2730

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 11, NOVEMBER 2002

The Complex Subband Decomposition and its Application to the Decimation of Large Adaptive Filtering Problems
James P. Reilly, Member, IEEE, Matt Wilbur, Member, IEEE, Michael Seibert, and Nima Ahmadvand
AbstractIn this paper, we show that a near perfect reconstruc-channel filterbank with a diagonal system inserted tion (NPR) between the analysis and synthesis filterbanks may be used to decompose a finite impulse response (FIR) system of order into complex subband components, each of order , where is the downsampling rate. This decomposition is at the expense of using complex arithmetic for the subband processing. The theory surrounding the proposed filterbank structure leads to a new understanding of subbanded adaptive filtering implementations. It also leads naturally to a delayless subbanded adaptive filter scheme. Using conditions on the analysis and synthesis filters, the formulas for the subband components and their respective properties are developed. Simulation results for an acoustic echo cancellation (AEC) example are given to support the developed theory. Index TermsAcoustic echo cancellation, adaptive filtering, decimation, filterbanks, subbanding.

I. INTRODUCTION ILTERBANK theory [1] is by now a mature topic. The classical approach to this subject has concentrated on the development of perfect reconstruction (PR) or near perfect reconstruction (NPR) filterbank systems, salient features of which are that the overall frequency response of the filterbank is distortionless and that the aliasing error introduced from downsampling in the subbands is cancelled at the output. This theory requires that there be no processing in the subbands. This requirement has limited the application of filterbank theory to such areas as data compression and related topics. system In this paper, we show that a diagonal can indeed be inserted between the analysis and synthesis filters of an channel filterbank in a meaningful way. By this, we of order (which may mean that an arbitrary FIR system be large) can be decimated by a filterbank structure shown in Fig. 1, where the subband components of are each , where is the downsampling rate of the of order filterbank. In the approach adopted in this paper, the analysis and synthesis filters have single-sided frequency responses. This decimation can render computationally intractable problems feasible and is useful in applications such as audio
Manuscript received March 28, 2000; revised June 12, 2002. The associate editor coordinating the review of this paper and approving it for publication was Prof. Arnab K. Shaw. J. P. Reilly is with the Department of Electrical and Computer Engineering McMaster University, Hamilton, ON, Canada L8S 4K1 (e-mail: reillyj@mcmaster.ca). M. Wilbur is with Sirenza Microdevices, Ottawa, ON, Canada. M. Seibert is with Zarlink Semiconductor, Ottawa, ON, Canada. N. Ahmadvand is with Peleton Photonic Systems, Ottawa, ON, Canada. Digital Object Identifier 10.1109/TSP.2002.804068.

and acoustics, which typically involve determination of a very large number of parameters describing an acoustic channel. Examples include blind identification of room impulse responses [2], adaptive beamforming [3], and active noise control [4]. In this paper, however, we concentrate on the acoustic echo cancellation (AEC) problem in telephony. This problem is associated with the hands-free telephone, which requires an acoustic echo canceller to eliminate the acoustic feedback from the loudspeaker, through the room, and back to the microphone. This device is an adaptive filter that estimates the acoustic impulse response (AIR) of the room and uses this estimate to suppress the echo signal, which is picked up by the microphone and sent back to the far end. The difficulty with this particular application is that the AIR is very long (typically on the order of several thousand samples), and the required adaptation rate is very fast, due to the rate at which the acoustic properties of the room can change. The implementation must also be very inexpensive. This later requirement stipulates the use of the LMS algorithm, but its performance is typically not adequate in these adverse circumstances. The classical filterbank theory has recently been extended through the use of subbanded adaptive filters [5][10], which have proven very successful in the AEC application. Subbanding using the single-sided filterbank approach proposed here has been used previously for adaptive filtering [9] and for transmultiplexing [11], [12]. However, the concept of subbanded adaptive filtering has generally evolved from an ad hoc motivation, despite the fact it has been well researched. In this paper, we present new theoretical results using single-sided filterbanks that rigorously justify subband adaptive filtering. One of the difficulties associated with a subbanded approach to the AEC problem is filterbank delay. Excessive delay in the speech path leads to difficulty in conducting a natural, full-duplex conversation. Thus, there have been delayless subbanded methods proposed [13][15]. The theoretical development presented in this paper leads naturally to a delayless adaptive filter structure, which is similar to that presented in [13]. The proposed method is briefly discussed in this paper. In Section II, we show that aliasing in the subbands can be (approximately) eliminated through the use of filterbanks with single-sided1 frequency responses. We then show that this property permits the system to be diagonal. In contrast, for must be considered a dense the general case, the system matrix in order to completely suppress aliasing errors
1This term follows from the term single sideband modulation. Here, we retain, e.g., the positive frequency components corresponding to a real signal and suppress the negative components.

1053-587X/02$17.00 2002 IEEE

REILLY et al.: COMPLEX SUBBAND DECOMPOSITION

2731

Fig. 1. Arbitary FIR system S (z ). (b) Subband approximation of the system S (z ).

[5], [6]. We then show that a sequence of subband components can approximate an arbitrary linearly time invariant (LTI) system of interest. A straightforward method of identifying the subband components is then presented, and it is shown that the subband components are indeed of order . We show there exists a many-toone correspondence be. tween the subband components and the model response We also show that convolution inside the subbands of the subband components representing two signals of interest is approximately equivalent to the convolution outside the subbands of the signals themselves. The delayless adaptive filter implementation is briefly discussed. These theoretical insights offer new approaches and flexibility for implementation of adaptive filtering problems. In Section III, we briefly discuss procedures for the design of the analysis and synthesis filters. Verification of the theory and simulations of the AEC problem are presented in Section IV. Conclusions are then presented in Section V. denote a polynomial in Notation: Symbols such as , whereas the lowercase version of the symbol, e.g., , represents the corresponding sequence in time. Bold characters, such and , represent a vector and a matrix, respectively, as whose elements are polynomials in . A bold lowercase character such as ( ) represents a vector whose elements define a sequence in time. The superscript represents the transpose opmeans the argument has been downeration. The notation sampled by a factor of . II. COMPLEX SUBBAND DECOMPOSITION In this section, we develop the complex subband decomposishown in Fig. 1(a) using tion of an arbitrary LTI system the oversampled channel filterbank shown in Fig. 1(b).

A. Complex Filterbanks To begin, we define a set of over the range overlapping frequency bands such that

(1) where is a transition bandwidth corresponding to a practical is taken to be even. analysis filter response. The integer Consider a set of FIR analysis filters of Fig. 1(b), each , having single-sided frequency responses of length , with the response specification finite magnitude otherwise (2)

as shown in Fig. 2(a). These filters are related in the time do, main to a real lowpass prototype with cut-off frequency as shown in Fig. 2(b), by (3) For the filterbank under discussion, we make the following set of assumptions. . a) . b) and . c) Under these assumptions, the -domain version of (3) is (4)

2732

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 11, NOVEMBER 2002

Fig. 2. (a) Example of the magnitude response of

H(

z)

. (b) Corresponding prototype filter H (e

).

where we have

. Similarly, for the synthesis filters,

(5) and the model reWe further assume that the input signal are real. sponse The analysis and synthesis filters are designed so that the filterbank satisfies the following properties, which we show later to be sufficient for the subband structure of Fig. 1(b) to reprefor appropriately chosen . sent the system 1)

term is equivalent to taking the real part in the time domain as shown at the output of Fig. 1(b). Equation (6) can be satisfied by making not too large with respect to . Equation (7) can be satisfied by making a root-Nyquist filter. resulting from downsamFig. 3 shows an example of pling and then upsampling by a factor of . From this figure, it is greater is clear that (6) can only be satisfied if the shift , i.e., than the bandwidth occupied by (8) or (9)

(6) 2) (7) where denotes complex conjugation, is a constant, and is a delay introduced for causality.2 The inclusion of the second
2With reference to (7), the quantity A (z ) = sequence a[n]:
a

[n]z

for a given

In the sequel, we will refer to an ideal prototype filter. By is such that (6) and (7) this, we mean that the prototype are satisfied with equality. Note that a necessary condition for (6) is that the frequency and be single sided. For double-sided responses copies of , each translated by responses, radians, cannot exist without overlapping with in their passbands for some values of , regardless of the value of . It is straightforward to show that this filterbank corresponds -times oversampled generalized DFT (GDFT) filterto a subband signals retained. For bank [16] with only the first

REILLY et al.: COMPLEX SUBBAND DECOMPOSITION

2733

Fig. 3.

Magnitude spectrum

H (e

) after downsampling and upsampling by a factor of

K.
(17)

real input signals, the remaining subband signals are comsignals and are, therefore, replex conjugates of the first dundant. Thus, without loss of generality, we ignore the sub. bands Since the analysis and synthesis filters have single-sided frequency responses, the time-domain signals after the analysis filters are complex. This necessitates the use of complex arithmetic with any algorithm operating in the subbands. We now use our conditions on the analysis and synthesis filin Fig. 1 so ters to specify the subband components approximates the output that the output of the filterbank of our model system of interest. In this vein, the sigdefined in Fig. 1(b) are given as [1] nals (10) We can combine the into vector form and write (11) where component matrix given by , is the alias

and

Let us now consider the th element given as

of

in (16). It is

(18) and From (6), each set of frequency responses overlap only in their stopbands for . For the time being, we assume an ideal prototype filter. Therefore (19) Equation (19) is a statement that aliasing error is suppressed in the subbands. From (15), (18), and (19), we then have (20) We now show that if the are chosen to satisfy (21)

. . . and the input vector . The signals which are arranged into the vector as

. . .

(12) in Fig. 1(b) is equivalent to in Fig. 1(a). Acthen cordingly, we substitute (21) into (20) to obtain (22) of By taking the real part of the time-domain equivalent and recalling that and are both pure real, and using (7), we have (23) is within a scale and a delay of , as required. Thus, In summary, as a consequence of (6) and (7), with an ideal prototype filter, we see that the filterbank system of Fig. 1(b) inserted bewith an appropriately chosen diagonal system tween the analysis and synthesis filterbanks is indeed equivalent . The satisfying (21) are to an arbitrary FIR system . referred to as the subband components of

after the upsampler, , can then be expressed (13)

where diag Thus, the filterbank output can be expressed as (15) where (16) (14)

2734

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 11, NOVEMBER 2002

In the case where the prototype is a nonideal practical filter whose stopband attenuation is finite, (6) holds only approximately, with the result that (19) to (23) also hold only approximately. The approximations become arbitrarily good as the proapproaches an ideal characteristic. totype filter response is periodic with peWe now interpret (21). Since , there are images in the range , each riod . By definition, of which can have a bandwidth of at most is completely specified by one of its periodic imis smaller than (or at most as ages. From (8), the width of . This allows wide as) the bandwidth of one period of , such that us to define the complement of is a continuous (i.e., nondisjoint) frequency band that is radians wide and such that is negligible for . contains one period of . EquaTherefore, over , leaving tion (21) restricts the response of over arbitrary. The result is that, the response of corresponding to in general, the subband components corresponds a fullband response are not unique. Note that . to the stopband region of the downsampled version of and have the same magnitude responses, Since in are suppressed by the syncomponents of thesis filters. Thus, we have the following: In the context of Fig. 1, where , (6) and (7) are satisfied, then given any FIR scalar system we can find many systems for each satisfying . Conversely, for each if we find many whose frequency responses differ only for , then each ( fixed) will correspond to the same . of the We now discuss the computation of the subband components for the case of a practical prototype filter. In this case, it is possuch sible to find FIR subband components in a least-squares that the system in Fig. 1(b) approximates when is FIR. Starting with the approxisense for mation (21), downsampling on both sides gives

Using Parsevals equality, this is equivalent in the time domain to

(27) in We define the quantity denotes the convolution operation). This (27) (where in quantity is the time-domain version of . Likewise, we define Fig. 1(b) when the input . This is the time-domain version when . Thus, the least-squares of given by (27) is the , which forces the solution to be closest in the least-squares sense corresponding . to the signal Before solving (27), we consider the length of the subband . We denote to be the length of an associcomponents 3 (Exceptions are and , which are to be ated sequence is given by defined). Then, (28) denotes the ceiling operator, is the length of where is the length of . Since we require that is constrained to be then , ,

(29) It is important to notice that the length of the subband compo, which may be significantly smaller than nents is of order for large . Given that the lengths are now established, the comay now be found in a minimum least squares efficients sense of (27) by solving (30)

(24) where If aliasing in the subbands is kept small, then a good choice of in this equation will be a good choice of in (21). Define the error in the approximation (24) as , and is the Toeplitz convolution matrix ,

. . . (25) so that we may find the least-squares approximation for on the unit circle as (26) . . .

Toeplitz

(31)

3For ease of notation, we drop any associated characters # and [n](unless ambiguity arises).

REILLY et al.: COMPLEX SUBBAND DECOMPOSITION

2735

Thus, the least-squares subband components of an FIR impulse are given by response (32)

whose least-squares subband components are subband

. If for each (35)

then is the MoorePenrose pseudo-inverse of the matrix where . Since the MoorePenrose pseudo-inverse is unique for a given full column rank matrix, we can conclude this discussion of length , a set of as follows. Given a fullband system FIR analysis filters and synthesis filters of length satisfying (6) and (7) and downsampling factor satisfying (9), a may be approximated in a leastfull-band LTI FIR system squares sense by a sequence of inband filters of length . is the From (32), we see that the subband component with . deconvolution of the quantity corresponding to a set of subThe overall system response 4 may be computed as band components (36) Proof: Since we assume that the stopband attenuation of the analysis filters is perfect, then there exist subband compo, , and for the th subband satisfying nents (37) (38) (39) Since , from (37), we have

(33)

which corresponds to the impulse response of Fig. 1(b). Even though the development of this section has treated is a single-input single-output only the case where system, we can extend the development to the multiple-input multiple-output case in a straightforward way. This can be done by incorporating a separate analysis filterbank for each required input and a separate synthesis filterbank for each required output. B. Properties of the Complex Subband Decomposition In the following, we assume an ideal prototype filter. Property 1: The filterbank structure of Fig. 1(b) is LTI. Proof: Using (20), the transfer function of the filterbank is given by

(40) When there is perfect stopband attenuation, the last equation must be equal to for only shows that . For practical filters, these equalover the frequency range ities (except for the last one) must be replaced with approxima. tions, giving This property indicates that we can approximate subband by components for a composite system and . In convolving the subband components of general, these subband components are not the least-squares , but in practice, they allow us to subband components of approximate the system shown in Fig. 4 with that in Fig. 5. C. Applications to the Adaptive Filtering Problem We now apply the complex subband decomposition to the adaptive filter problem [20]. The objective is to determine a filter to minimize the quantity response

(34) From [17], a system is LTI if and only if it has a transfer function. In the case of a practical prototype filter, extraneous terms that prohibit the definition of a transfer function appear in (20). However, for a sufficiently good prototype filter, these terms may be considered negligible. Property 2: Let and be the -transforms of arand be the bitrary discrete-time signals, and let -transforms of the least-squares subband components of and , respectively. Furthermore, define ,
4Here, we denote c [n] as the least-squares subband component of some sequence a[n] in the mth subband.

(41) , , and are the (or an approximation of ), where input, the desired, and the output error signals, respectively. Because any signal can be decomposed into subband components, we can write (41) as

(42) Using Property 2, we can now write (43)

2736

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 11, NOVEMBER 2002

Fig. 4. Approximation of C (z ) using subband components C

( ).

Fig. 5. Approximation of C (z ) using C

( )

zC

( ).

where (44) Equation (43) can be implemented by an adaptive filter algorithm operating inside each subband. (This is the open-loop configuration referred to by various authors.) Thus, this analysis implies that a subbanded configuration, which has an adaptive filter algorithm in each subband, yielding a set of subband coef, which minimize each in (43), can replace a ficients minifullband adaptive filter algorithm whose coefficients mize in (41). Further, with the commonly used LMS algorithm, the convergence rate is inversely proportional to the number of taps [20]. For a subbanded implementation, the length of the sub. Thus, we band components is given from (29) to be expect the convergence rate of the subbanded implementation to improve roughly by a factor of for this choice of adaptive filter. This is in addition to computational gains, which are discussed in Section IV. becomes ideal, In the limit, as the filterbank prototype the subband components given by (44) represent the fullband exactly. Thus, the error introduced by the subband response adaptive filter implementation can become negligible for a good . enough specification on Delayless Subband Adaptive Filtering: The complex subband decomposition leads to a straightforward method of eliminating this filterbank delay. An implementation is shown in Fig. 6, where it is assumed the desired output is the . Subband components corresponding to signal the adaptive filter coefficients are determined by an adaptive

filter algorithm operating inside the subbands, as discussed previously. However, the desired output is computed with the outside the subbands by convolving the input , which are the filter coefficients corresponding to filter , given by (33). The the set of subband coefficients through signal path thus does not experience any delay induced by the analysis or synthesis filterbanks. Note that a synthesis filterbank is not required in Fig. 6 since the desired output is not computed in the subbands. signal This proposed delayless implementation is similar to the open-loop configuration of [13], except that our proposed technique uses (33) to update the coefficients, whereas the method of [13] uses a sequence of fast Fourier transform (FFT) operations. The development of this technique using the concept of subband components gives the method theoretical rigor and gives new understanding of the approach. The proposed delayless configuration exhibits the same improved convergence performance as the conventional subbanded adaptive filter implementation. However, in the basic form of Fig. 6, the delayless structure requires substantially more computations than the conventional subbanded adaptive filter. This is due , which requires to the cost of the operation real multiply/adds per sample: a figure that can dominate the complexity of the entire configuration. However, in [13], it is shown that the computational cost of this convolution can be reduced by a factor of approximately and remain delayless is using a blocked FFT approach. Here, the sequence segments of equal length. Then, the first partitioned into segment can process by direct convolution (thus giving the delayless property), whereas the remaining segments process

REILLY et al.: COMPLEX SUBBAND DECOMPOSITION

2737

Fig. 6. Delayless subbband adaptive filter implementation.

by fast convolutions using FFTs and inverse FFTs, sequentially, for each segment. In addition, more computational overhead is involved in upaccording to (33). To save computadating the coefficients tions, this must be done at periodic intervals, with the effect that will lag behind by one pethe adaptation of the coefficients riod. The expense of this update can be significantly reduced by using a polyphase realization [1] for the filterbanks. III. FILTER DESIGN The design of the filters and for the proposed complex filterbank technique involves fewer constraints than for conventional subbanding systems since no consideration need be given to cancellation of unwanted aliasing components. There are many design possibilities since the only requirement and satisfy (6) and (7). is that the We briefly present two such filter design techniques. The first is based on a spectral factorization of a Nyquist filter. Let be the length of analysis and synthesis time domain responses and , respectively. Let be a Nyquist response . (Any windowed sinc function is a Nyquist of length of response [1]). We interpret the Fourier transform

as the magnitude-squared response of the desired proto. We can then assign to the mintype filter . This spectral factorizaimum-phase spectral factor of tion can be achieved using the complex cepstrum, as outlined is asin [1, App. D]. The synthesis filter prototype . We note in this case, signed to be the paraconjugate of is Nyquist, (7) is satisfied with the equality, because whereas (6) holds only approximately. Another approach is a least-squares procedure for designing . Here again, we treat , where the corresponding is of length , as the magnitudetime domain quantity . Consider the objective function squared response of , which involves a criterion on the stopband and passband, respectively, as

(45) where controls the weighting between the stopband and the passband regions. Fig. 7 may aid in visualizing the limits on

2738

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 11, NOVEMBER 2002

Fig. 7.

Pertinent frequency values useful for filter design.

the integral in (45). The sequence is chosen so that the is the solution to corresponding (46) In practice, since the stopband has small magnitude, only the first few terms of the sum in (45) need to be considered. and By virtue of the constraint in (46), the prototypes can be calculated by spectral factorization of in the same manner as previous. An alternative design procedure based on convex optimization of the prototype filter is given in [19]. IV. RESULTS In this section, we present some examples demonstrating the performance of the proposed subbanding technique. First, we present simulations that verify the theoretical development of the previous sections. We then discuss a subbanded implementation of an acoustic echo canceller (AEC). Next, we demonstrate the performance of the delayless subbanded AEC technique. For the following examples, we use a GDFT filterbank , subbands with the upper eight igwith coeffinored. The analysis and synthesis filters have cients and are designed to satisfy (6) and (7) well. The magnitude spectra of the analysis filters are shown in Fig. 8. A. Calculation of the Subband Coefficients with For this example, we decompose an FIR system into eight subband components for variable length . In this example, is a white, zero mean, unit variance Gaussian noise sequence. This experiment demonstrates how approximates in Fig. 1 versus well the output signal when the subband components are calculated using (32) for the given filterbank configuration. The error is calculated as expressed in decibels. The results are given in Table I. As can be seen, for only modestly long filters, the approximation can be considered quite accurate. For the next example, we decompose the same FIR system for a fixed into subband components of length

Fig. 8. Magnitude response of 16-channel GDFT analysis filters. TABLE I APPROXIMATION ERROR VERSUS

(47) using the the least-squares (LS) approximation given in (32).

REILLY et al.: COMPLEX SUBBAND DECOMPOSITION

2739

Fig. 9. Illustration of LS subband decomposition for m = 5.

Fig. 10.
c

Real parts of (h [n])

3 c

[n]

(solid) and (h [n])

Fig. 9 shows further results for the fifth subband, which was arbitrarily chosen. The plot in the upper left-hand corner shows , the magnitude spectrum of the fifth subband component of . In the upper right-hand corner, the magnitude spectrum for the for the downsampled analysis filter (multiplied by sake of appearance) is shown. The lower left-hand plot gives . Fithe magnitude spectrum of is shown in nally, the amplitude spectrum of the lower right-hand plot. For comparison, magnitude spectrum is superimposed. As can be seen, the LS subband of , and the component allows a good approximation of normalized squared error is approximately 43.05 dB. Now, we wish to demonstrate the accuracy of Property 2. For and , this purpose, we use two random sequences samples, which are both similar to that each of length used in the previous section. The real parts of the time domain in representation of the signals, corresponding to the each of Figs. 4 and 5, respectively, are shown superimposed in . The imaginary Fig. 10, where we have arbitrarily set parts show similar characteristics. as the subband comThe error due to using instead of the least-squares subband component of is approximately 30.12 dB for this subband. ponent in Note that this is the difference in the respective signals Figs. 4 and 5 and not the difference in the subband components themselves. The norm of the difference between the subband component and its approximation (normalized ) is 6.76 dB, which is a significant to the norm of increase over that shown in Fig. 10. We see from Fig. 11, which shows the two subband components in the frequency domain, that most of this error is concentrated in the region where the is less significant. This is a magnitude response of direct consequence of the fact that the frequency domain com, as ponents of any subband component is not specified over discussed in Section II. The result is that when convolving the downsampled analysis filter response with the approximation to the least-squares sub, most of the error is suppressed. band component of

[n]

3c

[n]

(dashed).

Fig. 11. Frequency response of subband component approximations. Solid: Magnitude of the Fourier transform of c [n]. Dot-dashed: Magnitude of the Fourier transform of the approximation c [n] 3 c [n]. Dashed: Magnitude response of the downsampled analysis filter.

Since the downsampled synthesis filters have the same magnitude response as the downsampled analysis filters, errors in the subband component approximations are suppressed at the filterbank output as well. B. Application to the AEC Problem We now illustrate the proposed subband decomposition procedure with the AEC problem in telephony [5][10], [18]. The AEC problem is depicted in Fig. 12, where the objective is to cancel the far-end input speech component, which appears at point A in the figure. In this discussion, we do not address the associated double-talk detection problem [21], which has to do with detecting the presence of the near-end speech signal. This

2740

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 11, NOVEMBER 2002

Fig. 12.

Depiction of the AEC problem.

Fig. 14. Echoed speech signal for all three echo cancellation experiments.

Fig. 13.

Acoustic impulse response used for AEC example.

example is intended to demonstrate the effectiveness of the proposed complex subbanding scheme in decimating a large adaptive filtering problem into many smaller problems. We demonstrate three cases, each using an AIR, measured in a real room, of 2000 samples in duration at a sampling rate of 8 kHz. The AIR waveform is shown in Fig. 13. The input in each case is real speech (approximately 9 s in duration). In each case, the echoed speech (that received by the microphone) was generated by convolving the AIR with the input speech signal and is shown in Fig. 14. The three test cases correspond to a) the fullband case where no subbanding is used; b) the subbanding method proposed in [5]; c) AEC using the proposed complex subband decomposition. In each case, echo cancellation is achieved using the normalized LMS (NLMS) adaptive filtering algorithm [20] with a value . (For the latter two of the NLMS step-size parameter cases, the NLMS algorithm is applied in each subband with the error signal being generated in the respective subband.) The near-end talker in Fig. 12 was silent, and the AIR was assumed stationary over the entire interval. In each case, background noise at an SNR of 25 dB was added to the echo signal.

Fig. 15. Echo canceller results for the fullband case. Top: Residual echo signal versus time at the canceller output. Bottom: ERLE in decibels versus time.

Results showing echo canceller performance for the fullband echo cancellation case are shown in Fig. 15. The top part shows the canceller output versus time, whereas the bottom part shows echo return loss enhancement (ERLE) in decibels. The ERLE is defined as the ratio of the averaged powers after and before cancellation. Lower ERLE implies improved echo cancellation. In each of Figs. 1517, power levels have been obtained by averaging the respective signals over a sliding window of 200 samples in duration. It is seen from Fig. 15 that the convergence rate for the fullband case is not as fast as desired: just under 6 s are required to achieve an ERLE level of 30 dB. Further, the computational load is heavier than necessary. It is shown [18] that this configuration requires 32 MIPS5 at an 8 kHz sampling rate to implement. Simulation results for the method proposed by [5] are shown in Fig. 16. This method requires a critically downsampled PR or
5MIPS stands for millions of instructions per second. An instruction includes one real multiply plus one real add.

REILLY et al.: COMPLEX SUBBAND DECOMPOSITION

2741

Fig. 16.

Echo canceller results for the method of [5].

Fig. 17.

Echo canceller results for the complex subbanding case.

NPR filterbank with analysis and synthesis filters having conventional double-sided frequency responses. In principle, such matrix of a configuration requires a densely connected adaptive filters in the subbands to completely eliminate aliasing errors at the output. However, in practice, the authors have shown that ideal performance may be closely approximated using only a tridiagonal structure. This implies the adaptive filtering process in each subband requires crossfilters from its adjacent subbands, as well as the adaptive filter in the subband of interest (except, of course, for the end subbands, which require only one adaptive filter). The analysis and synthesis filters were generated by cosine modulation of a lowpass quadrature mirror filter prototype, designed according to the method of [1], to yield a PR filterbank. Here, the filterbank coefficients are real, and therefore, we de. We have used and . The refine sults of Fig. 16 show that the convergence behavior of this technique is not significantly improved over the fullband case. This is due to the additional parameters in the crossfilters that must

be estimated, thus slowing convergence. Further, the ERLE level obtained using this method is degraded over that obtained with the fullband approach (approximately 22 db versus 36 dB, respectively, over the final 1-s epoch). However, this structure does show a significantly improved MIP count over the fullband implementation. It is shown in [18] that this configuration requires 11.7 MIPS for the adaptive filtering operation and an additional 0.3 MIPS for the filterbank overhead (using a discrete cosine transform implementation), to give a total of 12.0 MIPS. Thus, this subband configuration offers a substantial saving in computational requirements over the original fullband implementation. The results using the proposed complex subband decomposition are shown in Fig. 17. In this case, we have used a downsampling rate of . Here, we use the same filterbank configuration as in Section IV-A. Because of the diagonal property of the decomposition, we require only one adaptive filter per subband. The significant improvement in convergence speed offered by this approach in comparison to the other two methods is apparent from the figure. A 30-dB cancellation level is achieved in just over 3 s. However, the point to be emphasized is the reduction in MIP count afforded by the proposed method. Here, the total MIP count is 7.9 [18], including filterbank overhead. This attractive MIP count figure is achieved despite the penalty of complex arithmetic. This penalty is abated by the use of the higher downsampling rate and to the fact that only one adaptive filter is required per subband, as opposed to nearly three for the method of [5]. The simulation scenario for the method of [7] is similar to that presented here. The simulation results of [7] indicate roughly equivalent performance as is achieved with the proposed singlesided subbanding scheme. We therefore observe that proposed single-sided subbanding approach to acoustic echo cancellation offers a competitive alternative to this problem, both in terms of performance and computation. Further, it offers the designer improved flexibility in filterbank design and in the choice of downsampling rates : an option that is not available with many other methods. We now demonstrate the performance of the delayless subbanded adaptive filter implementation. The following experiment is identical to the subbanded AEC implementation deof Fig. 6 are scribed previously, except the filter coefficients computed and updated using (33) from the coefficients inside samples in the downsampled time the subbands every scale. Results showing ERLE for both the subbanded and delayless implementations in decibels are shown in Fig. 18. Power levels at each time interval are determined by averaging over the previous 200 samples, as before. As can be seen, the delayless configuration experiences approximately 5-db degradation in suppression performance relative to the subbanded configuration but, nevertheless, has a more-than-acceptable level for practical applications. The overall echo suppression level relative to the remote signal level over the entire epoch from 4 to 9 s for the subbanded configuration is approximately 30.36 dB and 25.52 dB for the delayless configuration. Implementation Issues: Since we have used uniform DFT filterbanks, they can be implemented very efficiently using the

2742

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 11, NOVEMBER 2002

decrease, also increases so that (6) is satisfied well. This implies more memory for the filter coefficients. This can become a critical issue in applications where cost is important. Thus, is chosen as a tradeoff between performance and cost. In AEC appears typical, with the correapplications, the choice sponding set as large as possible. V. CONCLUSIONS In this paper, we have shown that the use of single-sided filterbanks leads to near elimination of aliasing error, which in turn permits a diagonal decomposition of a broad class of large problems into smaller, more tractable pieces. The application of filterbank theory is therefore significantly expanded. We have shown there exists a many-toone correspondence between the set of subband components and the model response of interest and that there exists an equivalence between convolution outside the subbands and convolution inside the subbands. The filterbank design process is simplified over conventional methods since the filters are not constrained by the need to cancel aliasing errors. The downsampling rate for a fixed filterbank is variable, up to the number of subbands for ideal filter responses. These results lead to new insights into the way signal processing algorithms can be implemented. The method has been verified using the acoustic echo cancellation problem. Performance of the single-sided subbanding technique has been shown to exceed or meet that of previous AEC methods, both with respect to convergence and with respect to computational requirements. The complex subband decomposition leads, naturally, to a delayless AEC implementation, whose performance has been verified to be almost equivalent to the conventional subbanded implementation. ACKNOWLEDGMENT The authors wish to acknowledge and thank the following institutions for their support of this project: Mitel Corporation, Kanata, ON, Canada; the Centre for Information Technology Ontario (CITO); and the Natural Sciences and Engineering Research Council of Canada (NSERC). REFERENCES
[1] P. P. Vaidyanathan, Multirate Systems and filterbanks. Englewood Cliffs, NJ: Prentice-Hall, 1993. [2] B. W. Gillespie, R. S. Malver, and D. A. F. Florencio, Speech dereverberation via maximumkurtosis subband adaptive filtering, in Proc. ICASSP, Salt Lake City, UT, May 2001. [3] W. Kellerman, Strategies for combining acoustic echo cancellation and adaptive beamforming microphone arrays, in Proc. ICASSP, Munich, Germany, Apr. 1997. [4] S. J. Elliot, I. M. Stothers, and P. A. Nelson, A multiple error LMS algorithm and its application to the active control of sound and vibration, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP35, pp. 14231434, Oct. 1987. [5] A. Gilliore and M. Vetterli, Adaptive filtering in subbands with critical sampling, IEEE Trans. Signal Processing, vol. 40, pp. 18621875, Aug. 1992. [6] Q. Jin, Z.-Q. Luo, and K. M. Wong, Optimum filterbanks for signal decompositions and its application in adaptive echo cancellation, IEEE Trans. Signal Processing, vol. 44, pp. 16691679, July 1996. [7] S. S. Pradhan and V. U. Reddy, A new approach to subband adaptive filtering, IEEE Trans. Signal Processing, vol. 47, pp. 655664, Mar. 1999.

Fig. 18.

ERLE versus time. Solid: Delayless method. Dashed: Subbanded configuration.

TABLE II COMPUTATIONAL COMPLEXITY OF THE VARIOUS IMPLEMENTATIONS

polyphase representation, which is discussed at length in [1]. The computational cost for one polyphase filterbank for our purreal multiply/adds per sample poses is (MAPS). Here, we assume that the proposed scheme is to be implemented on a machine that can perform a multiply/add operation in one instruction. In this case, a complex multiply can be configured as four real multiply/add operations. Note that by sample, we refer to the time scale outside the subbands. The conventional NLMS algorithm can be implemented in approximately 2 real MAPS, where is the length of the filter. On this basis, we can compose Table II, which gives the approximate computational complexity of the various AEC implementations in MAPS, using NLMS in the subbands, including the filterbank overhead (if appropriate). In Table II, is the prototype filter length. Recall the quantity in the second row is . The last term in the delayless expression results from evaluation the fast convolution of with . Here, the filter length is divided into blocks of . FFTs are then performed on each length such that is updated from its block as discussed in [13]. The filter subband coefficients periodically every samples. To convert the entries in Table II to MIPS, we multiply the respective figure 10 , where is the sampling rate at by the quantity the input (8 kHz for our experiments). From Table II, the computational expense and, as we have seen earlier, the convergence rate, both improve with increasing . Thus, ideally, we would like to see as large as possible. must also increase with increasing in order to However, satisfy (9). Since the bandwidth of the prototype filter must then

REILLY et al.: COMPLEX SUBBAND DECOMPOSITION

2743

[8] J. Reilly, M. Seibert, M. Wilbur, and N. Ahmadvand, The decomposition of large problems using single-sided subbanding, in Proc. ICASSP, Istanbul, Turkey, June 2000. [9] S. Weiss, M. Hartneck, and R. W. Stewart, Design of near perfect reconstruction oversampled filterbanks for subband adaptive filters, IEEE Trans. Circuits Syst. II, vol. 46, pp. 10811085, Aug. 1999. [10] W. H. Chin and B. FarhangBoroujeny, Subband adaptive filtering with realvalued subband signals for acoustic echo cancellation, Proc. Inst. Elect. Eng. Vis. Image Signal Process., vol. 148, no. 4, pp. 283288, Aug. 2001. [11] M. Thomlinson and K. M. Wong, Techniques for the digital interfacing of t.d.m.f.d.m. systems, Proc. Inst. Elect. Eng., vol. 123, no. 12, pp. 12851292, Dec. 1976. [12] M. G. Bellanger, J. L. Daguet, and G. P. Lepagnoi, Interpolation, extrapolation, and reduction of computation speed in digital filters, IEEE Trans. ACoust. Speech Signal Processing, vol. ASSP-22, pp. 231235, Aug. 1974. [13] D. R. Morgan and J. C. Thi, A delayless subband adaptive filter architecture, IEEE Trans. Signal Processing, vol. 43, pp. 18191830, August 1995. [14] R. Merched, P. S. R. Diniz, and M. R. Petralia, A new delayless subband adaptive filter structure, IEEE Trans. Signal Processing, vol. 47, pp. 15801591, June 1999. [15] Y. Bendel, D. Burshtein, O. Shalvi, and E. Weinstein, Delayless frequency domain acoustic echo cancellation, IEEE Trans. Signal Processing, vol. 49, pp. 589597, July 2001. [16] R. E. Crochiere and L. R. Rabiner, Multirate Digital Signal Processing. Englewood Cliffs, NJ: Prentice-Hall, 1983. [17] T. Kailath, Linear Systems. Englewood Cliffs, NJ: Prentice-Hall, 1980. [18] M. T. Seibert, Acoustic echo cancellation using single-sided subbanding, M.Eng. thesis, McMaster Univ., Dept. Elect. Comput. Eng., Hamilton, ON, Canada, 1999. [19] M. Wilbur, Application of subbanding methods to the decimation of large problems, M.Eng. thesis, McMaster Univ., Dept. Elect. Comput. Eng., Hamilton, ON, Canada, 2000. [20] S. Haykin, Adaptive Filter Theory, 3 ed. Englewood Cliffs, NJ: Prentice-Hall, 1996. [21] H. Cho, D. R. Morgan, and J. Benesty, An objective technique for evaluating doubletalk detectors in acoustic echo cancelers, IEEE Trans. Speech Audio Processing, vol. 7, pp. 718724, Nov. 1999.

Matt Wilbur (M94) was born in Ontario, Canada. He received the B.Eng.S. and M.Eng. degrees from McMaster University, Hamilton, ON, in 1998 and 2000, respectively. His thesis research investigated the use of subbanding for the decimation of large problems and the convex design of prototype filters for filterbanks. He is currently working as a radio frequency integrated circuit design engineer for Sirenza Microdevices, Ottawa, ON, on high-performance RF signal processing circuits for wireless basestation transceiver systems. Mr. Wilbur received an NSERC post-graduate scholarship and the prestigious Fessenden Postgraduate Scholarship from CRC/Industry Canada.

Michael Seibert received the B.Eng. degree in computer engineering and the M.Eng. degree in electrical engineering from McMaster University, Hamilton, ON, Canada, in 1997 and 1999, respectively. In 1999, he joined Mitel Semiconductor (now Zarlink Semiconductor), Ottawa, ON, to develop advanced echo cancellation algorithms. His work continues to focus on the development and integration of speech enhancement algorithms for Zarlinks high-density voice processing line.

James P. Reilly (S76M80) received the B.A.Sc. degree in electrical engineering from the University of Waterloo, Waterloo, ON, Canada, in 1973 and the M.Eng. and Ph.D. degrees from McMaster University, Hamilton, ON, in 1977 and 1980, respectively, both in electrical engineering. From 1973 to 1975, he worked in the CATV industry, designing RF distribution equipment and, from 1980 to 1985, at then Bell-Northern Research, Ottawa, ON, where he was involved in the development of terrestrial microwave radio and multiplexing systems for digital transmission systems. He joined the faculty at McMaster University in 1985 as an associate professor and was promoted to full professor in 1992. His research interests are in several aspects of signal processing, specifically blind signal separation, blind signal identification, array signal processing, and Bayesian methods. Dr. Reilly is a registered professional engineer in the province of Ontario.

Nima Ahmadvand received the B.S. and M.S. degrees in electrical engineering from Sharif University of Technology, Tehran, Iran, in 1988 and 1991, respectively, and Ph.D. degree in electrical and computer engineering from McMaster University, Hamilton, ON, Canada, in 1997. His Ph.D. research included media access control protocol and architecture designs for WDM cross-connect networks, where a new class of multiaccess multichannel optical communication networks is introduced. From 1997 to 1998, he worked as Post-Doctorate Fellow at the Communications Research Laboratory, McMaster University, in the field of digital signal processing with a focus on blind deconvolution using subbanding techniques. In 1998, he joined the Advanced Wireless Technology group at Nortel Networks, Ottawa, ON, Canada, where he was involved in the design of the third-generation wireless code division multiple access networks. Since 2000, he has been working at Peleton Photonic Systems, Ottawa, ON, as the Vice President of Research and Development, leading the R&D design team in optical networking subsystems. His current research interests include optical networking protocols, photonic passive and active module designs, laser sources, and telecommunication networks.