1 / 7

Applying Recent Advances in Speaker Recognition to the Field of Document Classification

Applying Recent Advances in Speaker Recognition to the Field of Document Classification. CS294-5 Class Project Kofi Boakye Andrew Hatch. Motivation. Speaker recognition can be framed as a document classification problem. output “document”. expected word/phoneme counts.

yaron
Download Presentation

Applying Recent Advances in Speaker Recognition to the Field of Document Classification

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Applying Recent Advances in Speaker Recognition to the Field of Document Classification CS294-5 Class Project Kofi Boakye Andrew Hatch

  2. Motivation Speaker recognition can be framed as a document classification problem output “document” expected word/phoneme counts Here 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 jw8i speech recognizer input speech document classifier

  3. 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): SVM-based techniques for speaker recognition

  4. The Task 20 Newsgroups: Data set consists of 20,000 articles partitioned nearly evenly across 20 Usenet newsgroups Examples: alt.atheism, comp.windows.x, talk.politics.misc For evaluation, data is divided into training (60%) and test (40%) Speaker recognition paradigm adopted for evaluation Obtain true/false decisions for test/target pairs Each document tested against all 20 topics Summary statistic is equal error rate (EER)

  5. The Process SVM classifier used with unigram statistics as features Unigrams selected from a ranked list according to TFIDF score(w) = p(w) * log(N/d(w)) Experimented with: Vocabulary size Smoothed (Good-Turing) counts One-versus-All approach to classifier training taken Experimented with various scaling factors and normalizations for feature vectors Variance, rank, kernelized LLR

  6. Results Note: Present results only from alt.atheism Smoothing yielded no performance improvement Vocabulary expansion made asymptotic gains Inclusion of word stems permitted gains beyond apparent asymptote

  7. Future Work • Expand analysis to documents from all topics • More sophisticated stemming • Reworking of smoothing • Multiclass SVMs

More Related