Skip to content

Add CTC-LM-beam-search decoder. #2230

@xinghai-sun

Description

@xinghai-sun
  • The current paddle.layer.beam_search decoder only supports frame-wise decoding, not CTC decoding.
  • Add paddle.layer.ctc_beam_search to support CTC beam search decoding, with external language model scorer and word-count scorer. Please refer to Prefix Beam Search, CTC paper, and DS2 paper
  • It would be better to re-design the above layers, and provide a more general interfaces for beam search with any external custom scoring function.
  • Keep in mind that it would be better if it is compatible with both English (word-based LM) and Madarin (character-based LM).
  • Please add a design-doc here first, illustrating:
    • the algorithm of CTC prefix beam search with LM and WordCount;
    • how to design a general beam search interface.
  • This can also be used to improve the current OCR model performance.

Metadata

Metadata

Labels

No labels
No labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions