Language Detection - Githubissues

argmaxinc / WhisperKit

On-device Speech Recognition for Apple Silicon

https://takeargmax.com/blog/whisperkit

MIT License

3.17k stars 268 forks source link

Language Detection #78

Closed Abhinay1997 closed 6 months ago

Abhinay1997 commented 6 months ago

PR for the issue #29.

Adds LanguageLogitsFilter that suppresses non-language logits
Implements detectLanguage in TextDecoder that does a single forward pass from <startoftranscript> to get the language token
Adds tests for LanguageLogitsFilter

ZachNagengast commented 6 months ago

Added a couple CI fixes here FYI @Abhinay1997 should be in a working state now

Abhinay1997 commented 6 months ago

Thanks for the feedback ! Yes, let me add tests for Spanish and Japanese. Will need to look into the filtering time !

Abhinay1997 commented 6 months ago

@ZachNagengast, for comparing the detection with and without logits filter, I'll add a applyFilter param to detectLanguage then ? To toggle it on and off.

Abhinay1997 commented 6 months ago

The filter adds only 2.9802322387695312e-06 seconds. Negligible imo