A Syntactic Pattern Recognition Method for the Automatic Location of Potential Enhancement Regions in Running Speech

A method for the automatic location of phonetically important events in continuous speech signals is presented. The events chosen are plosive bursts, fricatives, nasals and vocalic onsets and offsets. These events have been chosen to match the requirements of a system that enhances these regions to make the signal more robust to channel degradation. The technique uses a syntactic pattern recognition approach distinguished from automatic speech recognition or from knowledge-based approaches. Performance on read speech from many talkers was approximately 20% misses and 20% false alarms. This performance seems broadly similar to other published techniques, although direct comparisons on equivalent tasks have yet to be undertaken.