Colin W. Wightman, Nick Campbell
Automatic Labeling of Prosodic Structure
Abstract:This paper describes a series of enhancements to the algorithm recently presented by Wightman and Ostendorf for labeling prosodic
patterns in speech. Through a combination of improved methodology and enhanced feature extraction, significant reduction of the error
rates associated with labeling prominences and major phrase boundaries are achieved. In particular, we show that joint labeling of the
prominence and boundary phenomena can significantly improve the
performance on both types of labels while eliminating inconsistencies
which had been observed between the two label sets. Application
of the improved algorithm is illustrated with experimental results for
joint labeling of prosodic structure in a corpus of professionally read speech.