Integrating Statistical and Knowledge - based Methods for Automatic Phonemic Segmentation

This thesis presents a prototype system, which integrates statistical and knowledgebased methods, for automatic phonemic segmentation of speech utterances for use in speech production research. First, Aligner, a commercial speech alignment software, synchronizes the speech waveform to the provided text, using hidden Markov models that were trained on phones. Then, a custom built knowledge-based segmentation program is used to locate and label the segmentation boundaries. This thesis provides one example of automatic segmentation that specifically identifies the boundaries associated with voiceless stop consonants. However, the structure of the segmentation system that is discussed in this thesis provides a framework for further development of the automatic segmentation system. The results of this prototype system are highly successful. Thesis Supervisor: Joseph Perkell Title: Senior Research Scientist