Chapter 6 - Speech Perception within a Biologically Realistic Information-Theoretic Framework

Abstract

Conceptualizing speech perception as a process by which phonemes are retrieved from acoustic signals is tradition. Within this tradition, research in speech perception has been focused often on problems concerning segmentation and lack of invariance. The problem of segmentation refers to the fact that if phonetic units exist, they are not like typed letters on a page. Instead, they overlap extensively in time much like cursive handwriting. The problem of lack of invariance is related to the segmentation problem. Because speech sounds are produced such that articulations for one consonant or vowel overlaps with the production of preceding ones and vice versa, every consonant and vowel produced in fluent connected speech is dramatically colored by its neighbors. Some of the most recalcitrant problems in the study of speech perception are the consequence of adopting discrete phonetic units as a level of analysis, a level that is not discrete and may not be real. In connected speech, acoustic realization of the beginning and end of one word also overlaps with sounds of preceding and following words; hence the problems of invariance and segmentation are not restricted to phonetic units. Speech perception follows a handful of general principles that are implemented in both sophisticated and not-so-sophisticated ways through the chain of processing from periphery through central nervous system.