Keyword Spotting in Meeting Data

This paper describes several approaches to  keyword spotting (KWS)  for informal continuous speech. We compare  acoustic keyword spotting, spotting in word lattices generated by large vocabulary continuous speech recognition and a hybrid approach making use of phoneme lattices generated by a phoneme recognizer. The  systems are compared on carefully defined test data extracted from ICSI meeting database. The  acoustic and phoneme-lattice based KWS are based on a phoneme recognizer making use of long temporal split context feature extraction
and posterior estimation using neural nets. The advantages and drawbacks of different approaches are discussed.