Whisper decoder loops (one word or a prompt list cycling for minutes) are collapsed after transcription, and each collapse shows up as a transcript fix
Also released that day: v0.22.0, v0.21.0, v0.20.0, v0.19.1, v0.19.0
- #161 7fe603f Whisper decoder loops (one word or a prompt list cycling for minutes) are collapsed after transcription, and each collapse shows up as a transcript fix. The as-heard repeats stay in the transcript; rejecting the fix puts them back. Captions stay empty across that stretch unless you type replacement text on the flag. Re-transcribe to recover the missing speech.
- #159 bdd11ae Framing follows the source's cutting. Shot cuts are now detected for every project before clip selection (they used to exist only when a project opted into the scenes refine step, which is now just the on-screen text scan), clip boundaries snap to them, a render detects them on the clip when an older project or an imported report has none, and the layout may change at every cut: a multi-camera talk show that alternates a close-up with a wide shot every second or two gets a full-frame close-up, then the wide shot, then the close-up again, instead of one split held across a dozen shots with a pane framed on an empty seat. A wide shot of three or more people is shown whole rather than split between two of them. Name fixes and the on-screen text scan now run by default (instance gates TRANSCRIPT_CORRECTION and REFINE_SCENES, project settings refineEntities and refineScenes); filler removal stays opt-in. Importing an issue report whose project was re-rendered no longer fails on its timestamps.