Tid: Onsdag 27 maj 2015, kl. 15:00–17:00
Plats: C307, Södra huset, Frescati

Postseminarium följer direkt efter seminariet i institutionens pentry.

Abstract
For those interested in understanding the motion of speech articulators, this is an exciting era. Real-time MRI (rtMRI) allows us to observe relatively transitory configurations of the vocal tract multiple times, among multiple speakers, and from multiple angles. Each pixel in a resulting rtMR image provides us with a large number of dimensions to use in describing the postures and movements of speech. With data of such high dimensionality, some form of data reduction is necessary in order to answer even relatively simple questions of phonetic or phonological interest. Instead of referring to average, time-varying pixel intensity in a region of interest (ROI) or by detecting boundaries and structures in the vocal tract, our group has found success with techniques that leverage the intensity of all pixels in an ROI. N pixels in a ROI, e.g., placed at the velopharyngeal port in an oblique section, are fed into a principal component (PC) analysis with i observations of a variety of speech sounds. The number of PCs is optimized and the resulting data are fed into a linear discriminant analysis using k-means. The resulting model can be validated against supplementary observations and used in cluster analysis. In this presentation, I will discuss results from studies of coronal consonants in Beijing Mandarin and oral/nasal vowels in Northern Metropolitan French. Using large numbers of PCs presents us with unique problems, particularly in the interpretation of multiple abstract results, which may vary across speakers. While these and related techniques engender a number of technical and rhetorical challenges, their application to large rtMRI corpora appears to have considerable potential for the study of articulatory phonetics and phonology.

Ryan K. Shosted at University of Illinois at Urbana-Champaign

Hjärtligt välkomna
Hatice Zora