Automatic Labeling of Intonation Using Acoustic and Lexical Features

This paper proposes a framework of automatic intonation labeling which involves detection and classification of pitch accents and phrase boundaries. Four statistical models are designed to perform these tasks on the basis of a compact and simple representation consisting of features identified as the main acoustic correlates of accentual prominence and phrase boundaries or describing the acoustic-phonetic realization of different types of pitch accents and boundary tones. The features can be easily derived from utterance's acoustics (F0 and timing cues) and lexical features. The models yield high average accuracy between 80% and 87% depending on the task, and high consistency they approach the levels of agreement among human labelers in manual transcription of intonation.