MULTI-CHANNEL ECHO CANCELLATION WITH ROUND ROBIN REGULARIZATION
by Jack W Stokes
and JohnC Piatt
TECHNICAL FIELD
The present invention relates in general to signal processing and more particularly to method and system of canceling the echo of multiple playback channels in a captured channel using round robin regulanzation
BACKGROUND OF THE INVENTION
Echo cancellation is an important element in a variety of applications In general, echo cancellation rs the digital cancellation of electncal and acoustrc echoes such that the echoes are attenuated or eliminated Echo cancellation is essential in applications such as communications systems, where it is used to improve sound quality Echo cancellation is used to overcome several different types of echoes, including hybrid echoes, caused by an impedance mismatch along an electncal line (including a telephone line), and acoustic echoes, caused by acoustic coupling of sound from a loudspeaker to a microphone These types of echoes appear in several different technologies, such as wireless telephony, hands-free telephony, teleconferencing systems, Internet telephony, and speech recognition systems By using echo cancellation, the sound quality and usefulness of these and many other technologies is improved

One type of echo cancellation is acoustic echo cancellation, which is used to cancel out the echoes of acoustic sound waves Typically, these echoes are formed when sounds emitted by one or more loudspeakers is picked up by one or more microphones Acoustic echoes can be quite noticeable and even annoying to a user
In general, the acoustic echo cancellation works by obtaining one or more playback signals, each going to corresponding loudspeakers and subtracting an estimate of the echo produced by that playback signal from the one or more microphone signals More specifically, the playback signals through this echo loop are transformed and delayed, background noise and possibly near end speech are added at the microphone, and a subtraction process for the echo cancellation is used The signal obtained after subtraction is called the error signal, and the goal is to minimize the error signal when no near end speech is present in the microphone signal
The heart of acoustic echo cancellation system is adaptive filtenng In general, an adaptive filter is used to identify or "learn" a transfer function of the room that contains the loudspeakers and microphones This transfer function will depend a great deal on the physical charactenstics of the room environment The adaptive filter works by taking the playback signal sent to the speakers and adjusting in a recursive manner some coefficients that represent an impulse response of the room The error signal, which is the estimated echo subtracted from the actual echo, is used to change the filter coefficients such that the error is minimized
Traditionally, the playback signals are each processed as a single stream of temporal samples, with a single delay line and a single filter To improve upon this, the playback signal can be split into subbands and a plurality of adaptive filters can be run in parallel, one adaptive filter per subband Changing the
length of the adaptive filters in the different subbands depending on the echo length in that subband in order to reduce the computational complexity is discussed in a paper by A Gilloire entitled "Experiments with Sub-band Acoustic Echo Cancellers for Teleconferencing" in 1987 International Conference on Acoustics, Speech, and Signal Processing, 1987, pp 2141-2144 From that paper, the adaptive filters for the lower subbands can be made longer in order to save CPU computation cycles because the bass tends to reverberate longer In the upper subbands, the filters can be shorter Thus, Gilloire's paper implied that longer adaptive filters in the lower subbands and shorter adaptive filters in the higher subbands can be used
To cancel the echoes in a captured signal, each subband of the playback signal is stored in a digital delay line, where the delayed subband signals are separated into taps At each tap, the playback signal is sampled The number of taps of a filter describes the length of the digital delay line For example, four taps means that the playback signal is sampled at the current frame, current frame-1, current frame-2, and current frame-3 Each of the delays is equal to the frame length (which can be, by way of example, approximately 16 milliseconds or 20 milliseconds) Thus, if the frame length is 16ms, and there are four taps (or a 4-long adaptive filter), and if the adaptive filters are implemented using adaptive subband filtering in the frequency domain, the playback signal is examined at a current frame, the frame 16ms earlier, the frame 32ms earlier, and the frame 48ms earlier than the current time
Each sample gets multiplied by the complex conjugate of a weight (called a tap weight, W), the multiplied weight is summed, and then is subtracted from the microphone signal Each tap weight is adjusted to minimize the output power Minimizing the output power suppresses as much of the speaker signal as possible, thereby reducing echoes
Acoustic echo cancellation was first used on monaural (or mono) systems FIG 1 illustrates a single channel, acoustic echo cancellation (AEC) system 100 used to process a mono, playback signal A mono playback signal x 105 is copied into equal multi-channel signals and then played through a right speaker 110 and a left speaker 120 Echoes 130,140 from each of the speakers 110, 120 are reflected off a wall 150 in a room and captured by a microphone 160 The microphone also captures desired speech 165 (such as from a teleconference participant) and background noise 170
The echoes 130,140, desired speech 165 and background noise 170 combine to construct a microphone signal y The microphone signal y is processed by a first analysis filterbank 175 and the playback signal x is processed by a second analysis filterbank 180 such that signals x and y are transformed from the time domain into frequency domain signals X and Y, respectively It is important to perform AEC in the frequency domain because the echoes in AEC are quite long and the adaptive filters converge more often and faster in the frequency domain than in the time domain It should be noted that the analysis filterbanks 175,180 can be implemented as any complex frequency domain transform such as a windowed (including the box window) fast Fourier transform (FFT) or, in an exemplary embodiment, a modulated complex lapped transform (MCLT)
The transformed X and Y signals are input to an AEC mono processor 185 that uses an adaptive filter to learn the transfer function of the room to minimize an error signal The processed signal is sent to a synthesis filterbank 190 that transforms the echo-reduced, frequency domain signal containing near end speech back to the time domain Note that the mono AEC processor in FIG 1 only uses a single adaptive filter per subband
FIG 2 is a detailed block diagram of the mono AEC processor 185 shown in FIG 1 for a single subband m and frame n The mono AEC processor 185
contains a single adaptive filter 200 for each subband An adaptive filter coefficient update 210 is used to update the processed coefficients of the subband adaptive filter 200 When the mono playback signal x is played to the speakers 110,120, as shown in FIG 1, the single adaptive filter 200 is implemented In a typical embodiment, the adaptive filter uses a normalized, least mean square (NLMS) algorithm having regularization The NLMS algorithm with regularization is set forth in detail below
When dividing one number by a second number, regularization is the process of adding or subtracting a small value to the denominator to ensure that the denominator never becomes zero, which in turn would cause the fraction to become infinite An alternative way to regularize the fraction is to set the denominator equal to some threshold if the denominator is positive and less than the threshold Likewise, if the denominator is negative, set the denominator to a negative threshold if it is less than the negative threshold
The single channel AEC system 100 shown in FIG 1 is only for removing echo from a mono playback signal One of the first papers to discuss extending AEC to stereo was a paper by M Sondhi and D Morgan entitled "Acoustic echo cancellation for stereophonic teleconferencing" in Proc IEEE Workshop Appls Signal Processing Audio Acoustics in 1991 However, while the NLMS algorithm works well for the mono AEC problem, NLMS performs poorly in the stereo (or other multi-channel) AEC problem This is because NLMS does not consider the cross-channel correlation of the multi-channel playback signal which significantly slows down the convergence of the adaptive filters
Sondhi and Morgan suggested using recursive least squares (RLS) instead of NLMS to solve the stereo AEC problem The RLS algorithm is an alternative algorithm for adjusting the parameters (or weights) of the adaptive filters The reason RLS works better than NLMS is that RLS tends to decorrefate the playback channels Since RLS recursively computes an estimate of the
inverse of a correlation matrix of the input speaker data, it can learn the correlation between the speaker channels and quickly converge to the correct solution Shondhi and Morgan, however, merely proposed potentially using the RLS algorithm instead of NLMS, but provided no detail
FIG 3 illustrates a stereo AEC system 300 used to process a stereo playback signal The stereo AEC system 300 shown in FIG 3 is a subband-based system, meaning that the speaker signal is split into a plurality of subbands and an adaptive filter is supplied for each subband The adaptive filters are run in parallel It should be noted that FIG 3 illustrates the AEC system 300 having a stereo playback signal, although AEC systems can be designed that also work for multi-channel playback signals In addition, a single microphone is illustrated in FIG 3, but the AEC system 300 is easily extendible to multiple microphones
Refernng to FIG 3, the stereo playback signal x is composed of two channels, a right stereo channel x(0) 302 and a left stereo channel x(1) 305, for the stereo playback case For the multi-channel AEC case, the N channel playback case, the signal would be composed of channels x(0) to x(N-1) The playback signals 302 and 305 are converted to an analog signal by a digital-to-analog converter (D/A) (not shown)
The multi-channel playback signal (which includes the stereo signal) can be created in several different ways FIG 4 illustrates the AEC system 300 of FIG 3 used with a voice communications system, such as Microsoft® Windows Messenger or voice chat for internet gaming In FIG 4, a digital, far-end mono speech signal 400 arrives from a source The speech signal 400 is mixed locally with some stereo audio sounds such as MusiCLeft 410 and MusiCRight 420, computer game sounds, or the computer's system sounds
Alternatively, FIG 5 illustrates the case where the multi-channel playback signal includes stereo music In this case, the MusiCLeft 410 is assigned to a multi-channel playback channel and the MusiCRight 420 is assigned to another multi-channel playback channel There is no far-end speech that is mixed with the multi-channel sound
In another alternate case, FIG 6 illustrates the case where the multichannel signal includes mono speech This situation shown in FIG 6 may be used for a Microsoft® Windows messenger system The mono speech 600 is copied to each of the playback channels, but the multi-channel playback signal is monaural
Refemng back to FIG 3, the playback signals 302, 305 next are played through a nght speaker 310 and a left speaker 320, respectively A first echo 330 and a second echo 340 are reflected off a wall 350 in a room (not shown) to produce echoes at the microphone 355 In the case of multiple microphones, a separate instance of the stereo AEC system 300 can process the signal captured from each microphone independently or one AEC algonthm could be processed on the mono output of a microphone array algorithm In addition to the echo from the speakers, the audio signal that is captured by the microphone 355 is also composed of a desired speech 360 and background noise 365 The analog audio signal captured by the microphone 355 is converted into a digital microphone signal, y, by an analog-to-digital converter (A/D) (not shown)
Acoustic echo cancellation is often performed using adaptive subband filtering based on a frequency domain transform such as the fast windowed transform (FFT) or the modulated complex lapped transform (MCLT) A first filterbank 370 and a second filterbank 375 convert each of the stereo playback signals x(0) and x(1) from the time domain to the frequency domain signals X(0) and X(1), respectively Likewise, a third analysis filterbank 380 converts the mono microphone signal y from the time domain to the frequency domain signal
Y The signals are processed by the stereo AEC processor 385 and the output Z is run through a synthesis filterbank 390 A time domain signal z with reduced echo then is output
FIG 7 is a detailed block diagram of the stereo AEC processor 385 for a single subband shown in FIG 3 The stereo AEC processor contains a first adaptive filter 700 for the first multi-channel playback signal X(0) and a second adaptive filter 710 for the second multi-channel playback signal X(1) Note that separate single channel filters are run in parallel on each subband independent of the other subbands As descnbed above with regard to FIG 3, the playback signals X(0), X(1) are processed and an adaptive filter coefficient update 720 ts used to update the processed coefficients of the single channel filters 700, 710 The frequency domain signal Z, with reduced echo, then is output
However, one problem with the RLS algorithm for computing the adaptive
filter weights is that it has a high computational complexity This complexity is on
the order of 0(2NA2 + 6N) compared to 0(2N) for the least mean squares (LMS)
where /V = C*L, C is the number of playback channels, and L is the adaptive filter
length in the subband Previously, this computational complexity of RLS
prohibited its use in AEC in practical systems A paper by B Hatty entitled,
"Recursive Least Squares Algorithms using Multirate Systems for Cancellation of
Acoustical Echoes" in 1990 International Conference on Acoustics, Speech, and
Signal Processing, 3-6 April, 1990, vol 2, pp 1145-1148, was one of the first
papers that discussed using a fast RLS (FRLS) for mono AEC FRLS increases
the speed and decreases the complexity of RLS by avoiding the use of a
correlation matrix (or any other types of matnces) One problem, however, with
FRLS is that it is quite unstable As a result of this instability, the FRLS algorithm
can quickly diverge There have been several attempts to improve the stability of
FRLS However, to date, no one has come up with a satisfactory solution for the
multi-channel AEC problem Hatty, in an attempt to improve the stability of
FRLS, proposed using a round robin scheme by resetting the entire FRLS
algonthm periodically in a band-by-band fashion What Hatty did was to completely reinitialized a band by throwing away the entire state of the algorithm and restarting it from scratch
The problem with this reset technique, however, is that this resetting caused echo leakthrough for the band being reset, due to the FRLS algonthm having to reconverge and relearn the transfer function of the room after each reset In addition, the Hatty technique caused distortion on the playback signal due to the fact that at any given time there was at least of portion of the algonthm is being reset
In 1995, J Benesty, J , et al in a paper entitled, "Adaptive Filtenng Algorithms for Stereophonic Acoustic Echo Cancellation" in Proc ICASSP'95, pp 3099-3102 used fast RLS (FRLS) to try and solve the stereo AEC problem However, the Benesty paper suggested using FRLS in the time domain instead of using adaptive subband filtenng
In another paper by J Benesty, D Morgan, and M Sondhi entitled, "A Better Understanding and an Improved Solution to the Problems of Stereophonic Acoustic Echo Cancellation" in Proc ICASSP97, pp 303-306, an update was proposed In the Benesty '97 paper, in order to decorrelate the left channel from the nght channel (which were very similar), Benesty added a nonlineanty to both channels In one implementation, Benesty added the positive "portion of the nonlineanty to one channel and the inverse (or negative) portion of the nonlineanty to the other channel This introduced nonlineanty forced the channels to be different enough that the adaptive filters could learn the individual paths In this way, the channels were decorrelated and made different enough so that the non-uniqueness problem associated with having to track the far-end transfer functions from the far-end person to far-end stereo microphones, as well as the near-end transfer functions from the near-end speakers to the near-end microphones, could be avoided
The problem with adding a nonlinearity to the signal (as is done in the Benesty '97 paper) is that adding any type of the nonlineanty tends to distort the signal Basically, adding a nonlinearity is adding distortion to the signal Adding distortion, however, is undesirable if the AEC system is to work well with a system that involves music playback Ideally, for music playback, the signal should be free of distortion so that the music is played back faithfully
In the paper by P Eneroth, S Gay, T Gansler, and J Benesty entitled "A Real-Time Implementation of a Stereophonic Acoustic Echo Canceller" in IEEE Trans On Speech and Audio Processing, Vol 9 no 5, July 2001, pp 513 - 523, a solution to the stereophonic AEC problem is proposed using FRLS in subbands and adding non-linearities to the playback channels This paper attempts to increase stability by running parallel structures of the FRLS algorithm so that when one of the structures "blows up" or goes unstable, they can fall back on to another structure that is less than optimal This implementation helps them reinitialize the algonthm
In 1990, when the Hatty '90 paper proposed using FRLS for adaptive
subband AEC processing, microprocessors were much slower than today's
s
microprocessors As a result, RLS was not a practical solution for the multichannel AEC problem However, with the significant increase in speed of modem microprocessors, RLS can now be used However, the RLS algonthm will become unstable and diverge if the correlation matrix of the multi-channel playback signal becomes singular
Therefore, what is needed is an echo cancellation system and method that can be used for a multi-channel playback signal In addition, what is needed is a multi-channel echo cancellation system and method that avoids the use of FRLS to prevent the system from becoming unstable In addition, what is needed is a multi-channel echo cancellation system and method that avoids the use of
adding distortion to the playback signal What is also needed is a multi-channel echo cancellation system and method that avoid and overcomes the problems of the RLS algorithm discussed above to effectively eliminate echo while retaining a faithful reproduction of the original signal
SUMMARY OF THE INVENTION
The invention disclosed herein includes a multi-channel echo cancellation system and method having round robin regularization The present invention can be used on both mono and multi-channels signals, and overcomes the above-mentioned deficiencies of current adaptive filter algorithms (such as RLS) In particular, the round robin regulanzation ensures that the adaptive filter algorithm does not become singular and blow up by regulanzing the correlation matnx of the adaptive filter algonthm This regulanzation is performed in a round robin fashion In other words, the inverse correlation matrix of the subband playback data associated with each adaptive filter is regulanzed, in a round robin sequence, to prevent the inverse correlation matnx from diverging In an alternative embodiment, each inverse correlation matrix is examined for divergence If the inverse correlation matrix starts to diverge, then the matrix is regularized
The multi-channel round robin regulanzation echo cancellation method includes obtaining a captured signal containing echo of a plurality of playback signals The plurality of playback signals are decomposed into a plurality of subbands, and a plurality of adaptive filters are applied to the subbands Each adaptive filter has an inverse correlation matnx Next, each one of the plurality of adaptive filters is selected in a round robin sequence, so that every round each of the filters is selected The inverse correlation matrix associated with each selected adaptive filter then is regularized The regularized adaptive filter then is used to remove the echo from the captured signal
Regulanzation of the inverse correlation matnx may be achieved by regulanzing the correlation matnx Thus, the inverse correlation matnx may be inverted, the correlation matrix then regularized, the regularized correlation matrix then inverted to yield a regularized inverse correlation matrix
The correlation matrix may be regulanzed in at least two ways One way is to add a small value to the diagonal elements of the correlation matrix every time the matnx is regulanzed This ensures that the inverse correlation matrix does not diverge Another way is to define a threshold value and examine each diagonal element of the correlation matnx Anytime there is a danger of division by zero, then that diagonal element of the correlation matnx is set equal to the threshold value If the correlation matrix diagonal element is greater than the threshold value, nothing is done to that element in the correlation matrix
Regulanzation is implemented in a round robin manner Each subband in
the system is selected in turn so that the adaptive filter for that subband can be
examined This examination may include determining whether regulanzation is
necessary at that time The round robin scheme is used to make sure that each
subband is examined on a regular basis The round robin scheme may be
implemented such that a single subband is examined for each frame
Alternatively, the round robin scheme may be implemented so that several or
even all subbands per frame are examined In addition, the round robin scheme
may be implemented so that one subband every several frames is examined

Another feature of the multi-channel echo cancellation system and method disclosed herein ts dynamic switching between monaural and multi-channel echo cancellation In particular, this feature provides dynamic switching between a mono AEC algorithm (such as normalized least mean square (NLMS)) and the invention's AEC algorithm for multiple channels with efficient reinitialization of the invention's echo cancellation algorithm The switching is dynamic, such that little or no echo leaks through because of the efficient reinitialization of the RLS
algorithm This is achieved by shanng state vanables and using a novel estimation technique to estimate the inverse correlation matrix for the RLS algonthm using the round robin regulanzation This allows the invention to switch from NMLS (mono) back to RLS (multi-channel) without any echo leakage
Still another feature of the multi-channel echo cancellation system and method disclosed herein is mixed processing for lower and upper subbands In particular, in order to lower CPU resources requirements, round robin regulanzation RLS is used on the lower subbands while NLMS processing is used on the upper subbands Both the plurality of playback signals and the captured signal are first divided into lower subbands (containing the lower frequencies of the signal) and the higher subbands (containing the higher frequencies of the signal) On the lower subbands, the RLS round robin regulanzation echo cancellation disclosed herein is used On the higher subbands, the NLMS processing is used NMLS uses less CPU resources and is used at the higher subbands because the stereo effect is attenuated at these frequencies At lower frequencies, where the stereo effect is most noticeable, the supenor RLS round robin regulanzation of the invention is used
BRIEF DESCRIPTION OF THE DRAWINGS
The present invention can be further understood by reference to the following description and attached drawings that illustrate aspects of the invention Other features and advantages will be apparent from the following detailed descnption of the invention, taken in conjunction with the accompanying drawings, which illustrate, by way of example, the principles of the present invention
Referring now to the drawings in which like reference numbers represent corresponding parts throughout
FIG 1 illustrates a prior art single channel, acoustic echo cancellation (AEC) system used to process a mono, playback signal
FIG 2 is a detailed block diagram of the mono AEC processor shown in FIG 1
FIG 3 illustrates a pnor art stereo AEC system used to process a stereo playback signal
FIG 4 illustrates the AEC system of FIG 3 used with a voice communications system
FIG 5 illustrates the AEC system of FIG 3 where the multi-channel playback signal includes stereo music
FIG 6 illustrates the AEC system of FIG 3 where the multi-channel playback signal includes mono speech
FIG 7 is a detailed block diagram of the stereo AEC processor shown in FIG 3
FIG 8 is a block diagram illustrating an exemplary implementation of the multi-channel round robin regulanzation echo cancellation system and method disclosed herein
FIG 9 illustrates an example of a suitable computing system environment in which the multi-channel round robin regulanzation echo cancellation system and method shown in FIG 8 may be implemented
FIG 10 is a general flow diagram illustrating the general operation of the multi-channel round robin regulanzation echo cancellation system shown in FIG 8
FIG 11 is a general flow diagram illustrating an exemplary implementation of a round robin scheme of the multr-channel round robin regulanzation echo cancellation method shown in FIG 10
FIG 12 is a block/flow diagramlrilustrating an exemplary implementation of the multi-channel round robin regulanzation echo cancellation system and method shown in FIG 8
FIG 13 is a block/flow diagram illustrating another exemplary implementation of the multi-channel round robin regulanzation echo cancellation system and method shown in FIG 8
FIG 14 illustrates the mixed lower/upper subbands processing using the multi-channel round robin regulanzation method descnbed herein
DETAILED DESCRIPTION OF THE INVENTION
In the following descnption of the invention, reference is made to the accompanying drawings, which form a part thereof, and in which is shown by way of illustration a specific example whereby the invention may be practiced It is to be understood that other embodiments may be utilized and structural changes may be made without departing from the scope of the present invention
I Introduction
Current techniques for echo cancellation of signals suffer from a variety of shortcomings For example, normalized least mean squares (NLMS) is useful when the signal is a monaural signal, but does not work well on correlated, multichannel playback signals Processing the multi-channel playback signal with a non-lineanty decorrelates the individual channels but adds undesirable distortion to the playback signals This is unacceptable when the playback signals contains music that a user would like to hear in its original, undistorted version Recursive least squares (RLS) and fast recursive least squares (FRLS) also help to decorrelate multi-channel playback signals However, FRLS can be extremely unstable RLS is computationally intensive and can become unstable for highly correlated playback signals
The multi-channel round robin regulanzation echo cancellation method and system disclosed herein overcome these problems of current methods The multi-channel round robin regulanzation echo cancellation method processes a multi-channel playback signal (such as an acoustic signal) by using a plurality of recursive least squares (RLS) adaptive filters and using a round robin
regularization scheme to ensure that the RLS algorithm remains stable Stability is ensured by regulanzing the inverse correlation matnx in the RLS algorithm in a round robin manner on a subband-by-subband basis Essentially, the round robin regularization ensures that the determinant of the inverse correlation matnx for the RLS adaptive filters on each of the plurality of subbands does not "blow up" (or attain a large value)
The multi-channel round robin regulanzation echo cancellation method and system provides a novel way to prevent the inverse correlation matnx from diverging thereby causing the RLS adaptive filter to diverge The inverse correlation matrix is penodically regulanzed for each subband based on a round robin scheme in order to ensure the algonthm does not diverge when the channels of the speaker signal are dependent
The multi-channel round robin regulanzation echo cancellation method and system also includes a novel feature that allows dynamic switching between a multi-channel, subband echo cancellation process (such as the RLS round robin regularization method described herein) and single channel, subband echo cancellation process (using an adaptive filtering algorithm such as normalized least mean squares (NLMS)) The dynamic switching is achieved with efficient reinitialization of either process, to minimize distortion when switching from multichannel processing to monaural processing and vice versa The multi-channel round robin regulanzation echo cancellation method and system also includes another feature of one type of processing the lower subbands using RLS and NLMS on upper subbands This allows the multi-channel round robin regulanzation echo cancellation method and system to run in environments where processing power is limited
II General Overview
FIG 8 is a block diagram illustrating an exemplary implementation of the multi-channel round robin regulanzation echo cancellation system and method
disclosed herein ft should be noted that FIG 8 is one of several ways in which the multi-channel round robin regulanzation echo cancellation system and method may implemented and used

Refernng to FIG 8, multi-channel round robin regulanzation echo cancellation system 800 resides on a computing device 810 In FIG 8, playback signals P(1), P(2) to P(N) are generated internally (such as music from a compact disk) by the system 800 In an alternate embodiment, playback signals P(1), P(2) to P(N) are external and received (such as via input audio jacks) by the system 800 Each playback signal is played back over its respective speaker, so that P(1) is played back through speaker S(1), P(2) is played back through speaker S(2), and so forth fn addition, the playback signals P(1) to P(N) are obtained and input into the multi-channel round robin regulanzation echo cancellation system 800
As shown by the dotted lines in FIG 8, the output of the speakers S(1) to S(N) causes echoes 820, 830, 840 to be reflected off a wall 850 of a room (not shown) These echoes 820,830,840 are picked up by a microphone 860 In addition, the microphone picks up desired speech 870 and background norse 880 The signal from the microphone 865, also called the "captured signal", along with the outputs of the playback signal P(1) to P(N), are input to the multichannel round robin regulanzation echo cancellation system 800 After processing by the multi-channel round robm regulanzation echo cancellation system 800, a new output signal 890 is generated Output signal 890 is the captured signal 865 with echoes of speaker signals 820, 830, and 840 reduced
It should be noted that the invention is not limited to acoustic signals In general, there may be N transmitters S(1) through S(N) and one receiver 860, and the multi-channel round robin regulanzation echo cancellation system writ reduce the amount of transmitted signal in the output signal 890
It should also be noted that the invention is not limited to a single receiver or microphone 860 The computing device 810 may be connected to a plurality of receivers 860, producing a plurality of signals 865 Each of the plurality of signals 865 would be operated on by an single independent instance of the multi-channel round robin regulanzation echo cancellation system 800 Each of those single instances would then produce a single output with reduced echo 890
III. Exemplary Operating Environment
The multi-channel round robin regulanzation echo cancellation system and method are designed to operate in a computing environment and on a computing device, such as the computing device 810 shown in FIG 8 The computing environment in which the multi-channel round robin regulanzation echo cancellation system and method operates will now be discussed The following discussion is intended to provide a bnef, general description of a suitable computing environment in which the multi-channel round robin regulanzation echo cancellation system and method may be implemented
FIG 9 illustrates an example of a suitable computing system environment in which the multi-channel round robin regulanzation echo cancellation system and method shown in FIG 8 may be implemented The computing system environment 900 is only one example of a suitable computing environment and is not intended to suggest any limitation as to the scope of use or functionality of the invention Neither should the computing environment 900 be interpreted as having any dependency or requirement relating to any one or combination of components illustrated in the exemplary operating environment 900
The multi-channel round robin regulanzation echo cancellation system and method is operational with numerous other general purpose or special purpose computing system environments or configurations Examples of well known computing systems, environments, and/or configurations that may be suitable for use with the multi-channel round robin regulanzation echo cancellation system
and method include, but are not limited to, personal computers, server computers, hand-held, laptop or mobile computer or communications devices such as cell phones and PDA's, multiprocessor systems, microprocessor-based systems, set top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments that include any of the above systems or devices, and the like
The multi-channel round robin regulanzation echo cancellation system and method may be descnbed in the general context of computer-executable instructions, such as program modules, being executed by a computer Generally, program modules include routines, programs, objects, components, data structures, etc , that perform particular tasks or implement particular abstract data types The multi-channel round robin regulanzation echo cancellation system and method may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network In a distributed computing environment, program modules may be located tn both local and remote computer storage media including memory storage devices With reference to FIG 9, an exemplary system for implementing the multi-channel round robin regulanzation echo cancellation system and method includes a general-purpose computing device in the form of a computer 910
Components of the computer 910 may include, but are not limited to, a processing unit 920, a system memory 930, and a system bus 921 that couples vanous system components including the system memory to the processing unit 920 The system bus 921 may be any of several types of bus structures including a memory bus or memory controller, a peripheral bus, and a local bus using any of a vanety of bus architectures By way of example, and not limitation, such architectures include Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video
Electronics Standards Association (VESA) local bus, and Penpheral Component Interconnect (PCI) bus also known as Mezzanine bus
The computer 910 typically includes a vanety of computer readable media Computer readable media can be any available media that can be accessed by the computer 910 and includes both volatile and nonvolatile media, removable and non-removable media By way of example, and not limitation, computer readable media may compnse computer storage media and communication media Computer storage media includes volatile and nonvolatile removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data
Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by the computer 910 Communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a earner wave or other transport mechanism and includes any information delivery media
Note that the term "modulated data signal" means a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal By way of example, and not limitation, communication media includes wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared and other wireless media Combinations of any of the above should also be included within the scope of computer readable media
The system memory 930 includes computer storage media in the form of volatile and/or nonvolatile memory such as read only memory (ROM) 931 and random access memory (RAM) 932 A basic input/output system 933 (BIOS), containing the basic routines that help to transfer information between elements withm the computer 910, such as dunng start-up, is typically stored in ROM 931 RAM 932 typically contains data and/or program modules that are immediately accessible to and/or presently being operated on by processing unit 920 By way of example, and not limitation, FIG 9 illustrates operating system 934, application programs 935, other program modules 936, and program data 937
The computer 910 may also include other removable/non-removable, volatile/nonvolatile computer storage media By way of example only, FIG 9 illustrates a hard disk dnve 941 that reads from or wntes to non-removable, nonvolatile magnetic media, a magnetic disk dnve 951 that reads from or writes to a removable, nonvolatile magnetic disk 952, and an optical disk dnve 955 that reads from or writes to a removable, nonvolatile optical disk 956 such as a CD ROM or other optical media
Other removable/non-removable, volatile/nonvolatile computer storage media that can be used in the exemplary operating environment include, but are not limited to, magnetic tape cassettes, flash memory cards, digital versatile disks, digital video tape, solid state RAM, solid state ROM, and the like The hard disk dnve 941 is typically connected to the system bus 921 through a nonremovable memory interface such as interface 940, and magnetic disk dnve 951 and optical disk drive 955 are typically connected to the system bus 921 by a removable memory interface, such as interface 950
The drives and their associated computer storage media discussed above and illustrated in FIG 9, provide storage of computer readable instructions, data structures, program modules and other data for the computer 910 In FIG 9, for example, hard disk drive 941 is illustrated as storing operating system 944,
application programs 945, other program modules 946, and program data 947 Note that these components can either be the same as or different from operating system 934, application programs 935, other program modules 936, and program data 937 Operating system 944, application programs 945, other program modules 946, and program data 947 are given different numbers here to illustrate that, at a minimum, they are different copies A user may enter commands and information into the computer 910 through input devices such as a keyboard 962 and pointing device 961, commonly referred to as a mouse, trackball or touch pad
Other input devices (not shown) may include a microphone, joystick, game pad, satellite dish, scanner, radio receiver, or a television or broadcast video receiver, or the like These and other input devices are often connected to the processing unit 920 through a user input interface 960 that is coupled to the system bus 921, but may be connected by other interface and bus structures, such as, for example, a parallel port, game port or a universal serial bus (USB) A monitor 991 or other type of display device is also connected to the system bus 921 via an interface, such as a video interface 990 In addition to the monitor, computers may also include other peripheral output devices such as speakers 997 and printer 996, which may be connected through an output peripheral interface 995
The computer 910 may operate in a networked environment using logical connections to one or more remote computers, such as a remote computer 980 The remote computer 980 may be a personal computer, a server, a router, a network PC, a peer device or other common network node, and typically includes many or all of the elements descnbed above relative to the computer 910, although only a memory storage device 981 has been illustrated in FIG 9 The logical connections deptcted in FIG 9 include a local area network (LAN) 971 and a wide area network (WAN) 973, but may also include other networks Such
networking environments are commonplace in offices, enterpnse-wide computer networks, intranets and the Internet
When used in a LAN networking environment, the computer 910 is connected to the LAN 971 through a network interface or adapter 970 When used in a WAN networking environment, the computer 910 typically includes a modem 972 or other means for establishing communications over the WAN 973, such as the Internet The modem 972, which may be internal or external, may be connected to the system bus 921 via the user input interface 960, or other appropriate mechanism In a networked environment, program modules depicted relative to the computer 910, or portions thereof, may be stored in the remote memory storage device By way of example, and not limitation, FIG 9 illustrates remote application programs 985 as residing on memory device 981 It will be appreciated that the network connections shown are exemplary and other means of establishing a communications link between the computers may be used
IV. Operational Overview
The operation of the multi-channel round robin regulanzation echo cancellation system 800 shown in FIG 8 now will be discussed FIG 10 is a general flow diagram illustrating the general operation of the multi-channel round robin regulanzation echo cancellation system shown in FIG 8 The multichannel round robin regulanzation echo cancellation method begins by decomposing a plurality of playback signals into a plurality of subband signals (box 1000) Each of these signals can correspond to a channel of playback Next, a plurality of adaptive filters, each having an inverse correlation matnx, are applied to the plurality of subband signals (box 1010) Each of the plurality of adaptive fitters is selected in a round robin sequence, such that each of the filters is selected in turn (box 1020) The inverse correlation matnx associated with each selected adaptive filter then is regularized (box 1030) As explained in detail below, regulanzation means processing the inverse correlation matnx such
that the determinant of the correlation matrix is not close to zero The regulanzed adaptive filter then is used to reduce the echo in the captured signal 865 from the plurality of playback signals (box 1040)
V. Operational Details and Working Example
The multi-channei round robin regulanzation echo cancellation method shown in FIG 10 uses a round robin scheme and regulanzation to ensure that the adaptive filters in the echo cancellation algonthm do not diverge when the playback channels are dependent As noted above, the method includes implementing a round robin scheme of selecting each adaptive filter of a subband and then regularizing the inverse correlation matnx for that selected adaptive filter Together, the round robin scheme and the regulanzation process make up the round robin regulanzation process of the multi-channel round robin regulanzation echo cancellation method Various ways of implementing the round robin scheme and regulanzation will now be discussed The discussion includes exemplary working examples of tested implementations of a round robin scheme and a regulanzation process In particular, a exemplary working example is presented of the multi-channel round robin regulanzation echo cancellation method using round robin regulanzation in a recursive least squares (RLS) algonthm It should be noted that this working example is only one way in which the round robin regulanzation process may be implemented
Mathematical Descnption of the Adaptive Filtering Processes
For a single channel of playback data, the NLMS algonthm with regulanzation is as follows
For each frame of audio data, n = 1,2, , and each subband m = 0 MA, compute,
(Equation Removed)
where is the error signal between the microphone signal, Y(n,m) is the estimated echo, £ is the conjugate of the error signal,
(Equation Removed)
is the estimated echo filter coefficients in the transform domain, L is the number of taps in the estimated echo filter,
(Equation Removed)
is a column vector of the most recent L frequency domain playback samples, \x is the step size, PNLMS IS the regulanzation factor, and Pest(n,m) is the estimate of the power in the mm band For example, pESt(n,m) can be found as
(Equation Removed)
where
(Equation Removed)
is the square of the magnitude of the complex speaker signal X(n,m) and a is the averaging parameter
When dividing one number by a second number, regulanzation is the process of adding or subtracting a small value to the denominator to ensure the denominator never becomes zero and causes the fraction to become infinite In equation (2), PNLMS IS the regulanzation factor that ensures that the denominator never goes to zero PNLMS IS added to pEst(n,m) since PEst(n,m) is a power estimate and is always positive An alternative way to regulanze the fraction is to set the denominator to some threshold if the denominator is positive and less than the threshold
One problem with using the NLMS algonthm for multi-channel playback signals is that NLMS does not work very well This is because the cross-channel correlation of the multi-channel playback signal significantly slows down the convergence of the adaptive filters Recursive least squares (RLS) is an alternative algorithm for adjusting the parameters (or weights) of the adaptive filters One of the main benefits of RLS is that it tends to decorrelate the
channels Since RLS recursively computes an estimate of the inverse of a correlation matrix of the multi-channel input speaker data, it can learn the correlation between the speaker channels and quickly converge to the correct solution For multi-channel speaker signals, the RLS algonthm is as follows (Equation Removed)

For each frame n ~ 1,2, and subband m - 0 W-1, compute

(Equation Removed)
where 8 is a small constant, I is the identity matrix,
(Equation Removed)

is the multi-channel Kalman gain vector, C is the number of playback channels, L is the filter length in the subband,
(Equation Removed)
is the multi-channel speaker input vector, P(n,m) is the inverse of the input speaker signal correlation matrix R(n,m),
(Equation Removed)
X is the exponential forgetting factor,
(Equation Removed)
is the weight vector, and is the conjugate of the error
It should be noted that the RLS adaptive filter stores and updates the inverse correlation matrix P for every subband This is referred to m this document as "having an inverse correlation matrix"
Round Robin Scheme
In general, the round robin scheme as used in the multi-channel round robin regulanzation echo cancellation method ensures that each subband in the system is selected so that the adaptive filter for that subband can be regularized Alternatively, the adaptive filter can be examined to determine whether regulanzation is necessary at that time The round robin scheme is used to make sure that each subband is examined on a regular basis
The round robin scheme may be implemented in a variety of different ways By way of example, and not limitation, a round robin scheme can be descnbed as where a single subband is regularized each frame Alternatively, the round robin scheme may include regularizing several or even all subbands per frame In another implementation, the round robin scheme may include regulanzing one subband every several frames Depending on the actual round robin scheme implemented, a counter that is pointing to the current subband being examined is updated to point at the next subband or group of bands being examined and reset back to the first band if necessary
FIG 11 is a general flow diagram illustrating an exemplary implementation of a round robin scheme of the multi-channel round robin regulanzation echo cancellation method shown in FIG 10 As shown in FIG 11, the round robin scheme begins by inputting a current frame (box 1100) One subband is then selected for regulanzation (box 1110) In this embodiment, each subband is selected in turn as a new current frame is input However, in alternative embodiments, a single subband may only be selected every few frames, or several subbands may be selected per frame
The inverse correlation matnx for the selected subband is inverted to obtain the correlation matrix for the selected subband (box 1120) The correlation matrix is then regulanzed (box 1130) In an alternative embodiment, box 1130 first examines the correlation matnx to determine whether it needs to be regularized, and then regularizes only when necessary Next, the regularized correlation matnx is re-inverted (box 1140), and the regulanzed inverse correlation matrix for the selected subband is stored (box 1150) Finally, all of the adaptive filters for all of the subbands are applied to the plurality of subband, playback signals to reduce echo in the captured signal (box 1160) Control then flows back to box 1100, which obtains the next current frame
Round Robin Regulanzation
In order to more fully understand the round robin regulanzation process disclosed herein, the following discussion is provided The discussion includes an exemplary working example of a tested implementation of the round robin regulanzation process in a recursive least squares (RLS) algorithm It should be noted that this working example is only one way in which the round robin regulanzation process may be implemented
In general, the round robin regulanzation process keeps the correlation matnx for each adaptive filter from becoming singular This regulanzation ensures that the inverse correlation matrix will not become unstable In an exemplary implementation, round robin regulanzation was used with an RLS algorithm The details of this implementation are as follows
In general, RLS is a fast method to solve the normal equation w° = R-1p
where w° is the optimal weight vector, R is the correlation matrix for the input speaker data given in equation (9), and p is the cross-correlation vector between the input speaker data and the microphone data
(Equation Removed)
If R starts to become singular, then the values in its inverse become very large and a valid estimate of the weight vector cannot be found The correlation matrix becomes singular rf either its rows or the columns are dependent This condition occurs when one channel of audio data is a linear transformation of another channel For example, the correlation matrix is singular if one channel is equal to another channel multiplied by some gain This example corresponds to a mono signal being panned across the speakers in the system For a purely mono signal, the signals played to each speaker are identical and the gam between the channels equals 1 In both of these examples, the correlation matrix quickly becomes singular depending on the value of the forgetting factor and how long the mono signal has been played
In order to prevent the correlation matnx from becoming singular, the round robin regulanzatron process is used to regularize the correlation matrix for a particular subband As descnbed above, regulanzatron involves ensuring that the denominator of a fraction never gets too close to zero, thereby preventing the value of the fraction from becoming too large

Regulanzatron of the correlation matnx R, can be achieved in a vanety of different ways Since the diagonal elements of the correlation matrix are non-negative, the invention only needs to regularize positive or zero elements of the diagonal correlation matnx By way of example and not limitation, the correlation matrix can be regulanzed by doing the following (1) adding a small value to the diagonal elements of the correlation matnx, and (2) setting values on the diagonal that are close to zero to a small threshold The latter technique defines a threshold value and examines each element of the correlation matnx Anytime there is a danger of division by zero (such as when the element being examined gets smaller than the threshold value), then that element of the correlation matrix
is set equal to the threshold value The threshold value is a small number that is greater than zero If the correlation matrix element is greater than the threshold value, nothing is done to that element in the correlation matrix
The first regulanzation technique that involves adding a small value to the diagonal elements of the correlation matrix can be summarized mathematically as
(Equation Removed)
where R(n,m) is the correlation matnx, P-1(n,m) is the inverse of the inverse correlation matnx, ßRLS IS the regulanzation factor, and I is the identity matrix PRLS can be a small constant (such as 1500 with 16-bit input data), or PRLS can be chosen using other non-uniform techniques By regularizing each subband in a round robin scheme, the central processing unit (CPU) consumption involved with the two matnx inverses can be minimized
As described above, the round robin scheme can be implemented where a single band or several bands are regularized per frame or one band is regularized every several frames Depending on the actual round robin scheme, the counter which points to the current band to be regularized is updated to point to the next band or group of bands to be regularized and reset back to the first band if necessary Since R and P are symmetric, a Cholesky decomposition can be used to find R from P-1 instead of LU decomposition, which requires twice as many operations However, in a test implementation the LU decomposition was used with partial pivoting to compute the inverse instead of Cholesky decomposition due to its numencal properties Furthermore, this regulanzation
process was computed using double precision mathematics to achieve acceptable performance for the matrix inversions
As a result, the new RLS algonthm with round-robin regularization is as follows
(Equation Removed)
For each frame rt-1,2, and subband m~0 M-1, compute
(Equation Removed)

end end
RoundRobmCount = RoundRobinCount + 1 If (RoundRobmCount == MaxRoundRobinCount)
RoundRobmCount = 0 end
In this embodiment of the invention, MaxRoundRobinCount is selected to be equal to the number of subbands in the mixed subband processing descnbed
below Selecting a value for MaxRoundRobinCount depends upon the value of the exponential forgetting factor X The value of X is usually chosen as a trade off between convergence accuracy of the RLS solution and tracking speed If X is very close to 1, the RLS algorithm will obtain very accurate tap weights and hence cancel most of the echo provided nothing moves in the near end room However, if something or someone does move, then the RLS algorithm cannot track the changes in the acoustic environment quickly The value of MaxRoundRobinCount should be chosen such that the inverse correlation matrix of the speaker signal is regulanzed often enough so that MaxRoundRobinCount does not reach a value that is too small If MaxRoundRobinCount IS too small, the inverse correlation matrix will become singular between round robin updates In this case, the inverse correlation matnces of multiple subbands will need to be regularized per frame to increase the frequency of when the inverse correlation matrix for each subband is regulanzed
FIG 12 is a block/flow diagram illustrating an exemplary implementation of the multi-channel round robin regulanzation echo cancellation system 800 and method shown in FIG 8 This exemplary implementation uses the round robin regulanzation technique set forth above FIG 12 shows the implementation using one multi-channel filter per subband, operating on aggregated data In particular, refernng to FIG 12, subband signals from the multiple playback channels are aggregated into a single vector (box 1200) Using this aggregated vector, the inverse correlation matnx is updated using equation (7) and the round robin regulanzation (box 1210) The aggregated subband signals and the inverse correlation matnx are used to update the Kalman gam using equation (4) (box 1220) The Kalman gain is used to update the filter coefficients using equation (6) (box 1230) These filter coefficients are sent to the multi-channel filter (box 1240), applied to the aggregated vector The process continues recursively at every step At every step, the current state of the multi-channel filters is used to reduce the echo in the multi-channel signal The captured signal with reduced echo is output
FIG 13 is a block/flow diagram illustrating another exemplary implementation of the multi-channel round robin regulanzation echo cancellation system 800 and method shown in FIG 8 This exemplary implementation uses the round robin regulanzation technique set forth above and implements each multi-channel adaptive filter of FIG 12 as a plurality of single-channel filters In particular, referring to FIG 13, subband signals from the multiple playback channels are aggregated into a single vector (box 1300) The inverse correlation matnx is updated using equation (7) and the round robin regulanzation (box 1310) The aggregated subband signals and the inverse correlation matrix are used to update the Kalman gam using equation (4) (box 1320) The Kalman gam is used to update the filter coefficients (box 1330) These filter coefficients contain coefficients for both the left and the nght channels The channel weights are de-aggregated (box 1340), where the first-channel filter coefficients are sent to the first-channel single-channel filter (box 1350) and the second-channel filter coefficients are sent to the second-channel single-channel filter (box 1360) The process continues recursively at every step At every step, the current state of the single-channel filters is used to reduce the echo in the multi-channel signal The captured signal with reduced echo is output
Dynamic Switching Between Monaural and Multi-Channel Echo Cancellation The multi-channel round robin regulanzation echo cancellation method and system also includes dynamic switching between echo cancellation algorithms best suited for monaural playback signals and echo cancellation best suited for multi-channel playback signals The dynamic switching occurs between a mono AEC algonthm (such as NLMS) and the multi-channel round robin regulanzation echo cancellation method with efficient reinitialization of either process minimizing the distortion when switching from multi-channel processing to monaural processing and vice versa
In signals having penods of a mono signal (such as talking) and penods of a multi-channel signal (such as music), a dynamic switching algorithm is used to intelligently determine whether the signal is mono (or nearly mono) or a multichannel signal If it is a mono signal, a mono AEC system as shown in FIG 1 with normalized least mean squares (NMLS) algorithm is used If it is a multichannel signal, the multi-channel round robin regulanzation RLS echo cancellation method is used, as shown in FIG 12 The switching is dynamic, such that little or no echo leaks through because the states of the NLMS and RLS algorithms are efficiently reinitialized dunng switching from one algorithm to the next This is achieved by shanng state vanables and using a novel estimation technique to estimate the correlation matnx for the RLS algonthm This allows switching from NMLS (mono) back and forth between the multichannel round robin regulanzation RLS without any glitches
In particular, if the playback signal sent to the speakers is known to only be a purely mono signal, then using normalized least mean squares (NLMS) with a monaural adaptive filter (as shown in FIG 1) tends to perform better than using the RLS with round robin regulanzation (shown in FIG 8) for several reasons For mono signals copied to multiple playback channels, RLS with round robin regulanzation will eventually diverge, the divergence time depending on the forgetting factor X As a result, the AEC algorithm may drift away from the true solution However, NLMS with a single adaptive filter works well for mono signals and is known to be stable In addition, NLMS has significantly lower CPU consumption than any type of multi-channel RLS
Like RLS, NLMS can also be viewed as trying to solve the normal equation, W" = R-'p
For NLMS, the normal equation is solved in a statistical manner using the method of steepest descent, while for RLS, the normal equation is solved
determmistically using least-squares Both methods essentially compute an estimate of R'1 For RLS, the inverse of the correlation matnx is estimated by P For NLMS, R"1 can be viewed as a diagonal matrix with values,
(Equation Removed)
As a result, the state of the inverse correlation matnx estimate can easily be initialized when switching from mono to multi-channel AEC When transitioning from NLMS to RLS, the diagonal elements can be initialized as shown in equation (10) Likewise, when transitioning from the multi-channel round robin regulanzation echo cancellation method and system (shown in FIG 8) to the mono AEC using NLMS (shown in FIG 1), no state has to be initialized explicitly since the state is derived from the data history only
Mixed Round Robin Regulanzation RLS and NLMS Processing
In order to lower central processing unit (CPU) resources needed to run the mufti-channel round robin regulanzation echo cancellation method, a novel mixed RLS/NLMS processing algonthm is used First, the signal is divided into lower subbands (containing the lower frequencies of a multi-channel playback signal) and upper subbands (containing the higher frequencies of the signal) On the lower subbands, the multi-channel round robin regulanzation echo cancellation method using RLS is used On the upper subbands, the NLMS algonthm is used The NMLS algonthm uses less CPU resources, and is used at the upper subbands because the stereo effect is attenuated at these frequencies At lower frequencies, where the stereo effect rs most noticeable, the superior multi-channel round robin regulanzation echo cancellation method is used at the expense of higher CPU processing costs
FIG 14 illustrates the mixed lower/upper subbands processing using the multi-channel round robin regulanzation method described herein Generally, in order to reduce processing requirements, the lower subbands are processed
using the using the multi-channel round robin regulartzation method implemented in RLS and the upper subbands are processed using NLMS A mixed lower/upper subband echo cancellation system 1400 is shown that uses the multi-channel round robin regulanzation echo cancellation method As shown in FIG 14, at least two playback signals x(0) and x(1) are used The playback signal x(1) is played through a right speaker 1410 and x(0) is played through a left speaker 1420
A microphone 1430 captures unwanted echoes from a room (not shown) in which the microphone is located A microphone signal y is processed by a first analysis filterbank 1440 and the playback signals x(0) and x(1) are processed by a second analysis filterbank 1450 and a third analysis filterbank 1460 such that signals x(0), x(1) and y are transformed from the time domain into the frequency domain signals X(0), X(1), and Y, respectively The analysis filterbanks 1440, 1450,1460 can be a windowed fast Fourier transform (FFT), a modulated complex lapped transform (MCLT), or some other frequency domain transform
The transformed playback signals X(0) and X(1) are sent to a mixed processing separation module 1470 The module 1470 uses a separation process to divide the left playback signal X(0) into left lower subband signals 1475 and left upper subband signals 1480 Similarly, the right playback signal X(1) is divided into nght lower subband signals 1478 and nght upper subband signals 1483 Similarly, a mixed processing separation module 1445 divides the captured signal Y into upper subband signals 1448, and lower subband signals 1447 A plurality of NMLS processors 1485 receive the upper subbands of the transformed microphone signal Y, the left upper subband signals 1480 and the nght upper subband signals 1483 for processing Each NLMS processor 1485 processes a signal using the algorithm set forth in equations (1) and (2), where the mono input speaker signal in equation (3) is replaced with the multi-channel speaker signal in equation (8) A plurality of RLS processors 1490 using the multi-channel round robin regulanzation echo cancellation method (an example
of this processor is shown in FIG 13) are used to process the left lower playback subband signals 1475, the right lower playback subband signals 1478, and the lower subband capture signals, 1447 The processed lower subband and upper subband signals are sent to a synthesis filterbank 1495, which outputs a time domain signal z with reduced echo
The foregoing description of the invention has been presented for the purposes of illustration and descnption It is not intended to be exhaustive or to limit the invention to the precise form disclosed Many modifications and variations are possible in light of the above teaching It is intended that the scope of the invention be limited not by this detailed description of the invention, but rather by the claims appended hereto

WHAT IS CLAIMED IS:
1. A method for canceling echo of a plurality of playback signals in a captured signal, comprising:
decomposing the plurality of playback signals into a plurality of subband signals;
applying a plurality of adaptive filters each having an inverse correlation matrix to the plurality of subband signals;
selecting each of the plurality of adaptive filters in a round robin sequence;
regularizing the inverse correlation matrix of each selected adaptive filter; and
reducing the echo in the captured signal using the outputs of the adaptive filters.
2. The method as set forth in claim 1, wherein the plurality of adaptive filters comprises recursive least squares (RLS) adaptive filters.
3. The method as set forth In claim 2, further comprising applying the plurality of RLS adaptive filters in parallel.
4. The method as set forth in claim 2, further comprising having a length of the RLS adaptive filters longer in lower subband signals and shorter in upper subband signals.
5. The method as set forth in claim 1, wherein regularizing further comprises ensuring that the plurality of adaptive filters are stable.
6. The method as set forth in claim 1, wherein regularizing the inverse correlation matrix further comprises processing the inverse correlation matrix such that a determinant of the correlation matrix is not close to zero
7. The method as set forth in claim 1, wherein regularizing the inverse
correlation matrix comprises:
inverting the inverse correlation matrix to produce a correlation matrix;
regularizing the correlation matrix to produce a regularized correlation matrix; and
obtaining a regularized inverse correlation matrix by inverting the
Si
regularized correlation matrix.
8. The method as set forth in claim 7, wherein regularizing the correlation matrix further comprises adding a small positive value to diagonal elements of the correlation matrix.
9. The method as set forth in claim 7, wherein regularizing the correlation matrix further comprises:
defining a threshold value;
comparing each diagonal element of the correlation matrix to the threshold value; and
setting the diagonal element equal to the threshold value if the correlation matrix diagonal element is smaller than the threshold value.
10. The method as set forth in claim 1, wherein applying a plurality of
adaptive filters comprises:
adapting a weight vector for each of the plurality of adaptive filters;
de-aggregating the weight vector to produce a plurality of single-channel filter coefficient vectors such that there is one single-channel filter coefficient vector for each playback channel; and
applying the plurality of single-channel filter coefficient vectors to the plurality of subband signals.
11. The method as set forth in claim 1, wherein the captured signal
comprises an acoustic signal.
12. The method as set forth in claim 1, wherein the captured signal comprises an electrical signal.
13. The method as set forth in claim 1, wherein the plurality of playback signals is a stereo signal.
14. The method as set forth in claim 1, wherein applying a plurality of adaptive filters further comprises identifying a transfer function of a room in which the echo is present.
15. A computer-readable medium having computer-executable instructions for performing the method recited in claim 1.
16. A computer-implemented method for eliminating echoes of a plurality of playback signals in a captured signal, comprising:
decomposing the plurality of playback signals into subbands;
processing the subbands using a recursive least squares (RLS) adaptive filter for each of the subbands;
selecting each of the RLS adaptive filters in round robin scheme;
regularizing the selected RLS adaptive filter to ensure stability; and
reducing echoes in the captured signal using the outputs of the RLS adaptive filters.
17. The computer-implemented method of claim 16, wherein the captured signal comprises an acoustic signal.
18. The computer-implemented method of claim 16, wherein the captured signal comprises an electrical signal.
19. The computer-implemented method of claim 16, wherein
regularizing the selected RLS adaptive filter comprises:
computing a correlation matrix of the RLS adaptive filter; regularizing a correlation matrix of the RLS adaptive filter; and computing a regularized inverse correlation matrix of the RLS adaptive filter from the regularized correlation matrix.
20. The computer-implemented method of claim 19, wherein regularizing the selected RLS adaptive filter further comprises examining the correlation matrix of the RLS adaptive filter to determine whether regularization is necessary. * *
21. The computer-implemented method of claim 19, wherein regularizing a correlation matrix further comprises adding a small positive number to each diagonal element of the correlation matrix.
22. The computer-implemented method of claim 19, wherein regularizing the correlation matrix of the RLS adaptive filter further comprises:
defining a threshold value; and
setting a diagonal element of the correlation matrix equal to the threshold value if the diagonal element is smaller than the threshold value.
23. The computer-implemented method of claim 19, wherein computing a correlation matrix comprises inverting the inverse correlation matrix to obtain a correlation matrix.
24. The computer-implemented method of claim 19, wherein computing a regularized inverse correlation matrix comprises inverting the regularized correlation matrix to obtain a regularized inverse correlation matrix.
25. The computer-implemented method of claim 16, wherein the round robin scheme comprises selecting a single RLS adaptive filter for every frame.
26. The computer-implemented method of claim 16, wherein the round robin scheme comprises selecting a single RLS adaptive filter every several frames.
27. The computer-implemented method of claim 16, wherein the round robin scheme comprises selecting multiple RLS adaptive filters for every frame.
28. A method for automatically switching between two different types of echo cancellation processes when processing one or more playback signals and a captured signal containing echo* comprising:
detecting whether the one or more playback signals is monaural;
processing the captured signal and the one or more playback signals using a monaural echo cancellation process if the one or more playback signals are monaural;
processing the captured signal and the one of more playback signals using a multi-channel round robin regularization echo cancellation process if the one or more playback signals are not monaural; and
dynamically switching between the monaural echo cancellation process and the multi-channel round robin regularization echo cancellation process whenever the detection changes from one type to another.
29. The method as set forth in claim 28, wherein the monaural echo cancellation process is normalized least mean squares (NLMS) echo cancellation.
30. The method as set forth in claim 28, wherein the monaural echo cancellation process is least mean squares (LMS) echo cancellation.
31. The method as set forth in claim 28, wherein the monaural echo cancellation process and the multi-channel round robin regularization echo cancellation process share state variables.
32. The method as set forth in claim 28, further comprising using an estimation technique to estimate a correlation matrix for the multi-channel round robin regularization echo cancellation process.
33. The method as set forth in claim 32, wherein using an estimation technique further comprises solving a normal equation in a statistical manner using a method of steepest descent when dynamically switching from the multichannel round robin regularization echo cancellation process to the monaural echo cancellation process.
34. The method as set forth in claim 32, wherein using an estimation technique further comprises solving a normal equation deterministically using a least squares process when dynamically switching from the monaural echo cancellation process to the multi-channel round robin regularization echo cancellation process.
35. The method as set forth in claim 28, wherein the captured signal contains lower captured subbands and upper captured subbands and further comprising processing the upper captured subband using a normalized least mean squares (NLMS) process.
36. A process for canceling echoes of a plurality of playback signals in a captured signal, the captured signal comprising lower captured subbands and higher captured subbands and the plurality of playback signals comprising lower playback subbands and higher playback subbands, the process comprising:
using a first echo cancellation process to cancel echoes from the upper captured subbands; and
42. The process of claim 38, wherein regularizing the correlation matrix further comprises ensuring that the plurality of adaptive filters remain stable.