Acoustic Distance, Acoustic Absement, and the Lexicon

Abstract

It is common in linguistic analysis to treat words as strings of speech segments that are believed to be transduced from the speech signal. However, there are notable shortcomings with this approach, especially concerning word comparison. Principally, comparing speech segment strings does not directly assess the acoustic similarity of words, despite theories and evidence that words that sound similar compete for activation during spoken word recognition. The present dissertation aims to provide a perceptually-grounded method by which words can be represented acoustically and then compared with the dynamic time warping algorithm. The dissertation comprises three studies. The first study is a regression analysis to demonstrate the relationship between acoustic distance and spoken word recognition. It also investigates how to derive more abstract acoustic representations for words based on productions from multiple speakers. The second study investigates what sort of spectral distance function best reflects human perception of acoustic distance. It also examines human perceptual sensitivity to duration differences. The third study compares speech features that are learned by a neural network to mel frequency cepstral coefficients to determine which style of representation for the speech signal better reflects perception. The neural network features are an ensemble of features specific to certain regions of the speech spectrum, while the mel frequency cepstral coefficients are a summary of the entire spectrum. Together, these studies inform the processes of converting the speech signal to an acoustic representation and tuning acoustic comparisons so that they better relate to human cognition.