Automatic Detection of Prosodic Stress in American English Discourse

Due to the incompletely understood nature of prosodic stress, the implementation of an automatic transcriber is very diicult on the basis of the currently available knowledge. For this reason, a number of data driven approaches are applied to a manually annotated set of les from the OGI English Stories Corpus. The goal of this analysis is twofold. First, it aims to implement an automatic detector of prosodic stress with suuciently reliable performance. Second, the eeectiveness of the acoustic features most commonly proposed in the literature is assessed. That is, the role played by duration, amplitude and fundamental frequency of syllabic nuclei is investigated. Several data-driven algorithms, such as Artiicial Neural Networks (ANN), statistical decision trees and fuzzy classiication techniques, and a knowledge-based heuristic algorithm are implemented for the automatic transcription of prosodic stress. As reference , two diierent subsets from the OGI English stories database were hand labeled in terms of prosodic stress by two individuals trained in linguistics. The agreement between the two transcribers on a set of common les is only slightly higher than that obtained by the automatic systems. While the ANN based approach achieves the highest performance (77% primarily stressed vocalic nuclei vs. 79% unstressed vocalic nuclei in average for the two transcribers data sets), the other methods show that both transcribers grant a major role to duration and (to a slightly lesser degree) to amplitude. Pitch relevant features of the syllabic nuclei appear to play a much less important role than amplitude and duration.