Attila Kondacs 17 03 2009

From EECS

March 17: Dr. Attila Kondacs:

Sub-pitch-period spatio-temporal patterns for place of articulation separation in vocalic sounds.

Abstract:

Linear analysis techniques that assume quasi-stationarity of the speech signal do not provide sufficient time-frequency resolution to determine the place of articulation when the articulators move fast. I argue that the traditional low and high-level primitives of speech processing, frequency and phonemes, are inadequate and should be replaced by new representation layers: 1. short pitch period resonances and other repeating spatio-temporal patterns 2. articulator configuration trajectories 3. syllables. I will present progress made in this direction: a class of low-level features as an alternative to frequency-based quantities. I will introduce a mixed frequency- and time-domain feature-extraction technique on vocalic pitch periods that is applicable to one such feature set. I will demonstrate the technique by separating the steady state vowel sets [ae] [ao] and [uw, iy] (low-front, low-back and high vowels). Work in progress.

Short bio:

Attila Kondacs obtained a masters in mathematics from Eotvos University, Budapest, and a diploma in Economics from the University of Cambridge, UK. After completing his doctorate in artificial intelligence at MIT, he spent a year at the University of Chicago as a postdoctoral researcher modeling vocalic speech segments. He currently works for a hedge fund in Chicago.