A Light-Weight Autoregressive CNN-Based Frame Level Transducer Decoder for End-to-End ASR

Abstract

A convolutional neural network (CNN) transducer decoder was proposed to reduce the decoding time of an end-to-end automatic speech recognition (ASR) system while maintaining accuracy. The CNN of 177 k parameters and a kernel size of 6 generates the probabilities of the current token at the token level, at the token transition of the output token sequence. Two probabilities of the current token, one from the encoder and the other from the CNN are added to the frame level to reduce the decoding step to the number of input frames. An encoder composed of an 18-layer conformer was combined with the proposed decoder for training with the Librispeech data set. The forward-backward algorithm was used for training. The space and re-appearance tokens are added to the 300-word piece tokens to represent the token string. A space token appears at a frame between two words. A comparison with the autoregressive decoders such as transformer and RNN-T decoders demonstrates that this work provides comparable WERs with much less decoding time. A comparison with non-autoregressive decoders such as CTC indicates that this work enhanced WERs.

Keywords

• speech recognition
• autoregressive speech recognition
• end-to-end
• CNN
• transducer decoder