Multiresolution CNN for reverberant speech recognition

The performance of automatic speech recognition (ASR) has been greatly improved by deep neural network (DNN) acoustic models. However, DNN-based systems still perform poorly in reverberant environments. Convolutional neural network (CNN) acoustic models showed lower word error rate (WER) in distant speech recognition than fully-connected DNN acoustic models. To improve the performance of reverberant speech recognition using CNN acoustic models, we propose the multiresolution CNN that has two separate streams: one is the wideband feature with wide-context window and the other is the narrowband feature with narrow-context window. The experiments on the ASR task of the REVERB challenge 2014 showed that the proposed multiresolution CNN based approach reduced the WER by 8.79% and 8.83% for the simulated test data and the real-condition test data, respectively, compared with the conventional CNN based method.