[yashranaway]

generating captions decoded and transcribed every recording segment in full, ignoring the timeline. so a 30 minute recording trimmed down to 1 minute still sent all 30 minutes through the model, which is slow and was half of what #2719 hit. the long input crash itself was already fixed by #2539.

now it works out the source ranges the timeline actually keeps, merges overlaps, joins just those into one audio buffer and runs the model once, then maps word times back to source time through an offset table so caption placement doesn't change. if an undo brings back audio while it's running, it asks you to generate again instead of applying partial captions. the captions oracles pass.

fixes #2719

Confidence Score: 5/5

The PR appears safe to merge; no new actionable issues were found.

Summary

Caption generation now reads only audio kept by the timeline. Overlapping source ranges are merged, joined for one transcription call, and mapped back to source times.

• Caption generation transcribes only audio kept by the timeline.
• The editor checks that the timeline still fits the transcribed audio.

_{Reviews (3) · Last reviewed commit: "fix(desktop-gpui): transcribe kept range..." · Reviewed by Greptile}

[vercel]

@yashranaway is attempting to deploy a commit to the Cap Software Inc Team on Vercel.

A member of the Team first needs to authorize it.

[yashranaway]

@greptileai review

[yashranaway]

@greptileai review

[superagent-security]

Superagent found 1 security concern(s).

[yashranaway]

@greptileai review