DIKRAN@SU-CSLI.ARPA (08/02/84)
From: Dikran Karagueuzian <DIKRAN@SU-CSLI.ARPA> [Forwarded from the CSLI Newsletter by Laws@SRI-AI.] LEXICAL ACCESS USING PARTIAL INFORMATION By Daniel P. Huttenlocher, Massachusetts Institute of Technology, Friday, August 3, 2 p.m. in the Trailers' Conference Room next to Ventura Hall. ABSTRACT: Current approaches to speech recognition rely on classical pattern matching techniques which utilize little or no language knowledge. We have recently proposed a model of word recognition which uses speech-specific knowledge to access words on the basis of partial information. These partial descriptions serve to partition a large lexicon into small equivalence classes using sequential phonetic and prosodic constraints. The representation is attractive for speech recognition system because it allows all but a small number of word candidates to be excluded using only a crude description of the acoustic signal. For example, if the word ``splint'' is represented according to the broad phonetic string [fricative][stop][liquid][vowel][nasal][stop], there are only two matching words in the 20,000 word Webster's Pocket Dictionary, ``splint'' and ``sprint.'' Thus, a partial representation can both greatly reduce the space of possible word candidates, and be relatively insensitive to variability in the speech signal across utterance situations. This talk will discuss a set of studies examining the power of such partial lexical representations.