exponentially modulated, as discussed in more detail below. The outputs of these filters are fed through an FFT process3l0 to produce subband signals Xm(fl) m=0,1, M-1 which are subband components of x(n). It can be summarised as being a polyphase FEE implementation in which the incoming signal (i.e. x(t)) is downsampled and filtered through a prototype filter which is exponentially modulated by the amount of: ake m = 0,1,.., M -1 which leads to a set of single sideband adaptive filters. A fast algorithm such as FF1' can be used to calculate the subband signals, which will be in the form of complex signals The number of filters can be equal to the decimation factor D (which is not general).

Consider a general filter function Hm(Z), m=0,1, ..., M-1 with the following property: Hm(Z)=Ho(ZeJ27rmIM) m=0,1,",M -1 This can be rewritten in the time domain for the sake of ease of understanding, giving: hm(fl) = where ho(n) is the time domain coefficients of the prototype filter. Ho(z) is called the prototype filter, and it can be seen that the filter bank is made up of a single filter which is exponentially modulated. In the polyphase FET structure of figure 2, where E1(z) m=0,1,..., M-1 the individual filter functions in the bank of filters can be easily derived from the prototype filter, as will now be explained. If the prototype filter is an FIR filter with L taps: H0(z) = Em(Z) is then defined as: L/M -1 Em(Z) = e,(n)z m =0,1,...,M -1 with em(n) h0(Mn + m) It is supposed that L is chosen such that L/M is an integer.

Figure 3, a delayless subband adaptive filter according to a first embodiment of the invention The first embodiment of the invention involves a filter bank having a special type of modulation to provide subband decomposition with real outputs to replace the exponential modulation of figure 2 which produces complex outputs. This leads to reduced complexity and computation load, and an improved convergence. As shown in figure 3, a subband filter bank 320 is provided to operate on an incoming signal x(t). It derives real values for the subband signals. A similar bank 330 is provided for deriving subband error signals from the error signal E(t).

As in figure 2, the subband signals are used to derive adapted weights for each subband, using the element 340 illustrated. As a result of the subband signals being real values, this element can be considerably simplified and use much less computation than is required for operating on complex values. The subband filters may be double sideband filters rather than the single sideband filters of figure 2.

This follows from having filters that produce real value outputs. If the weights are to be combined for use by a wideband filter, as in figure 2, they can be combined by converting to the frequency domain as illustrated by 350. The combination involves taking a selection of the weights from each subband, and stacking them as illustrated by 360. The selection and order of stacking of the frequency domain weights can be chosen depending on the type of subband decomposition. This will therefore differ from the selection and stacking used in figure 2, and shown in more detail in the Morgan paper. It should correspond to the type of decomposition, to ensure not too much frequency information is lost, otherwise the filter will not converge, or will converge only slowly. Finally, as before, the single set of weights output by the combiner is used by the adaptive filter 380.

Two ways of implementing the subband filter block will now be described in more detail with reference to figures 4 and 5, representing the best modes of implementation known to the inventors, as they can achieve good filter performance in terms of rate of convergence for a given computation load. Other ways can be conceived which may gain some of the advantages, and are within the scope of the claims. As shown in figure 4, a polyphase filter structure has a series of parallel data streams derived by delay elements z1 280, in a similar manner to the structure of figure 2. The decimation factor applied by decimation elements 390 is M/2, (for the case of critical sampling) and the filter bank 400 is denved based on a different modulation scheme which will be explained in more detail below. The filtered data streams are fed to an FDCT block 410 to produce the subband signals. This can be summarised as a special type of cosine modulated filter bank in a polyphase DCT structure.

Discrete cosine transform type II of signal x(n) is defined as: X(k) =x(n)A(k)cos((n+0.5)k/N A(k) = X(k) = x(n)A(k)cos((k+0.5)n/N A(k) = where N is the length of the x(n).

Polyphase DCT is an efficient implementation of uniform double sideband filter bank where subband filters are related by: hk(n)= h(n)cos(z(m+0.5)n/M) m -1 The prototype filter is modulated by the amount of: 2akcos(7z(m+0.5)n/M) m=0,1,",M-l where M is the number of subbands. In this case the result is a set of double sideband filters and the output is real signal. This can be efficiently implemented by using fast discrete cosine transform type 11. In the previous structure the prototype cut-off frequency is set to l/M but in this case because of the double sideband effect the cut-off frequency must be set to 1/(2*M). The whole set of double sideband adaptive filters can effectively span the whole frequency range, which is one of the main criterion for some applications of the subband adaptive filters. However the general technique can accommodate any number of subbands and decimation factor. The structure has the advantage of using real adaptive filters in the subbands. In principle other types of filter bank can be used, and the modulation scheme can be applied in other ways to achieve a similar outcome.

The cosine modulation can be represented as: H, (z) = O.5H0(ze'' 5)/M)) + O.5H0 (z e_ 05)/M)) m = 0,1,... ,M -1 It can also be rewritten as: h,(n) = h0(n)cos(zn(m + 0.5)/M) in this case Em(Z) is defined as:

K M-I

Em(Z)= em(n)z m=0,1,...,M-1 n=O with em(n) k(Mn/2+ m) It is supposed that K (the length of the prototype filter) is such that K/M is an integer.

A sine modulation can also be used, in which case Hrn(Z) = _0.5jHo(zeJm+o5M))+ O.5Ho(ze_m+osM)) m = -1 The same considerations as above apply equally to the sine modulation.

One preferred way of selecting and stacking the subband weights to correspond to the subband decomposition set out above, will now be described. It is similar to Morgan in that it involves a mapping of selected ones of the frequency domain weights from all the subbands. It differs in the type of mapping. The new mapping is set out below in table 1. In this table, each column represents the 32 weights for one subband which are input to the mapping. The leftmost column does not represent any subband, but merely indicates the number of each of the rows which make up the other columns. Where there is a number in a box of a column, this shows that input has been selected. If there is no number, that input is not selected.

The value of the number represents which point in the output sequence that input is mapped to. Hence in the table, there are 64 numbers, meaning there are 64 outputs, for a wideband filter having 64 taps.

The pattern of the mapping is such that a quarter of subbands PET bins (representing the frequency domain weights for each subband)are mapped to the wideband filter. Each of the subbands have a quarter of their sequence of values selected. For successive subbands, successive quarters are chosen. In the table it is assumed that M=8.

Subband Sub Sub Sub Sub Sub Sub Sub Sub FFT 0 1 2 3 4 5 6 7 0 0 ---32 --- 1 1 ---33 --- 2 2 ---34 --- 3 3 ---35 --- 4 4 ---36 --- 5 ---37 --- 6 6 ---38 --- 7 7 ---39 --- 8 -8 ---40 -- 9 -9 ---41 -- -10 ---42 -- 11 -11 ---43 -- 12 -12 ---44 -- 13 -13 ---45 -- 14 -14 ---46 -- -15 ---47 -- 16 --16 ---48 - 17 --17 ---49 - 18 --18 ---50 - 19 --19 ---51 - --20 ---52 -21 --21 ---53 - 22 --22 ---54 - 23 --23 ---55 - 24 ---24 ---56 ---25 ---57 26 ---26 ---58 27 ---27 ---59 28 ---28 ---60 29 ---29 ---61 ---30 ---62 31 ---31 ---63 Tabie 1: Frequency mapping from subbands to fuilband This way of decomposition and frequency mapping still has a slight frequency distortion between subbands. This can be reduced if desired by using well known noise suppression techniques such as wavelets. Overall, the method is more effective than the complex subband adaptive filters, and can achieve faster convergence particularly when noise reduction is used. Even with such noise reduction the real subband adaptive filter is still less computationally heavy and can use less memory than the prior art complex subband types.

The same mapping can be used for the sine modulation set out above.

The combination of sine and cosine Figure 5 shows how a combination of cosine and sine modulation can be used. In this case the modulation can be represented as: hm (n) = h0 (n)[cos(Jln(m + 0.5)! M) + sin(nn(m + 0.5)! M)) m = 0,1,..., M -1 It is then possible to have a greater decimation factor (M). As shown in the figure, the data streams are generated by delay elements 280 in the same way as shown in figure 4. The parallel streams are decimated by decimation elements 290 and filtered by the filter bank 440 derived from a prototype filter by the modulation set out above. Instead of the single FDCT block of figure 4, a pair of blocks are provided, one 450 an FDCT, the other 460 an FDST. The filtered datastreams are all fed to both blocks. The outputs of both blocks are summed for each subband individually, to produce the real subband signals, using adders 465. For the combining of the frequency domain weights, the mapping is different to that set out in table 1, as shown in table 2 below. This structure tends to have poorer convergence performance than the structure of figure 4, but can still provide an improvement over the Morgan paper.

Subband Sub Sub Sub Sub EFT 0 1 2 3 0 0 -32 - 1 1 -33 - 2 2 -34 - 3 3 -35 - 4 4 -36 - 5 -37 - 6 6 -38 - 7 7 -39 - 8 8 -40 - 9 9 -41 - 10 -42 -11 11 -43 - 12 12 -44 - 13 13 -45 - 14 14 -46 - 15 -47 - 16 -16 -48 17 -17 -49 18 -18 -50 19 -19 -51 -20 -52 21 -21 -53 22 -22 -54 23 -23 -55 24 -24 -56 -25 -57 26 -26 -58 27 -27 -59 28 -28 -60 29 -29 -61 -30 -62 31 -31 -63 Table 2, mapping corresponding to sine and cosine combination modulation Examples of applications and implementations of the adaptive filters will now be described with reference to figures 6 and 7.

Figure 6, Telephone Network.

Figure 6 shows an application of the echo canceller of the invention in a conventional telephone network. In this figure, a long-distance telephone network is shown, for making a telephone call from one subscriber to another. For convenience, one side of the network is denoted the near end, and the other side is denoted the far end. A subscriber's handset 90 is coupled to a private branch exchange (P B X) by a 2-wire subscriber line 45. In the P B X, a hybrid coil 60 is used to convert between the two wire subscriber line and a 4 -wire line to the Central Office or local exchange 10. The conversion to 4-wire enables the voice signals in two directions to be a separated, which is useful for digitising and further processing. Each P B X may support tens or hundreds of subscribers, and will have sufficient hybrid coils according to how many calls are to be supported simultaneously.

Connections from many PBXs and many subscriber lines may beconcentrated at a Central Office 10, which maybe many miles away from the subscriber. The central office contains the echo canceller 70, and a switch 80. For the sake of clarity, many other functions of the Central Office are not illustrated. There may be many echo cancellers provided, according to how many calls are to be handled simultaneously.

Conventionally, each Central Office concentrates many calls on to one or more or trunk routes 130 which make up the long distance telephone network 50. At the far end, similar elements and functions are provided. A far end Central Office 20 contains an echo canceller 110 and a switch 100. 4-wire lines 150 are provide to connect the Central Office to one or more P B Xs 30. Each will contain a hybrid 120. Two-wire subscriber lines 160 couple handsets 165 to the hybrid.

As the echo cancellers are intended to cancel echoes arising from the hybrids at each end of the circuit, in principle, they can be located anywhere in between the hybrids. They are in practice usually located in a central office where many lines are switched and concentrated. This is convenient to enable them to be shared to make more efficient use of limited processing resource, and for ease of access.

Figure 7, typical DSP implementation of elements of a central office including an adaptive filter.

Figure 7 shows in schematic form a typical D S P implementation including an adaptive filter and a coefficient adapter for the filter, which can make use the embodiments set out above with reference to figures 3,4 and 5. A switching fabric or data bus 1100 is used to multiplex or switch many individual voice channels. A number of channel interfaces 1110,1120,1130,1140 are shown. These may carry one or more voice channels. Many may be carried simultaneously if they are TDM ( time division multiplexed), or multiplexed in other ways, such as by the use of ATM (Asynchronous Transfer Mode) cells, or any other data stream for carrying voice.

The D S P has a large number of selectable software modules 1150; some of the principal ones are shown. They can be stored in off- chip memory. Individual channels may choose which of the modules are needed for particular voice calls.

The appropriate software module is called and executed in a prioritised order. They may all be implemented or executable by a single D S P device, or may be spread across any number of such devices, to increase the throughput of calls.

The modules are shown include the adaptive filter module 1160, suitable for carrying out echo cancellation, or other filtering tasks. An associated coefficient adaptation module 1170 is also shown. This may be implemented using the real subband delayless coefficient calculation techniques set out above. An echo cancellation module 1180 is provided for overall control of the echo cancellation process, making use of the adaptive filter module as appropriate. A noise reduction module 1190 may be provided, using similar techniques to those used for echo cancellation. A non-linear processing module 1200 may be provided for further processing of voice signals after echo cancellation. A DTMF processing module 1210 may be used to recognise the key pad inputs and take appropriate action.

Switching fabric control modules 1220 and voice recognition modules 1230 may be provided. Many other modules (not illustrated) may be provided for carrying out other functions typically carried out within a D S P in a Central Office.

The various modules including the adaptive filter and the coefficient adaptation module can be implemented in well known programming languages such as C or Ada, or others, as would be well known to those skilled in the art. The resulting code can be cross-compiled into a lower level language appropriate to run on a DSP, such as the fixed or floating point types made by TI or Motorola or others, or on a general purpose microprocessor, or any type of firmware, or programmable or fixed hardware, or any combination. The software can in principle be implemented as instructions or as combinations of data, instructions, rules, objects and so on.

Other variations and concluding remarks Other variations to the embodiments described above can be conceived, within the scope of the claims. Although described with reference to telecommunications applications, of course it is applicable to many other applications. It may be used in active noise control or other adaptive controllers including those shown in Morgan.

As has been described above, an adaptive filter has a set of subband filters for decomposing the incoming signal and the error signal into a number of subband signals, and generates coefficient weights from the subband signals. The subband filters have a polyphase structure having a modulation of sinusoidal form or combined sine and cosine form, to create double sideband filters having real outputs. Coefficient weights for adapting the filter are derived from the subband coefficient weights according to a frequency mapping which corresponds to the modulation of the set of subband filters. Such modulations enable real outputs of the filter to be used in a delayless adaptive filter without undue distortion or without an unduly small decimation factor. Having real outputs rather than complex outputs enables the processing load in subsequent stages to be reduced by a factor of two.

Keeping the decimation factor high and keeping the calculation load low is particularly important in applications such as telecommunications signal processing such as echo cancellation where the throughput is often limited by the processing capacity of existing installed hardware.

Claims

Claims 1. An adaptive filter for filtering an incoming signal using an

error signal to adapt the filter, the filter having: a set of subband filters having a modulation of sinusoidal form or combined sine and cosine form, to create double sideband filters having real outputs for decomposing the incoming signal and the error signal into a number of real subband signals, a coefficient adaptor for generating subband coefficient weights from the subband signals, and a combiner for deriving coefficient weights for the filter from the subband coefficient weights according to a frequency mapping which corresponds to the modulation of the set of subband filters.

2. The filter of claim 1, the set of subband filters having a polyphase structure.

3. The filter of claim 1, the subband filters being arranged to use an FDCT algorithm.

4. The filter of claim 1, the modulation being cos (it(m+O.5)nfM) or sin (7r(m+O.5)n/M) where M is the number of subbands, and m=O,1,2,M-1.

5. The filter of claim 2, the modulation being modulation of a prototype filter of the polyphase structure.

6. The filter of claim 1, having a transformer for transforming the weights into the frequency domain, the combiner being arranged for combining the transformed weights into a wideband set of weights, and an inverse transformer, for returning the wideband set to the time domain, for use as updated coefficients in the adaptive filter.

7. The filter of claim 1, the frequency mapping involving selecting a proportion of the subband coefficient weights according to an amount of frequency information held in each weight.

8. The filter of claim 7, the mapping involving a selection of successive quarters of bins of successive subband signals, for use in sequence as wideband frequency bins.

9. The filter of claim 1, the filter being in the form of software.

10. The filter of claim 1, in the form of a system or apparatus.

11. An echo canceller having the adaptive filter of claim 1.

12. An echo canceller having the adaptive filter of claim 4.

13. Central office apparatus for a telecommunications network, arranged to process voice or data transmissions, and having the adaptive filter of claim 1.

14. Central office apparatus for a telecommunications network, arranged to process voice or data transmissions, and having the adaptive filter of claim 4.

An adaptive controller having the adaptive filter of claim 1.

16. A method of producing enhanced voice or data signals using an adaptive filter for filtering an incoming signal using an error signal to adapt the filter, the method having the steps of: decomposing the voice or data signals and an error signal into a number of subband signals using a set of subband filters having a modulation of sinusoidal form or combined sine and cosine form, to create double sideband filters having real outputs, generating subband coefficient weights for the adaptive filter from the subband signals, the set of subband filters, and deriving coefficient weights for adapting the filter from the subband coefficient weights according to a frequency mapping which corresponds to the modulation of the set of subband filters.

17. A method of operating subscriber equipment coupled to a telecommunications network, to transmit or receive voice or data signals after the signals have been enhanced by the method of claim 16.

18. A method of offering to subscribers a voice or data transmission service using signals enhanced by the method of claim 16.

19. An adaptive filter for filtering an incoming signal using an error signal to adapt the filter, the filter having: a set of subband filters having a polyphase structure arranged to create double sideband filters outputting real signal values for decomposing the incoming signal and the error signal into a number of real subband signals, a coefficient adaptor for generating coefficient weights from the subband signals, and a combiner for deriving coefficient weights for the filter from the subband coefficient weights according to a frequency mapping which corresponds to the modulation of the set of subband filters.

20. A method of producing enhanced voice or data signals using an adaptive filter for filtering an incoming signal using an error signal to adapt the filter, the method having the steps of: decomposing the voice or data signals and an error signal into a number of subband signals using a set of subband filters having a polyphase structure arranged to create double sideband filters outputting real signal values, generating coefficient weights for the adaptive filter from the subband signals, and deriving coefficient weights for the filter from the subband coefficient weights according to a frequency mapping which corresponds to the modulation of the set of subband filters.