Combination of Word and Phoneme Approach for Spoken Term Detection
Burget Lukáš, doc. Ing., Ph.D. (DCGM FIT BUT)
Karafiát Martin, Ing., Ph.D. (DCGM FIT BUT)
spoekn term detection
The poster deals with Combination of Word and Phoneme Approach for Spoken Term Detection
The goal is to provide a technique for search for an out-of-vocabulary (OOV) term or keyword in a speech. The spoken term detection is based on lattices generated from word and phoneme recognizers. The term is a sequence of words (quoted query). The results are evaluated on the Broadcast news subset of NIST STD 2006 development set. Our goal is to propose an upper-bound word/phoneme combination technique for Spoken Term Detection (STD). We want to estimate the upper bound results for OOV terms, which can be achieved using rescoring by strong (and slow) LVCSR system. Then we can decide how much worse is an simple (and fast) sub-word method. We also evaluated different approaches for term confidence estimation and two differen approaches for phoneme lattice generation.