# Trackdub

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Michael Anderson](https://cerebralvalley.ai/u/canopylions)
- **GitHub:** https://github.com/MesilimLLC/Trackdub-Director
- **Demo video:** https://www.youtube.com/watch?v=6dTZjwIx-nE
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5032

Trackdub-Director is an open-source dubbing workbench (C# / ASP.NET Core + React) built for the Build for Your World track. You already posted a Spanish Reel. You can hear it. You cannot check it. This is the check.

Give it a clip in English. It dubs it into Spanish, then verifies its own work line by line. Every line must pass two gates: a timing gate that estimates the spoken length against the original window, and a meaning gate that judges whether the translation still says what the original said. If timing fails first, meaning is recorded as NotEvaluated and no tokens are spent: a check that never ran never reads as a pass. Passing lines are synthesized (ElevenLabs, disclosed; the Meta catalog has no TTS endpoint, and no performer is imitated or cloned), the real audio is measured against the window, and a 2.46 s take against a 2.2 s window is rejected with the numbers printed. Every run leaves a verify-only receipt: source, shipped Spanish, measured audio, verdict, and the gate's reason in English. One-click verification, deliberately not one-click dubbing.

The Meta models also read the scene: SAM 3.1 grounds elements like the coffee-shop sign (verified against its own reflection), scene context can request a register change that must re-pass both gates, and Muse Image repaints the sign itself on still frames (CAFETERÍA), with every pixel outside the box machine-checked against the source: difference zero. Emotion and pacing tags extracted by the models are carried verbatim into the ElevenLabs synthesis, so the delivery follows the performance. The whole flow runs in a browser workbench a creator can actually use, ending in a dubbed MP4 export.

Open weights, measured: Muse Spark Open judged independently and agreed with the hosted judge on 73 of 75 verdicts, stricter on both disagreements; Muse Glimmer recovered 4 of 4 refused lines (two winners telegraphic, disclosed in the repo). Three native Spanish speakers rated the gate's output blind; all three failed both gate-passed pressure lines on sight, and two of three failed a shipped line for a plural the gate passed, recorded in the limitations rather than quietly fixed. Fluency is the known, named gap.

## More from Meta Global AI Developer Hackathon

- [Physica](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5029)
- [straw hats](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5030)
- [SOUS AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5031)
- [JupiterOne](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5033)
- [Telltale](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5034)
- [Quanticomp](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5035)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5032. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt