An Efficient Method To Accelerate Rnn-T Decoding For Asr

End-to-end automatic speech recognition has been widely studied in recent years, and the transducer model has attracted a lot of attention. However, the decoding process of transducer models is slow and allocate large amount of memory. In order to speed up the decoding of the transducer model, this paper adopts a series of methods to limit the dimensions of the transducer model to a reasonable range: First, we use Byte BPE modeling instead of character-level modeling for Mandarin. Second, we design a simple but effective frame reducer module that speeds up the decoding of transducer models with only a slight compromise in accuracy. Finally, a K-bank CNN stateless method is applied to the prediction network, which vests the prediction network with the ability to focus on different levels of linguistic dependencies and further improves the recognition accuracy. We show that the proposed method can accelerate the decoding speed and improve the recognition accuracy on AISHELL-1 dataset.