Unsupervised extraction of prosodic structure

Authors: Uwe D. Reichel

Abstract:

Our approach for unsupervised extraction of prosodic structure in spontaneous speech consists of the four steps: chunking into interpausal units, syllable nucleus extraction, prosodic boundary detection, and pitch accent detection. The extraction is based on acoustic features derived from F0 parameterization, and on energy and segment duration features. Phrase boundaries and accents are detected by means of nearest centroid classifiers which are bootstrapped from the data.