PIMS-CSC Seminar: Measuring the phonetic similarity between speech sounds

This talk demonstrates a new metric for measuring phonetic similarity between speech sounds, based on several types of phonetic data collected for a set of 130 consonants and vowels. Phonetic similarity is frequently invoked for explaining a wide range of phonological observations, and the purpose of this project is to provide a resource for quantifying phonetic similarity, distinguishing different types of phonetic similarity (e.g. articulatory, acoustic, and perceptual) and for distinguishing phonetic similarity from phonological notions of similarity, such as those based on features (e.g. Frisch, Broe, and Pierrehumbert 1997, Kondrak 2003).

Subjects were recorded by the following methods: digital audio, video of the face, mid-sagittal lingual ultrasound video, oral and nasal airflow, and electroglottography. Acoustic similarity was calculated using a dynamic time warping algorithm, and vocal tract data were generated by tracing tongue, palate, and lip surfaces in each of the video frames within each target segment. Each type of phonetic data was used to generate a distance matrix. Acoustic and vocal tract-based distances were subjected to a principal component analysis in order to generate independent acoustic and articulatory dimensions. The resulting data are used to measure the similarity of the speech sounds in various dimensions, as part of a similarity calculator application, which calculates distances and generates maps based on user input.