Violin: An open-source video translation skill that breaks language barriers
Together AI
Together AI released Violin, an open-source tool that translates videos into other languages, dubbing and all. It's free to use and build on, tackling the fact that most big YouTube channels only publish in English.
Based on reporting by Together AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Most of the internet's biggest video channels speak one language, and it isn't yours unless you happen to speak English. Together AI wants to chip away at that with Violin, a fully open-source pipeline that takes a video in one language and spits out a dubbed, translated version in another, complete with a chatbot that can answer questions about what's on screen.
The mechanics are refreshingly unglamorous. Violin runs a video through Whisper V3 for transcription, hands the timestamped text to DeepSeek V4 Pro for translation, then feeds the result to Cartesia's Sonic 3 for text-to-speech. Users can even type plain-language instructions to pick a voice style, and the system supports natural-sounding speakers in languages like Korean, Dutch, Italian, and Chinese. Notably, Together AI built in a guardrail against voice cloning — the dubbed audio uses a distinct synthetic voice rather than mimicking the original speaker, layered quietly over the source audio instead of replacing it outright.
The chat assistant is the part that separates Violin from a basic dubbing script. It's powered by a vision-language model, Qwen3.5-397B-A17B in this case, that samples recent video frames alongside subtitle text so it can answer questions grounded in both what's said and what's shown. Ask it to summarize a section or clarify a point buried ten minutes into a talk, and it pulls from audio and visuals rather than transcript alone.
Together AI is shipping this as a web app for people who just want to upload a file, a CLI for anyone scripting batch jobs, and a packaged agent skill for developers wiring it into autonomous workflows. The whole thing is MIT-licensed, so anyone can fork it, tweak the translation rules, or swap in different underlying models. The team demoed it on a talk from Stanford's Percy Liang, translating an English tech presentation into Chinese as a proof of concept.
The motivation traces back to a lopsided stat: by one estimate, 66 percent of videos on the top 250 YouTube channels are in English, with Spanish trailing at just 15 percent. That's a lot of content sitting behind a language wall for most of the planet. Violin doesn't solve that on its own, but it's a concrete, open attempt at lowering the wall a little.
My take — AI-written commentary, not fact-checked reporting
I like that this exists as open infrastructure rather than another walled-off dubbing SaaS charging per minute — that's exactly the kind of unglamorous, useful tool the open-model ecosystem should be producing instead of chasing another leaderboard win. The voice-cloning guardrail is a smart, responsible touch too, though I'd bet someone forks this within a month and rips it out.
Read more about this at: Together AI