Information Distance Based Self-Attention-BGRU Layer for End-to-End Speech Recognition

The common utilization of bidirectional gated recurrent unit (BGRU) architectures for end-to-end speech recognition suffers from long-term dependence and information redundancy. The reason lies in that the BGRU architectures model speech data according to time distance, which implicitly assumes that speech data is continuous. In this paper, we propose a new hypothesis, i.e., speech data possess the feature of being locally continuous and globally discrete. Based on this hypothesis, we propose to model speech data according to information distance. To support this hypothesis, we design an information distance based modeling architecture. Via the incorporation of self-attention mechanism, the proposed architecture is termed self-attention bidirectional gated recurrent unit (SABGRU). Experiment results show that SABGRU increases more than 10% speech recognition accuracy over conventional BGRU.