eduzhai > Applied Sciences > Engineering >

Transformer with Bidirectional Decoder for Speech Recognition

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 5 pages

Abstract: Attention-based models have made tremendous progress on end-to-end automaticspeech recognition(ASR) recently. However, the conventional transformer-basedapproaches usually generate the sequence results token by token from left toright, leaving the right-to-left contexts unexploited. In this work, weintroduce a bidirectional speech transformer to utilize the differentdirectional contexts simultaneously. Specifically, the outputs of our proposedtransformer include a left-to-right target, and a right-to-left target. Ininference stage, we use the introduced bidirectional beam search method, whichcan not only generate left-to-right candidates but also generate right-to-leftcandidates, and determine the best hypothesis by the score.To demonstrate our proposed speech transformer with a bidirectionaldecoder(STBD), we conduct extensive experiments on the AISHELL-1 dataset. Theresults of experiments show that STBD achieves a 3.6 relative CERreduction(CERR) over the unidirectional speech transformer baseline. Besides,the strongest model in this paper called STBD-Big can achieve 6.64 CER on thetest set, without language model rescoring and any extra data augmentationstrategies.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×