ContentRoof Lab · Research notes
Notes from inside a video engine
ContentRoof turns one sentence into a narrated, animated video in eleven languages. These notes are written by the people who build it: the problem, what we tried, what failed, what we measured, and where it still falls short.
3D simulation · the heart
3D simulation · a jet engine
Explainer · Arabic
Cinematic storyResearch notes
04
A billion is not a billion in Spanish
Spanish billón, German Billion and French billion mean a million million. How we check every number when a video is rebuilt in another language: 8 of 8 planted errors caught.
8 of 8planted number errors caught
03
Sixty seconds is a different length in every language
Asked for 60 s, our videos came out 44–105 s. Measuring each language's voice per letter now lands 43–62 s in eleven languages, and a cloned voice took Arabic errors from 10% to 0.
44–105 s → 43–62 sasked for 60 s, in eleven languages
02
Balanced benchmarks hid a 95% error rate
Our sentence classifier scored 79–95% on balanced tests. On real narration it gave 95% of plain statements a shape they don't have. What we measured and what fixed it.
5% → 88%plain statements left alone
01
We took the browser out of video export
Our exports spent 80% of every frame waiting for Chrome's screenshot. We draw the same React components with Skia instead: 2.4× faster on a server, 5× on a laptop GPU.
2.4×faster export, the same frames