Agnieszka Wagner, Adam Mickiewicz University
This paper proposes a framework of automatic determination of phrasing using acoustic features derived from the speech signal. The feature vectors were defined in a series of analyses investigating the acoustic-phonetic realization of minor and major phrase boundaries and different boundary types. The resulting representation was used to train statistical classifiers to automatically determine phrase boundary position and type. The output of the classifiers can be used to provide speech corpora with phrasing information to enhance the performance of TTS or ASR systems, or to generate a comprehensive feedback in prosody tutoring systems. Apart from providing an efficient means for automatic phrase boundary detection, the study presented in this paper sheds also light on the role of timing and F0 cues in signaling phrase boundaries.