Parallel Gated Recurrent Unit Networks as an Encoder for Speech Recognition

Listen, Attend and Spell (LAS) network is one of the end-to-end approaches for speech recognition, which does not require an explicit language model. It consists two parts; encoder part receives acoustic features as inputs, decoder produces character at a time step, based on output attention mechanism. Multi-layer recurrent neural networks (RNN) are used in both parts. Hence, LAS architecture can be simplified RNN decoder, another encoder. Their shapes layer sizes different. In this work, we examined performance using multi RNNs part. Our baseline uses with hidden size 256. We 2 4 128 64 each case. The main idea behind proposed approach to focus different patterns (phonemes case) data. At encoder, their outputs concatenated fed decoder. TIMIT database compare mentioned networks, phoneme error rate metric. experimental results showed that achieve better than network. However, increasing number guarantee further improvements.