New efficient fillers for unlimited word recognition and keyword spotting

Describes our complete results for improved lexical fillers as well as two new kinds of fillers, gives their results in unlimited speech recognition as well as for keyword spotting and compares them to the acoustic-phonetic filler in the case of keyword spotting. Tests have been conducted on different vocabularies derived from ATIS (Air Travel Information System) and the Wall Street Journal database. Results for keyword spotting show the superiority of the independent lexical phonemic filler that combines accuracy (92% for a false alarm rate of 1.2 FA/h/kw) as well as task-independent training. As for new-word detection, the syllabic and the independent lexical fillers perform quite well, and allow relevant detection of the phonetic transcription.