Improve Mandarin Speech Recognition Using Simple Recurrent Units and SpecAugment

Simple Recurrent Units (SRU) is one kind of Recurrent Neural Networks, which has been explored in some tasks. However, SRU hasn't been widely used in automatic speech recognition task. In this paper, we explore one SRU to improve the hybrid Connectionist Temporal Classification and Attention based Mandarin speech recognition system using bidirectional long short-term memory (Bi-LSTM). Besides, we also use the data augmentation strategy of SpecAugment to further improve the system. The experiments on Aishell-1 dataset show that bi-directional (Bi-SRU) outperforms BiLSTM as encoder, and some optimization strategies like label smooth and scheduled sampling also improve the accuracy. The model trained using Bi-SRU and SpecAugment achieves the best performance in the experiments.