Seminar: 2/9 - Keith Johnson
6:00 to 6:00 AM at:
*** THIS TALK IS POSTPONED UNTIL FALL 2007 SEMESTER. ***
In this paper I examine the assumption that people recognize words by first recognizing the alphabetic segments that make up the word. Behavioral studies in the psycholinguistic literature have been inconclusive on this point, so this paper takes a different approach to this question - we will examine pronunciation variation in conversational speech. Data from a corpus of thirteen hours of natural conversational speech finds that over 10% of all word tokens in normal speech have at least one segment deletion. This is not a fact that can be easily accommodated in current word recognition theory and this figure even obscures variation somewhat because some words have a dizzying number of variants. For example, "Ohio" was pronounced in 15 different ways and "yknow" had over 100 pronunciation variants.
Interestingly, the number of variants a word has is a function of the word's frequency of occurrence - suggesting that the most frequently produced words in conversation would be the most difficult to recognize in a segment-by-segment strategy. The paper concludes with a discussion of two phenomena which suggest that words are phonetic gestalts - words have phonetic "islands of reliability" which do not vary, and homophones are usually not homophonous in conversational speech.