Use CutSet for whisper annotation workflow #834
Conversation
…instead of touching the audio file
desh2608
changed the title
… feature/whisper_cut
|
@pzelasko I have separated it into two helper functions. Let me know if it looks reasonable now. |
lhotse/workflows/whisper.py
Outdated
| id=f"{cut.id}-{segment['id']:06d}", | ||
| recording_id=cut.recording_id, | ||
| start=round(segment["start"], ndigits=8), | ||
| duration=round(segment["end"], ndigits=8), |
There was a problem hiding this comment.
I think with this single change we can prevent timestamps going out of cut's duration; can you also include _postprocess_timestamps in the cut-based workflow?
| duration=round(segment["end"], ndigits=8), | |
| duration=max(cut.duration, round(segment["end"], ndigits=8)), |
There was a problem hiding this comment.
BTW, why are we using "end" as the duration? Shouldn't it be "end-start"?
There was a problem hiding this comment.
Because in my own testing I discovered that their "end" is actually a "duration" 😂 but if you could triple-check that I got it right, maybe using some longer recording, that'd be great.
There was a problem hiding this comment.
So I ran the model on one of the AMI headset recordings (~5000s) and it seems like the "end" actually shows the end of the segment, not the duration. Here is the JSON containing the results["segments"]: https://drive.google.com/file/d/169igkcDY2SmMs5k3hOhHip89T4MQDnKs/view?usp=sharing
|
Thanks, LGTM |
This PR allows transcribing a CutSet with the whisper workflow. It is currently WIP since it relies on #832 and #822. But I have verified that it works.