Keyword extraction based on syllabic information

Abstract

The syllabic structure of a spoken word in continuous speech can be extracted by a pattern of gross phonetic classes. A computer program has been written to transform continuous speech into these segmental patterns. In keyword extraction, portions of the speech utterance that display similar syllabic characteristics to the keyword reference pattern are identified. One approach to keyword extraction is to start with the phonemic transcription of the word and then utilize known phonological rules to accommodate syntactical environments and speaker variations. Another approach which yields comparable results is based on a Markov model incorporating time duration information and segment transitions to determine statistically whether the input segmental pattern is a candidate keyword. This approach which has been tested on continuous speech from a variety of different speakers and different keywords has application in a word recognition or word spotting system.