Multi-Channel Speech Recognition: LSTMs All the Way Through
Hakan Erdoğan, Tomoki Hayashi, John R. Hershey, Takaaki Hori, Chiori Hori, Wei-Ning Hsu, Suyoun Kim, Jonathan Le Roux, Zhong Meng, Shinji Watanabe · 2016
Long Short-Term Memory recurrent neural networks (LSTMs) have demonstrable advantages on a variety of sequential learning tasks. In this paper we demonstrate an LSTM “triple threat” system for speech recognition, where LSTMs drive the three main subsystems: microphone array processing, acoustic modeling, and language modeling. This LSTM trifecta is applied to the CHiME-4 distant recognition challenge. Our previous state-of-the-art ASR systems for the previous CHiME challenge employed LSTM mask estimation based beamforming, noise robust features, in addition to DNN/RNNLM based back end. The proposed system refines each module of the previous system including bidirectional LSTM (BLSTM) mask estimation based beamforming, BLSTM-DNN hybrid acoustic model, and language model rescoring based on LSTM. We perform constrained re-estimation based speaker adaptation, and also prepare several complementary systems by changing the beamforming strategy and the acoustic model configurations, and combine these systems based on word-posterior based system combination. The final system achieved 2.98% WER for the real test set in the 6-channel track, which reduces the WER from the baseline by 8.5% absolute, and also outperforms our previous CHiME-3 system by 6.1% absolutely.