The acoustics of place of articulation in English plosives

Abstract

This thesis investigates certain aspects of the acoustics of plosives’ place of articulation that have not been addressed by most previous studies, namely: 1. To test the performance of a technique for collapsing F2onset and F2mid into a single attribute, termed F2R. Results: F2R distinguishes place with effectively the same accuracy as F2onset+F2mid, being within ±1 percentage point of F2onset+F2mid at its strongest over most of the conditions examined. 2. To compare the strength of burst-based attributes at distinguishing place of articulation with and without normalization by individual speaker. Results: Lobanov normalization on average boosted the classification of individual attributes by 1.4 percentage points, but this modest improvement shrank or disappeared when the normalized attributes were combined into a single classification. 3. To examine the effect of different spectral representations (Hz-dB, Bark-phon, and Bark-sone) on the accuracy of the burst attributes. The results are mixed but mostly suggest that the choice between these representations is not a major factor in the classification accuracy of the attributes (mean difference of 1 to 1.5 percentage points); the choice of frequency region in the burst (mid versus high) is a far more important factor (13 percentage-point difference in mean classification accuracy). 4. To compare the performance of some traditional-phonetic burst attributes with the first 12 coefficients of the discrete cosine transform (DCT). The motivation for this comparison is that phonetic science has a long tradition of developing burst attributes that are tailored to the specific task of extracting place-of-articulation information from the burst, whereas automatic speech recognition (ASR) has long used attributes that are theoretically expected to capture more of the variance in the burst. Results: the DCT coefficients yielded a higher burst classification accuracy than the traditional phonetic attributes, by 3 percentage points.