A Syllable-Based Prominence Detection Model Based on Discriminant Analysis and Context-Dependency

Abstract

On the basis of our previous work, we propose a syllable-based prominence detection model within the framework of ex-ploratory data analysis and discriminant learning in the acoustic domain. This paper investigates two hypothesis on the acoustic data processing: a linear discriminant analysis in which the rel-ative discriminant ability of single prosodic cues are combined into prosodic patterns and a context-dependant model that ac-counts for phonological dependencies (phonetic intrinsic prop-erties and coarticulation effect). The proposed approach signif-icantly outperforms a baseline method on a corpus of French read speech with a performance of 87.5 % in f-measure for the prominent syllables (respectively 90.4 % in global accuracy). 1.