Evaluation of Energy and Duration on Malay Phrase Breaks

This paper presents evaluation of energy and duration features in detection of phrase breaks. The training feature set is developed from evaluation of targeted phrase break in listening test. This cheaper and faster approach is proven useful for under-resource language such as Malay which do not have comprehensive prosodic-annotated corpus. In addition, instead of labeling the phrase break using linguistic and phonetic meaning, the listening test allow labeling of phrase break as perceived by listener. Then, word-related energy and duration features are extracted from the targeted phrase break. Evaluation of the features with RBF, MLP and logistics models reveal best detection accuracy of 84.6% which is comparable to existing context-based algorithm. This simpler approach of using energy patterns and duration features for detection of phrase break, can be used to segment lengthy spontaneous speech into smaller meaningful utterance without analysis of linguistic information. In addition, the results can be use as preliminary information for evaluation of boundary salience at the targeted boundary locations.