A non-parametric minimax approach for robust speech recognition

Robust statistical procedures are studied and applied to speech recognition. The goal is to improve recognition under general nonparametric mismatch situations between training and testing conditions. Towards this end, we develop an M-hypotheses decision rule for a statistical model related to hidden Markov model. The decision rule employs two hypotheses robust likelihood ratio tests between all pairs of the M hypotheses, and is shown to be asymptotically optimal for the worst case mismatch condition. The proposed approach is experimentally evaluated and compared to a known parametric approach where the mismatch is modeled parametrically, and to the standard MAP approach, where no mismatch is assumed.