The Fawaispeech System for Multi-Channel Speech Recognition in ICMC-ASR Challenge

This paper describes our FawAISpeech system for the Automatic Speech Recognition (ASR) task in the In-car Multi-Channel Automatic Speech Recognition Challenge (ICMC-ASR). Specifically, the weighted prediction error (WPE), Minimum Variance Distortionless Response (MVDR) and Guided Source Separation (GSS) are used to enhance the target speaker speech. Then, we explore the effectiveness of E-Branchformer based multi-channel cross-attention ASR systems. Experiments show that our system achieves a character error rate (CER) of 14.72% on the Eval set, and obtains a third place in the challenge.