High fidelity analysis of vowel acoustic space

Vowel acoustic space is often characterized by polygons, whose vertices are determined by summary statistics such as mean values of the formant frequencies of distinct phonemes. The F1-F2 quadrilateral is the most familiar of these. However, using summary statistics to represent formant-frequency data presents fundamental limitations. These data are inherently lossy—summarizing large amounts of data with single values; mean itself is a non-robust statistic, highly sensitive to outliers; and even robust statistics ignore distributional information within the data, which can vary markedly among different phonemes and age groups. We introduce a new approach characterizing and measuring change in formant spaces statically and developmentally. This approach treats acoustic spaces as point clouds of data, in which no information is abstracted or lost. Within this framework, we measure the spatial overlap of sets of formant data using an approach combining optimization theory and computational statistics. This provides highly sensitive measures of both similarity and extent of temporal change. This novel approach is robust with respect to outliers, noise, and missing values. It also has a strong intuitive and rigorous mathematical foundation and is easily visualized. Finally, it enables detailed examination of individual phonemes and clustering of speakers identifying shared developmental patterns. [Work was supported by NIH grants # R01-DC 006282 & P30-HD03352.]