Towards automatic language processing and intonational labeling in European Portuguese

Abstract

This work describes a framework that encompasses multi-layered linguistic information, focusing on prosodic features (pitch, energy, and tempo patterns), uses such features to distinguish between sentence-form types and disfluency/fluency repairs, and contributes to the characterization of intonational patterns of spontaneous and prepared speech in European Portuguese. Different machine learning methods have been applied for discriminating between structural metadata events, both in university lectures and in map-task dialogues, containing large amounts of spontaneous speech. Results show that prosodic features, and particularly a set of very informative features, are crucial to distinguish between sentence-form types and disfluency/fluency repair events. This is the first work for European Portuguese on both fully automatic processing of multilayered linguistically description of spoken corpora and intonational labeling.