Non-Autoregressive Speech Recognition with Error Correction Module

Abstract

Autoregressive models have achieved good performance in the field of speech recognition. However, the autore-gressive model uses recursive decoding and beam search in the inference stage, which leads to its slow inference speed. On the other hand, the non-autoregressive model naturally cannot utilize the context since all tokens are output at one time. To solve this problem, we propose a position-dependent non-autoregressive model. And in order to make better use of contextual information, we propose a pre-trained language model for speech recognition, which is placed behind the non-autoregressive model as an error correction module. In this way, we exchanged a smaller amount of calculation for the improvement of the recognition rate. Our method not only greatly reduces the computational cost, but also maintains a good recognition rate. We tested our model on the public Chinese speech corpus AISHELL-1. Our model achieves a 6.5% character error rate while the real-time factor is only 0.0022, which is 1/17 of the autoregressive model.