Hybrid HMM/BLSTM-RNN for Robust Speech Recognition

The question how to integrate information from different sources in speech decoding is still only partially solved (layered architecture versus integrated search). We investigate the optimal integration of information from Artificial Neural Nets in a speech decoding scheme based on a Dynamic Bayesian Network for noise robust ASR. A HMM implemented by the DBN cooperates with a novel Recurrent Neural Network (BLSTM-RNN), which exploits long-range context information to predict a phoneme for each MFCC frame. When using the identity of the most likely phoneme as a direct observation, such a hybrid system has proved to improve noise robustness. In this paper, we use the complete BLSTM-RNN output which is presented to the DBN as Virtual Evidence. This allows the hybrid system to use information about all phoneme candidates, which was not possible in previous experiments. Our approach improved word accuracy on the Aurora 2 Corpus by 8%.