I built a fully local AI pipeline that attempts to dub entire films unattended. Looking for people with NVIDIA GPUs to break it and contribute
I've been working on Dubline, an open-source project that tries to take a foreign-language film and produce a complete English dub entirely locally.
No cloud inference, hosted APIs or paid services.
The basic pipeline is:
video → dialogue/M&E separation → ASR + forced alignment → speaker/character tracking → translation/adaptation → voice cloning/TTS → timing/QC → final mastered MKV
It currently uses a slightly ridiculous collection of models/tools including:
- Qwen3-ASR / Whisper
- IndexTTS 2.5 / Qwen3-TTS
- Hy-MT2 for translation/adaptation
- pyannote + CAMPPlus for speaker identification
- Bandit v2 / RoFormer / Demucs for separation
- OpenCV face tracking
- llama.cpp
- FFmpeg / Rubber Band
- Optional MuseTalk lip-sync
The aim isn't just "translate subtitles and run TTS". I'm trying to preserve character identity, timing, pauses, emotional delivery, non-verbal sounds, the original music/effects bed, loudness, etc., while having the pipeline automatically retry or flag lines when something looks wrong.
It's designed around consumer NVIDIA hardware, with the heavier models isolated into separate processes so VRAM can actually be released between stages. 8GB is the theoretical minimum, 12GB+ is much more sensible.
It's still very much a work in progress, and that's basically why I'm posting.
I'd really like some other people to install it, throw horrible real-world material at it and tell me where it falls apart.
Things I'm particularly interested in testing:
- Different NVIDIA GPUs / VRAM amounts
- Windows vs Linux
- Different source languages
- Bad/noisy/old film audio
- Multiple speakers and overlapping dialogue
- Subtitle-less material
- Long films rather than tiny demos
- Speaker identification failures
- TTS/emotion/timing problems
- Installation/setup failures I've accidentally made invisible on my own machine
And if anyone finds the project interesting and wants to contribute, PRs/issues are very welcome. There are plenty of areas where somebody who actually knows audio, ML, Python, video processing or UI development better than I do could improve it.
I'm especially interested in people trying to break it rather than telling me it looks cool from the README.
GitHub:
https://github.com/leighrobertabbott/Dubline
If you try it, even "it exploded at step 3 on my 4070 because of X" would genuinely be useful feedback.