Sparse Adaptive Filters for Echo Cancellation Synthesis Lectures on Speech and Audio Processing Editor B.H.Juang,GeorgiaTech SparseAdaptiveFiltersforEchoCancellation ConstantinPaleologu,JacobBenesty,andSilviuCiochina˘ 2010 Multi-PitchEstimation MadsGræsbøllChristensenandAndreasJakobsson 2009 DiscriminativeLearningforSpeechRecognition:TheoryandPractice XiaodongHeandLiDeng 2008 LatentSemanticMapping:Principles&Applications JeromeR.Bellegarda 2007 DynamicSpeechModels:Theory,Algorithms,andApplications LiDeng 2006 ArticulationandIntelligibility JontB.Allen 2005 Copyright© 2010byMorgan&Claypool Allrightsreserved.Nopartofthispublicationmaybereproduced,storedinaretrievalsystem,ortransmittedin anyformorbyanymeans—electronic,mechanical,photocopy,recording,oranyotherexceptforbriefquotationsin printedreviews,withoutthepriorpermissionofthepublisher. SparseAdaptiveFiltersforEchoCancellation ConstantinPaleologu,JacobBenesty,andSilviuCiochina˘ www.morganclaypool.com ISBN:9781598293067 paperback ISBN:9781598293074 ebook DOI10.2200/S00289ED1V01Y201006SAP006 APublicationintheMorgan&ClaypoolPublishersseries SYNTHESISLECTURESONSPEECHANDAUDIOPROCESSING Lecture#6 SeriesEditor:B.H.Juang,GeorgiaTech SeriesISSN SynthesisLecturesonSpeechandAudioProcessing Print1932-121X Electronic1932-1678 Sparse Adaptive Filters for Echo Cancellation Constantin Paleologu UniversityPolitehnicaofBucharest,Bucharest,Romania Jacob Benesty INRS-EMT,UniversityofQuebec,Montreal,Canada Silviu Ciochina˘ UniversityPolitehnicaofBucharest,Bucharest,Romania SYNTHESISLECTURESONSPEECHANDAUDIOPROCESSING#6 M &C Morgan &cLaypool publishers ABSTRACT Adaptivefilterswithalargenumberofcoefficientsareusuallyinvolvedinbothnetworkandacoustic echocancellation.Consequently,itisimportanttoimprovetheconvergencerateandtrackingofthe conventionalalgorithmsusedfortheseapplications.Thiscanbeachievedbyexploitingthesparse- nesscharacteroftheechopaths.Identificationofsparseimpulseresponseswasaddressedmainlyin thelastdecadewiththedevelopmentoftheso-called“proportionate”-typealgorithms.Thegoalof this book is to present the most important sparse adaptive filters developed for echo cancellation. Besides a comprehensive review of the basic proportionate-type algorithms,we also present some of the latest developments in the field and propose some new solutions for further performance improvement,e.g.,variablestep-sizeversionsandnovelproportionate-typeaffineprojectionalgo- rithms.Anexperimentalstudyisalsoprovidedinordertocomparemanysparseadaptivefiltersin differentechocancellationscenarios. KEYWORDS network and acoustic echo cancellation, adaptive filters, sparseness, Wiener, LMS, ± NLMS,VSS-NLMS,PNLMS,IPNLMS,EG ,VSS-PNLMS,APA,PAPA vii Contents 1 Introduction.................................................................1 1.1 EchoCancellation..........................................................1 1.2 Double-TalkDetection......................................................3 1.3 SparseAdaptiveFilters......................................................4 1.4 Notation...................................................................5 2 SparsenessMeasures ......................................................... 7 2.1 VectorNorms..............................................................7 2.2 ExamplesofImpulseResponses..............................................9 2.3 SparsenessMeasureBasedonthe(cid:2)0 Norm....................................9 2.4 SparsenessMeasureBasedonthe(cid:2)1 and(cid:2)2 Norms...........................10 2.5 SparsenessMeasureBasedonthe(cid:2)1 and(cid:2)∞ Norms..........................11 2.6 SparsenessMeasureBasedonthe(cid:2)2 and(cid:2)∞ Norms..........................12 3 PerformanceMeasures ...................................................... 15 3.1 Mean-SquareError........................................................15 3.2 Echo-ReturnLossEnhancement ...........................................16 3.3 Misalignment.............................................................17 4 WienerandBasicAdaptiveFilters............................................19 4.1 WienerFilter .............................................................19 4.1.1 EfficientComputationoftheWiener-HopfEquations..................22 4.2 DeterministicAlgorithm...................................................24 4.3 StochasticAlgorithm......................................................28 4.4 VariableStep-SizeNLMSAlgorithm .......................................31 4.4.1 ConvergenceoftheMisalignment ....................................33 4.5 SignAlgorithms...........................................................34 viii 5 BasicProportionate-TypeNLMSAdaptiveFilters ............................37 5.1 GeneralDerivation........................................................37 5.2 TheProportionateNLMS(PNLMS)andPNLMS++Algorithms .............39 5.3 TheSignedRegressorPNLMSAlgorithm...................................40 5.4 TheImprovedPNLMS(IPNLMS)Algorithms..............................41 5.4.1 TheRegularIPNLMS...............................................42 5.4.2 TheIPNLMSwiththe(cid:2)0 Norm .....................................44 5.4.3 TheIPNLMSwithaNorm-LikeDiversityMeasure....................45 6 TheExponentiatedGradientAlgorithms.....................................47 6.1 CostFunction.............................................................47 6.2 TheEGAlgorithmforPositiveWeights.....................................48 ± 6.3 TheEG AlgorithmforPositiveandNegativeWeights ......................49 ± 6.4 LinkBetweenNLMSandEG Algorithms.................................51 ± 6.5 LinkBetweenIPNLMSandEG Algorithms...............................53 7 TheMu-LawPNLMSandOtherPNLMS-TypeAlgorithms ..................55 7.1 TheMu-LawPNLMSAlgorithms .........................................55 7.2 TheSparseness-ControlledPNLMSAlgorithms.............................59 7.3 ThePNLMSAlgorithmwithIndividualActivationFactors...................60 8 VariableStep-SizePNLMSAlgorithms ...................................... 65 8.1 ConsiderationsontheConvergenceoftheNLMSAlgorithm..................65 8.2 AVariableStep-SizePNLMSAlgorithm....................................70 9 ProportionateAffineProjectionAlgorithms...................................73 9.1 ClassicalDerivation .......................................................73 9.2 ANovelDerivation........................................................75 9.3 AVariableStep-SizeVersion...............................................79 10 ExperimentalStudy .........................................................87 10.1 ExperimentalConditions...................................................87 CONTENTS ix 10.2 IPNLMSVersusPNLMS..................................................88 10.3 MPNLMS,SC-PNLMS,andIAF-PNLMS.................................92 10.4 VSS-IPNLMS............................................................95 10.5 PAPAs...................................................................96 Bibliography...............................................................103 Index......................................................................111 Authors’Biographies.......................................................113