Combination of strongly and weakly constrained recognizers for reliable detection of OOVS

This paper addresses the detection of OOV segments in the output of large vocabulary continuous speech recognition (LVCSR) system. First, standard confidence measures based on frame-based wordand phone- posteriors are investigated. Substantial improvement was however obtained when posteriors from two systems - strongly constrained (LVCSR) and weakly constrained (phone posterior estimator) were combined. We show that this approach is suitable also for the detection of general recognition errors. All the results are presented on WSJ task with reduced recognition vocabulary.

• Type: Conference
• Year: 2008
• Where: ICASSP
• Authors: Lukas Burget, Petr Schwarz, Pavel Matejka, Mirko Hannemann, Ariya Rastrow, Christopher M. White, Sanjeev Khudanpur, Hynek Hermansky, Jan Cernocký
• Keywords: Continuous Speech Recognition, Frame-based Wordand Phone, ICASSP 2008, Phone Posterior Estimator, Signal Processing