Priberam Machine Learning Lunch Seminar: An Iterative, Constrained Approach for Pitch Component Extraction

Speaker: Gopala Anumanchipalli (INESC-ID, LTI/CMU)
Venue: IST Alameda, Sala PA2 (Edifício de Pós-Graduação)
Date: Tuesday, November 30th, 2010
Time: 13:00
Lunch will be provided

Abstract:
In this talk I will describe an approach for automatic extraction of global and local patterns of pitch(F0) contours taking into account the overall trends of these phenomena in the presented data. We propose an iterative algorithm to optimally extract these components to minimize the reconstruction error of the F0 contour. Furthermore, we present a constraint specification strategy to incorporate known constraints on these phenomena to converge on better realizations of the components (like the Phrase and Accent commands of the physiologically motivated Fujisaki Model of F0). The extracted components are shown to be correlated to established theoritical notions of declination, metrical feet and accent tones.

Bio:
Gopala is a Ph.D. student in the LTI, Carnegie Mellon University and INESC-ID Lisboa, IST. He is advised by Dr. Alan W Black and Dr. Luis Oliveira. He is currently at INESC-ID. He is interested broadly in everything to do with language, but specifically works on building models and transformation approaches for prosody in Speech synthesis. He is working in the PT-Star project aiming to do Speech-to-Speech machine translation of video lectures.