Adaptive windows multiple deep residual networks for speech recognition

作者：

Highlights：

• We propose a new method that changes the structure of CNN and Res-Net.

• The method adapts speaking and speaker variations and deals with very deep models.

• The network can model the phone duration variations using different window sizes.

• This method preserves the input information in its deep layers.

• Achieving high recognition performance over the state-of-the-art approaches.

摘要

•We propose a new method that changes the structure of CNN and Res-Net.•The method adapts speaking and speaker variations and deals with very deep models.•The network can model the phone duration variations using different window sizes.•This method preserves the input information in its deep layers.•Achieving high recognition performance over the state-of-the-art approaches.

论文关键词：Speech recognition,Deep neural network,Adaptive windows convolutional neural network (AWCNN),Multiple Residual Networks (MRes),Adaptive windows convolutional neural network and Multiple Residual networks (AMRes)

论文评审过程：Received 12 July 2018, Revised 18 July 2019, Accepted 24 July 2019, Available online 29 July 2019, Version of Record 2 August 2019.

论文官网地址：https://doi.org/10.1016/j.eswa.2019.112840