Learning to predict Pitch Accents using Bayesian Belief Networks for Greek Language


Panagiotis Zervas, Manolis Maragoudakis, Nikos Fakotakis, George Kokkinakis

Wire Communications Laboratory, Department of Electrical and Computer Engineering, University of Patras, 26500 Rion, Patras, Greece




Any text-to-speech (TTS) system that aims at producing understandable and natural-sounding output needs to have a module for predicting prosody. In natural speech, some words are said to be stressed, or to bear Pitch Accents (PA). Errors at this level may impede the listener in the correct understanding of the spoken utterance. Regarding the performance of data driven methods, the scale and quality of the corpus are important. Since there is no suitable corpus available for modeling Modern Greek (MG) prosody, we created a corpus consisted of 5.500 words, distributed in 500 paragraphs. In describing pitch accent in particular and intonation features in general, we use Pierrehumbert's theory adopted for MG. In the present study, we try to predict four categories of PA H*, L*, L+H* and unaccented.


Prosody, Pitch Accent Tones, Speech Synthesis, Bayesian Networks, Naive Bayes, CART

Language(s) Greek
Full Paper