Prosodic Cues for Automatic Word Boundary Detection in ASR

This article presents a cross-lingual study for agglutinative, fixed stressed languages, like Hungarian and Finnish, about the segmentation of continuous speech on word level by examination of supra-segmental parameters. We have developed different algorithms based either on a rule based or a datadriven approach. The best results were obtained by data-driven algorithms (HMMbased methods) using the time series of fundamental frequency and energy together. This HMM based method will be described in this article. Word boundaries were marked with acceptable accuracy, even if we were unable to find all of them. On the base of this study a word level segmentationer has been developed which can indicate the word boundaries with acceptable precision for both languages. The evaluated method is easily adaptable to other fixed-stress languages.