The SHNU system for the CHiME-5 Challenge

Abstract

This paper is a description of our system submitted to the 5th CHiME challenge. In this challenge, we only focus on the A-single-array task. Several improvements over the conventional ASR baseline have been used, including the speech processing front-end, the automatic training data augmentation of the official training data, and the acoustic model combination with different structures. The final overall WERs of both the development and evaluation sets only using the reference Kinect array are around 64%.