kensho-technologies / pyctcdecode

A fast and lightweight python-based CTC beam search decoder for speech recognition.
Apache License 2.0
421 stars 89 forks source link

Adding an abstract state to better support extending the package. #73

Closed lopez86 closed 1 year ago

lopez86 commented 2 years ago

The idea here is to reduce the explicit reliance on kenlm in the decoder class. This should handle everything except the save and load functions. With this change, it will be much easier to build new language models while still respecting the type hints.

lopez86 commented 1 year ago

LGTM -- do we definitely want to do a bump to 0.6.0 for this, though? Or just a 0.5.0 or 0.5.1?

I think we want 0.6.0 - there are some API changes in public functions even if they're things that probably won't really matter to current users