Event

Prosodic Pitch Features for Sentence Segmentation

James Fung

ICSI

Tuesday, August 18, 2009
12:30

In this talk, we explore the potential of using statistical learning algorithms to aid in the design of prosodic features. In particular, we examine the application of classifiers such as linear discriminant analysis (LDA) and multi-layer perceptron (MLP) to generate pitch features for sentence segmentation in English. We show that the resulting features are comparable to an established baseline set of pitch features and explore some of the issues concerning the system, including training and robustness. We also present experiments using lexical tone modeling for Mandarin pitch features. We fit the pitch contour using a Legendre polynomial basis and examine their behavior to create a tritone pitch contour model. This lexical contour is then removed so prosodic features can make use of higher level pitch structure.