Enriching Speech Recognition by Recovering Punctuation and Performing Capitalization

From HLT@INESC-ID

Fernando Batista

Date

  • 14:30, March 09, 2007
  • 3rd floor meeting room

Speaker

Abstract

This presentation describes my work on inserting punctuation marks and capitalizing the output of an Automatic Speech Recognition System (ASR). The output of an ASR often consists on a raw text, usually in a lower-text format, without any punctuation marks. This work aims to provide more usable transcriptions both for humans and machines.

Different experiments were performed: using transducers; the SRILM toolkit; and maximum entropy models. The presentation will describe the advantages and major difficulties on applying each one of the methodologies. Results of experiments conducted both over written newspaper corpora and the speech output will be presented.

As this work is not concluded yet, I will present the future work on this matter.