Our work is focus on robust acoustic model for the 5th CHiME5 challenge. To boost the performance to be better than baseline, we designed different network structure that can be suitable for CHiME5 challenge dataset. Our final result can achieve 13% relative improvement compare to the baseline.