基于矢量量化(vq)的说话人识别的研究.pdf

Abstract 硕士学位论文 Abstract Speal(errecognition tecllIlology is animportant research branch inspeech recognition isbased on thefeatureparameters,which are extracted toea’ectively renect thepersonality recognition process includes pre-processing of thespeech signal,feature extraction,modeling呻d model p印er made the following research: In tenlls ofspeech enhancement,due totheinfluence ofnoise on theperfbnnance of speaker recognition system,FaStICA based on thenegatiVe entropy was this paper,FastICA bined with a11dwas used speech show thatthee肫ct ofspeech e1111aJlcement isobvious. Inendpoint detection stage,the paper studied thetraditional endpoint detection、矿hich isbased on dualthreshold and cepstmm (}谢ng ideafrom t11ismethod,aJl inlproVed cepstnⅡn distance endpoint detection algorithm was experiments show thattlleimproved method can achieve betterresults. In stage,the cepstral fIeanJreand pitch of speech signal bined asfeatureparameters inspeaker recognition ,if mese feature p猢eters were superimposed directly,me锄ount ofcalculation、析ll increases,thereby both and therecognition timewillalso Fisher critedon、vas used to select feature dimensions. First,the Fisher criterion ratio to each dimension offeatureparameters was ,seveml dimensions ofeachfeature, which correspond toseVeral leading biggest Fisher criterion ratios,was bined features,one group which can achieve thebestrecognitionpe怕mlaJlce was results show bined features,wIlich were selectedby Fishercriterion,caJl remoVe redundancy a11dmnher improve thepe—’omance ofspeal(er system. The vector Quantization Model wausintroduced LBG algoritllm would besensitiVe tooutliers,impulse noise and salt&pepper noise iIuringVQ codebook gener