Confusion-Based Entropy-Weighted Decoding for Robust Speech Recognition

An entropy-based feature parameter weighting scheme was proposed previously [1], in which the scores obtained from different feature parameters are weighted differently in the decoding process according to an entropy measure. In this paper, we propose a more delicate entropy measure for this purpose considering the inherent confusion among different acoustic classes. If a set of acoustic classes are easily confused, those feature parameters which can distinguish them should be emphasized. Extensive experiments with the Aurora 2 testing environment verified that this approach is equally useful for different types of features, and can be easily integrated with typical existing robust speech recognition approaches.