A Novel Noise-Robust ASR Method by Applying Partially Connected DNN Model and Mixed-Bandwidth Concept

In recent years, deep neural networks achieve significant improvements in automatic speech recognition. In this paper, we propose a deep structure used for robust ASR. The model has several partially connected layers which can suppress noise in different frequency bands. In order to recognize the speech data which has been distorted by noise seriously, we try to use parts of their frequency bands with a mixed-bandwidth model. The results have shown that the partially connected network could suppress noises in different frequency bands properly. The model's phone recognition on TIMIT corpus outperforms the state-of-the-art DNN model. Keywords-robust; ASR; DNN; mixed-bandwidth