Skip to content

STE-58: narrate the demo video, and stop it sitting still - #72

Merged
Lin1er merged 1 commit into
mainfrom
feat/ste-58-voiceover
Sep 28, 2026
Merged

Lin1er merged 1 commit into
mainfrom
feat/ste-58-voiceover

Conversation

@Lin1er

@Lin1er Lin1er commented Sep 28, 2026

Copy link
Copy Markdown
Owner

The first cut was silent and static. Three minutes of screencast with captions and hard cuts is a long time to watch nothing move and hear nobody speak.

What changed

  • Narration across all twenty scenes. Scene lengths are now driven by the narration rather than the reverse, so no sentence is clipped. Where a line runs longer than its footage the last frame is held — which is why the poisoned refusal and the 401 now sit on screen long enough to actually read.
  • Motion. A slow push on the cards, a few pixels of drift on the footage. No shot is frozen any more.
  • Transitions. A different wipe, slide or dissolve between each pair of scenes instead of fading to black nineteen times.
  • Audio normalised to −16 LUFS with 1.2 dB of headroom, so it will not clip on a laptop speaker.

Same footage, same transactions, same 2 min 54 s.

The voice is synthetic, and the README says so

Generated with Piper (en_US-ryan-high) from narration.json, which is committed beside the video with the script word for word. Nobody in the team recorded it. That is stated in the README rather than left for a viewer to wonder about, and the committed script means every sentence can be checked against the frame it plays over.

The narration claims nothing the earlier cut did not: submission is still operator-run, every wallet belongs to the team, and the model never decides a verdict.

cc @axel @fable @Nabil-Aufa

The first cut was silent and static: three minutes of screencast with captions
and hard cuts, which is a long time to watch nothing move and nobody speak.

Adds narration across all twenty scenes, a slow push on the cards, a few pixels
of drift on the footage so no shot is frozen, and a different transition between
each pair of scenes instead of a fade to black. Scene lengths are now driven by
the narration rather than the other way round, so nothing is clipped mid
sentence; where a scene's line runs longer than its footage the last frame is
held, which is why the poisoned refusal and the 401 now sit on screen long
enough to read.

The voice is synthetic — Piper, en_US-ryan-high — and the script is committed
beside the video as narration.json, so the claim in every sentence can be checked
against the frame it plays over. The README says so plainly rather than letting a
viewer assume someone recorded it.

Audio is loudness-normalised to -16 LUFS with 1.2 dB of headroom, so it will not
clip on a laptop speaker. Same footage, same transactions, same 2 min 54 s.
@vercel

vercel Bot commented Sep 28, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
sterish Ready Ready Preview Sep 28, 2026 3:09pm UTC
sterish-landing Ready Ready Preview Sep 28, 2026 3:09pm UTC

@Lin1er
Lin1er merged commit 4337308 into main Sep 28, 2026
4 checks passed

This branch was successfully deployed

2 active deployments
Preview – sterish — 376bf39e Deployed Sep 28, 2026 by vercel[bot]
Preview – sterish-landing — 376bf39e Deployed Sep 28, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant