PowerPoint PresentationSlide 2SVM-based techniques for speaker recognitionSlide 4Slide 5Slide 6Slide 7Applying Recent Advances in Speaker Recognition to the Field of Document ClassificationCS294-5 Class ProjectKofi BoakyeAndrew Hatchspeech recognizerHere are some words and shit, and some phones… iuf aweufb aijf aiuwedf aiu faiuf awiuef awiue faiuw juie ui uji iuf aiuw iuaw iuaw fiuoaw efoi iua iu dfiouas doia doifab sjkbd iuas uijos djhg wei woie oie aiouw efoiwbef oiweb foiwbe oiw efiow fojaw wue jw8ioutput “document”expected word/phoneme countsinput speechdocument classifierSpeaker recognition can be framed as a document classification problemMotivationSVM-based techniques for speaker recognitioniikijikjdpsdpsdpssLLR)()|(log)|(),(●Use relative frequencies of phone/word n-grams as features for each utterance, utti:–Symmetry-based feature transformations:●variance normalization:●rank normalization (Andreas Stolcke):–Kernelized log-likelihood ratio (Campbell et al., NIPS 2003):)()|(,...,)()|()(1100NiNidputtdpdputtdpitorfeatureVec))|(var()|(,...,))|(var()|()(1100jNiNjiuttdputtdputtdputtdpitorfe at ureVec))|(()|(jijiuttdprankuttdp The Task20 Newsgroups:Data set consists of 20,000 articles partitioned nearly evenly across 20 Usenet newsgroupsExamples: alt.atheism, comp.windows.x, talk.politics.miscFor evaluation, data is divided into training (60%) and test (40%)Speaker recognition paradigm adopted for evaluationObtain true/false decisions for test/target pairsEach document tested against all 20 topicsSummary statistic is equal error rate (EER)The ProcessSVM classifier used with unigram statistics as featuresUnigrams selected from a ranked list according to TFIDFscore(w) = p(w) * log(N/d(w)) Experimented with:Vocabulary sizeSmoothed (Good-Turing) counts One-versus-All approach to classifier training takenExperimented with various scaling factors and normalizations for feature vectorsVariance, rank, kernelized LLRResultsSystem EER (%)3K-word 10.346K-word 8.7812K-word 7.8325K-word 7.8311K-word w/ stems 8.122K-word w/ stems 7.54System EER(%)Baseline 7.54Variance norm. Running…Rank norm. 10Kernelized LLR 6.58Note: Present results only from alt.atheismSmoothing yielded no performance improvementVocabulary expansion made asymptotic gainsInclusion of word stems permitted gains beyond apparent asymptoteFuture Work• Expand analysis to documents from all topics• More sophisticated stemming• Reworking of smoothing • Multiclass
View Full Document