An event-driven f(o) weighting for prosody control in a large corpus-based TTS system

Abstract

This letter presents event-driven weighting methods concentrating on f(0) for prosody control in a large corpus-based text-to-speech (TTS) system. We determine the f(0) weighting factor for a given target using its linguistic features by automatically using classification and regression trees (CART). The target predicted as perceptually important is weighted more than others. This results in more natural synthetic speech from a prosodic viewpoint.