Focused state transition information in ASR

We present speech recognition graphical models that use "focused evidence" to directly influence word and state transition probabilities in an explicit graphical-model representation of a speech recognition system. Standard delta and double delta features are used to detect loci of rapid change in the speech stream, and this information is applied directly to transition variables in a graphical model. Five different models are evaluated, and results are given on the highly mismatched training/testing condition tasks in Aurora 3.0. The best of these models gives an average 8% reduction in word error rate over baseline, significant at the 0.05 level