Dub Studio：本地 AI 视频配音

Free offline AI video dubbing studio for Windows — voice cloning, translation, subtitles & on-screen-text localization. 100% local, one n…

Dub Studio Free, offline AI video dubbing studio for Windows — re-voice any video into another language with a cloned voice, translated captions, and on‑screen‑text localization. 100% local, zero Python: one native .exe (Rust + C++/CUDA); every model and engine downloads with a button. English · Русский · 中文 · Español · Português · Français 🌐 Live demo & before/after video showcase → What it is Dub Studio turns any video into a dubbed version in another language — with the speaker's own voice cloned, captions translated, and on‑screen text localized right on the frame. Drop a clip, and a smart auto‑pass builds the first draft; then a live editor puts every caption, voice, blur box, font and title under your control with an instant preview. By default everything runs locally on your machine — no cloud, no subscription; your footage and your voiceprint never leave your computer. And if your PC is weak (can't run the local Gemma/Higgs) or you want more speed and quality, the heavy stages (translation, vision, TTS, transcription) can optionally be offloaded to the cloud via OpenRouter — each engine picked independently (local ↔ cloud), with voices auto-cast by speaker gender (beta). The key is stored locally; everything is off by default. It's a fully native rewrite. No embeddable Python, no torch, no CUDA wheels. The whole pipeline is Rust + native C++/CUDA engines (GGUF/ONNX): one process, fast startup, low VRAM. Models, engines, CUDA/VC++ runtime and ffmpeg are downloaded and installed by the app itself on first run. An NVIDIA GPU is what the app is built and tested for: voice, translation and vision run on CUDA locally, and subtitles are burned into the video with NVENC. Separation, diarization and recognition can run on the CPU and the heavy stages can go to OpenRouter, but a machine without NVIDIA is not a tested setup - see What runs where below. See it in action ▶ Watch the before/after video showcase → — real clips dubbed end‑to‑end on a local GPU: different videos, different modes, different languages, nothing left the machine. 🎙️ Full dub · EN→RU 🗣️ Voice-over · EN→ES 🈶 Full dub · →中文, CJK on frame 📝 Subtitles · original-lang 🎬 Full dub · widescreen 16:9 🔤 Transcript · diarized For AI agents While Dub Studio is open it serves MCP at http://127.0.0.1:8793/mcp: an agent such as Claude Code, Claude Desktop, Cursor or Codex does everything the window does, through the same code — makes projects of video files, analyzes them, fixes the translation, timing and speakers line by line, casts the voices, styles the subtitles, adds titles and blur boxes, renders, exports the same video in more languages, writes SRT and TXT, and saves the results to a folder. Settings, Agent (MCP) shows whether an agent is connected and what to paste into the client. Given this repository, an agent can set everything up and drive the studio by itself: Install the studio from the latest release and start it. Connect to its MCP server: claude mcp add --transport http dub-studio http://127.0.0.1:8793/mcp Other clients: { "mcpServers": { "dub-studio": { "type": "streamable-http", "url": "http://127.0.0.1:8793/mcp" } } } Read the skill the server serves (resource studio://skill, prompt studio), the same text as docs/mcp-skill.md — every tool, the ground rules and step-by-step recipes — and start with the tool studio_status. llms.txt says the same for tools that look for it. To keep the skill in Claude Code, save docs/mcp-skill.md as ~/.claude/skills/dub-studio/SKILL.md. Only agents on this computer and the studio's own window may connect. Five modes, switchable on the fly Mode What it does 🎙️ Dub Full re-voice into the target language with the original timbre cloned — auto-cast per speaker or pick a voice 🗣️ Voice-over Translated voice over the ducked original — the source is still audible underneath; balance is adjustable 📝 Subtitles Burn original-language captions, keep the original audio — no dubbing, no translation ✨ Funny remix Give a theme ("as a pirate", "as a news report") → the model rewrites the whole script, then re-dubs 🎬 Transcript Clean diarized transcript with per-speaker layout, karaoke play-along, one-click voice creation, .srt/.txt export Load a clip once and send it into any mode — right inside the editor. Features Voice cloning — the original timbre is cloned and speaks the new language (native Higgs Audio v3 engine, GGUF). Auto-cast by speaker or bring your own voice from a pack. Speaker diarization — who speaks and when (NVIDIA Nemotron 3 Diarization, up to 8 voices), a distinct voice per speaker. Character casting (beta) — a character is a face + voice pair. The app gathers faces across the whole video, recognizes the same person and binds them to a speaker by co-occurrence (the one on camera in close-up gets the voice, a background listener doesn't); it auto-picks the clearest avatar frame and saves a casting profile for the whole series — assign voices and character descriptions once, and the next episode applies them automatically. A real-faces / cartoon·anime toggle switches the face detection accordingly. Choice of ASR engine — transcribe with Parakeet-TDT (GPU, default) or Whisper (Purfview faster-whisper standalone, runs on CPU) — pick the model size (tiny … large-v3-turbo) and quant (compute type) right in settings. Import ready-made subtitles — bring your own .srt/.ass as the exact transcript: text and timing come straight from the file instead of auto-recognition (speakers are still auto-assigned by diarization). Tick “subtitles already in the target language” and translation is skipped too — an English clip + your Russian subs → a Russian dub straight from them, no ASR and no MT. Multi-language export — the ▾ next to Export sends one video into several languages at once; each inherits all your edits (subtitle layout, styles, blur boxes, cloned voice) — only the text is re-translated and re-voiced. Save & reopen projects — autosave, a list of recent projects on the start screen, an