sherpa-onnx icon indicating copy to clipboard operation
sherpa-onnx copied to clipboard

[WIP] add CTC prefix beam search / hotwords / shallow fussion

Open pkufool opened this issue 1 year ago • 2 comments

This PR implements the core part (c++/python/JNI) of CTC prefix beam search related decoding methods, including hotwords and rnnlm shallow fussion.

  • [x] offline prefix beam search
  • [x] offline hotwords
  • [ ] offline rnnlm shallow fussion
  • [ ] online prefix beam search
  • [ ] online hotwords
  • [ ] online rnnlm shallow fussion

BTW we release our recent progress on CTC models, see https://arxiv.org/pdf/2410.05101 for details.

pkufool avatar Oct 17 '24 10:10 pkufool

请问这个request计划什么时候合并

fuyanzhe avatar Feb 21 '25 03:02 fuyanzhe

Hello. I trained CR-CTC model and decoded streaming CTC model and got token repetition (ex. ref: 안녕하세요 / hyp: 안녕녕하세요)

So, I really need online prefix beam search.... Do you have any plans to release online ctc prefix beam search? Thank you!

dohe0342 avatar Jul 22 '25 12:07 dohe0342