SYSTRAN / faster-whisper

Faster Whisper transcription with CTranslate2
MIT License
11.35k stars 948 forks source link

Get alternatives for each segment #444

Open Ar770 opened 1 year ago

Ar770 commented 1 year ago

Hi I was wondering if it would be possible to return for each segment more than one prediction?

phineas-pta commented 1 year ago

333

guillaumekln commented 1 year ago

By default the decoding is running a beam search which can return multiple predictions. However, the segments lengths are defined by the predictions themselves. So the different predictions could have a different segments structure which can not be aligned into alternatives.

On the other hand, it could be possible to return the alternatives when using without_timestamps=True. In that case all segments have a fixed 30-second duration.