A DYANAMIC CROSS-REFERENCE PRUNING STRATEGY FOR MULTIPLE FEATURE FUSION AT DECODER RUN TIME

摘要

Although multiple cues, such as different signal processing techniques and feature representations, have been used in speech recognition in adverse acoustic environment, how to maximally utilize the benefit of these cues is largely unsolved. In this paper, a novel search strategy is proposed. During parallel decoding of different feature streams, the intermediate outputs are cross-referenced to reduce pruning errors. Experiment results show this method significantly improved recognition performance on a noisy large vocabulary continuous speech task.