Comparison of Keyword Spotting Approaches for Informal Continuous Speech

This paper describes several approaches to keyword spotting (KWS) forinformal continuous speech. We compare acoustic keyword spotting,spotting in word lattices generated by large vocabulary continuousspeech recognition and a hybrid approach making use of phoneme latticesgenerated by a phoneme recognizer. The systems are compared oncarefully defined test data extracted from ICSI meeting database. Theadvantages and drawbacks of different approaches are discussed. Theacoustic and phoneme-lattice based KWS are based on a phonemerecognizer making use of temporal-pattern (TRAP) feature extraction andposterior estimation using neural nets. We show its superiority overtraditional HMM/GMM systems. A posterior probability transformationfunction is introduced for posterior based acoustic keyword spotting.We also propose a posterior masking algorithm to speed-up acoustickeyword spotting.