abus-aikorea/voice-pro

The problem

Dubbing a video into another language normally means stitching together several unrelated tools: a downloader, a vocal separator, an ASR model, a translator, and a TTS engine, each with its own install and audio format quirks. Creators, podcasters and localization researchers end up gluing these together by hand or paying per-minute cloud services. Voice-Pro bundles the whole chain into one Gradio web app that runs on your own machine, from YouTube URL to finished dubbed audio.

What's different

The usual alternative is a hosted service such as ElevenLabs, or a self-built pipeline of Whisper plus a TTS library. Voice-Pro's distinguishing move is breadth in one installer: it wraps Whisper/Faster-Whisper/Whisper-Timestamped for recognition, Deep-Translator for 100+ languages, Demucs/MDX-Net for vocal isolation and several zero-shot cloning models (F5-TTS, E2-TTS, CosyVoice) alongside Edge-TTS and kokoro. In v4 the installer moved from Miniconda/pip to uv with a committed lockfile, and CUDA Toolkit/Visual Studio build tools are no longer needed because dependencies ship as prebuilt wheels. That packaging work, not any single model, is what makes it more than a demo.

Who it's for

Individual creators, dubbing hobbyists and small localization teams on Windows with an NVIDIA GPU who want an offline, no-subscription workflow and are willing to accept model download sizes and quality that varies by language and voice. It is not for high-volume production pipelines needing SLA-backed output, nor for Linux/Mac-first teams: the README explicitly says only Windows + NVIDIA is verified, and even the license line is inconsistent (GPL-3.0 in metadata, LGPL in the header). Developers who only need one component should use that library directly rather than adopting a whole monolith.

Cost of adoption

You need Windows, an NVIDIA GPU with a substantial amount of VRAM for the cloning models, and disk space for the model zoo; first runs download weights from Hugging Face. Translation defaults to a free Google endpoint that can rate-limit, and optional Azure TTS/Translator require your own keys. Install is via the bundled start.bat, which bootstraps uv and Python locally; see the README for details. Most importantly, the README states development is paused for a separate product, so treat this as a frozen, source-available tool rather than an actively maintained platform, and check the license file before redistributing.

Verdict

Voice-Pro is best read as a well-packaged integration layer, and the packaging has clearly mattered: 13.3k stars and a 271-point Show HN thread from late 2024 show real interest, and a +339 week on the Python trending board in 2026 indicates the v4 rewrite revived attention. But the authors say updates are on hold, so the risk is drift: upstream model APIs and Gradio versions will move and this repo may not follow. If you need an offline dubbing GUI now, on Windows with an NVIDIA card, it is a reasonable one-stop choice; if you need long-term maintenance or a scriptable pipeline, borrow the model choices and build your own.