Bidirectional LSTM networks for improved phoneme classification and recognition

Abstract. In this paper, we carry out two experiments on the TIMIT speech cor-pus with bidirectional and unidirectional Long Short Term Memory (LSTM) net-works. In the first experiment (framewise phoneme classification) we find that bidirectional LSTM outperforms both unidirectional LSTM and conventional Re-current Neural Networks (RNNs). In the second (phoneme recognition) we find that a hybrid BLSTM-HMM system improves on an equivalent traditional HMM system, as well as unidirectional LSTM-HMM.