Accessing Videos and Implementing Speaker Recognition System Using Speech Processing

Abstract

Modern speechunderstandingsystemsmergeinterdisciplinarytechnologiesfromsignalpro- cessing, patternrecognition,naturallanguage,andlinguisticsintoaunifiedstatisticalframework. These systems,whichhaveapplicationsinawiderangeofsignalprocessingproblems,representa revolutioninDigitalSignalProcessing(DSP).Onceafielddominatedbyvector-orientedproces- sors andlinearalgebra-basedmathematics,thecurrentgenerationofDSP-basedsystemsrelyon sophisticated statisticalmodelsimplementedusingacomplexsoftwareparadigm.Suchsystems are nowcapableofunderstandingcontinuousspeechinputforvocabulariesofseveralthousand wordsinoperationalenvironments.Weexploredthecorecomponentsofmodernstatistically- based speechrecognitionsystems.Theobjectiveofthisprojectistoimplementaspeechrecogni- tion engineanddevelopasystemforspeakerrecognitionusingMelFrequencyCepstrumsand VectorQuantization.ThiswouldinvolvethedesignofanefficientMATLABcodeonaPC. Throughout thedevelopment,measureswillbetakentokeepthememoryrequirementandthe processing timeofthesoftwareassmallaspossible.EverySpeechRecognitionsystemmustbe judged ontwobasicfactorswhichgovernitsusability-accuracyandspeed.Unfortunately,one of themalmostinvariablycomesatthecostoftheother.Ahigheraccuracyrateimpliesawider training sequenceandahighernumberofiterationsinthelearningalgorithm.Ontheotherhand, accuracyremainsanimportantobjectiveofourproject.Theprecisionoftheabovetwomentioned algorithms thathavebeenuseddependalmostentirelyonthemodelparametersforeveryisolated wordwhichneedstobecalculatedattheveryoutset.Toimproveaccuracy,wecalculatethese parameters inaMATLABenvironmentderivingourresultsonalargenumberoftestsequences recorded inatypicalnoisyenvironment.