Phoneme Lattice Based Confidence Measures in Keyword Spotting

Phoneme lattice based Confidence Measure (CM) is proposed in this paper. It makes use of phoneme lattices generated by a phoneme recognizer. Acoustic Model (AM) based CM is also introduced. For a decoded speech frame aligned to an HMM state, the CM based on AM is calculated. These two confidence measures are combined using a weighting factor to obtain a hybrid CM as they had different dynamic scales. On spontaneous conversational telephone database, the Figure Of Merit (FOM) achieves 17.0% relative improvement comparing to AM based CM.