End-to-End Speech Recognition Methods Based on Convolutional Neural Networks and Bidirectional Gating Recurrent Unit

In view of the problem that the traditional acoustic model is complex and cannot be trained uniformly, and the data must be pre-aligned, this paper proposes a Chinese end-to-end speech recognition model based on convolutional neural network (CNN) and bidirectional gating recurrent unit (Bi GRU). This model uses the one-dimensional convolutional neural network to model the acoustics model, which can effectively extract the high-level abstract features with the context information, reduce the gradient disappearance, and uses the CTC beam search algorithm to realize the end-to-end training and decoding. Experimental results on the open-source dataset AISHELL-1 show a significant improvement in speech recognition performance compared to the baseline model, with 15.8% character error rate (CER) on the test set.