Doctoral Dissertation Research: Lexical Acoustics

点击翻译按钮获取中文摘要

This project will document which acoustic properties of the speech signal English listeners use to communicate by looking at patterns in both their correct recognition of words, and their confusions between words, when listening in a noisy environment. The project will enhance understanding of speech production and perception, particularly regarding how acoustic features are distributed throughout the words of a language to minimize misperceptions and therefore facilitate reliable communication. More broadly, this work is relevant for research on the relation between speech sounds and word knowledge in both adults and children, contributing to the identification of unique distinctions between languages that affect second language learners, and ultimately informing models focusing on various types of hearing impairment in everyday communication.Two general models of the structure of the sound system of a language are contrasted. The first model, referred to as the "inventory model", is the canonical approach that forms the basis for most phonetic descriptions of sound systems in speech perception. Under the inventory model, the sound system is structured as a fixed set of contrastive consonants and vowels, where this set is treated as homogeneous in weight such that all elements play an equal role in defining the system and determining the relative importance of individual acoustic cues. The second model, referred to as the lexicon model, considers the system of contrastive sounds to be critically distributed over the lexicon, such that the relative weight of different cues now depends on the number and configuration of words those cues serve to distinguish. In contrasting these two approaches, two main questions are asked: (1) how well do the models agree in their estimates of cue weights, and what accounts for points of agreement and disagreement; and (2) what does each model entail for the stability of the system in the presence of background noise in the environment. The inventory model uses a database of controlled productions of syllables to predict prior-published confusion patterns between sounds, with cue weights then measured as the relative importance of each cue in the model predicting listener perception patterns. For the lexicon model, a database of nearly 27,000 words produced by a single speaker is used to predict listener word recognition patterns when these words are embedded in noise in a series of six experiments utilizing different tasks (open, fill-in-the-blank recognition, and closed choice between two similar-sounding words) and different stimuli (naturally produced, enhanced, and degraded). The cue weights from these two model fits are then compared to address the first question. To answer the second question, simulations of noise added to different cues, and the resulting change in predicted similarity between sounds (in the inventory case) and words (in the lexicon case), are used to study the stability and adaptive response of each system to uncertainty in the acoustic signal. As such, these experiments will be able to address key claims about the structure and organization of sound systems in language.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.