Flexible discriminative training based on equal error group scores obtained from an error-indexed forward-backward algorithm

Abstract Thisarticlepresentsanewapproachtodiscriminativetrain-ing that uses equal error groups of word strings as the unit ofweighted error modeling. The proposed approach, MinimumGroupError(MGE),isbasedonanovelerror-indexedForward-Backward algorithm that can be used to generate group scoresefficiently over standard recognition lattices. The approach of-fers many possibilities for group occupancy scaling, enabling,for instance, the boosting of error groups with low occupan-cies. Preliminary experiments examined the new approach us-ing both uniformly and non-uniformly scaled group scores. Re-sults for the new approach evaluated on the Corpus of Spon-taneous Japanese (CSJ) lecture speech transcription task werecompared with results for standard Minimum Classification Er-ror (MCE), Minimum Phone Error (MPE) and Maximum Mu-tual Information (MMI), in tandem with I-smoothing. It wasfound that non-uniform scaling of group scores outperformedMPE when no I-smoothing is used.Index Terms: speech recognition, discriminative training