Adaptive beamforming and adaptive training of DNN acoustic models for enhanced multichannel noisy speech recognition

This paper describes our contribution to the development of an ASR system for the CHiME 2015 Challenge. We applied a new adaptive beamforming method of multichannel alignment for enhancing speech recorded with six microphones. Then we trained an effective CD-DNN-HMM acoustic model using CMVN for noise robustness as well as fMLLR and i-vectors for speaker and environment adaptation. As a result, our system provides 7.33% WER on the development set and 14.34% WER on the test set (58% WER reduction compared to the baseline system).