Text-based and Signal-based Prediction of Break Indices and Pause Durations

Abstract

The relation between symbolic and signal features of prosodic boundaries is experimentally studied using prediction methods. Text-based break index prediction turns out to be fairly good, but signal-based prediction and pause duration prediction perform worse. A possible reason is that random signal feature variations, as usually produced by humans, are hard to predict. Speakers divide their utterances into prosodic phrases separated by weak or strong prosodic boundaries. These boundaries can be perceptually classied according to the ToBI break indices, or they can be described by measurements of prosodic features of the speech signal, i.e. the pause duration and the local contours of F0, speech rate, amplitude, and voice quality in the vicinity of the prosodic boundary. The goal of the present study is to investigate the relationship between symbolic and signal representations of prosodic boundaries by means of three prediction methods (see Fig. 1). Prosodic phrase break prediction can be decomposed into two tasks: break localization and prediction of its phonetic realization. For the latter this study concentrates on pause duration.