The Munich Feature Enhancement Approach to the 2nd Chime Challenge Using Blstm Recurrent Neural Networks

We present a highly efficient, data-based method for monaural feature enhancement targeted at automatic speech recognition (ASR) in reverberant environments with highly non-stationary noise. Our approach is based on bidirectional Long Short-Term Memory recurrent neural networks trained to map noise corrupted features to clean features. In extensive test runs, enhanced features are evaluated with gradually refined recognition back-ends, reaching from simple maximum likelihood (ML) trained recognisers to state-of-the-art ASR using discriminative training and model adaptation techniques. In the result, consistent improvements over the baseline ASR systems on both the small and medium vocabulary tasks of the 2nd CHiME Speech Separation and Recognition Challenge demonstrate the efficacy of the proposed method, achieving up to 52 % relative reduction of word error rate with respect to the multi-condition ML training baselines.