On the Contribution of Instance-Specific Characteristics to Speech Perception

Author: Ann R. Bradlow, Lynne C. Nygaard And David B. Pisoni

Abstract:
Utterance-, speaker-, and listener-related correlates of speech intelligibility were investigated using data from the Indiana Multi-Talker Sentence Database and from a talker identification study. The sentence database consists of 100 Harvard sentences produced by 20 speakers, as well as intelligibility data in the investigate some of the sentence- and talker-specific correlates of speech intelligibility. Results showed that talker-related characteristics such as gender and individual differences in phonetic implementation were correlated with the observed variability in talker intelligibility. Specifically, the data showed that female talkers, who tend to exhibit fewer instances of phonological reduction phenomena, were generally more intelligible than male talkers, who may be less precise in their phonetic implementation of phonological forms. An investigation of the sentence-related factors that correlate with variability in overall sentence intelligibility revealed that the number and type of words that comprise a particular sentence were important factors controlling intelligibility. Sentences with relatively high overall intelligibility were generally shorter (had fewer words), and had more easy-to-recognize words (i.e. shorter, more frequent, and more phonetically distinctive words) than sentences with relatively low overall intelligibility. The talker identification study trained listeners to identify 10 talker by name over a period of 9 days, after which the listeners performed a word identification task with novel words spoken by the now familiar voices as well as by novel voices. These data provided us with the means to investigate the effect of talker familiarity on speech perception, and the relationship between talker identifiability and talker intelligibility. Results of this investigation showed that listeners who learned to identify the voices showed an advantage in the word identification task with words spoken by familiar talkers relative to their performance in the task with unfamiliar talkers. These data also showed that talkers who were easily identified within the group of 10 talkers were not the most intelligible talkers as measured by the word identification task, implying that talker distinctiveness and talker intelligibility are not necessarily related.