// the find
liuzhao1225/YouDub-webui
开源 AI 视频本地化工具：自动完成 YouTube/Bilibili 视频下载、字幕识别与翻译、语音克隆配音、音轨混合和字幕压制。
YouDub WebUI is a local video dubbing pipeline that takes a YouTube or Bilibili URL and outputs a new video with cloned voice dubbing in the target language, complete with translated subtitles and the original background audio preserved. It's built and battle-tested by a creator with 800k+ Bilibili subscribers who uses it in daily production. The primary supported path is English-to-Chinese; Chinese-to-English works but is less mature.
The pipeline is genuinely end-to-end: download, vocal separation (Demucs), ASR (Whisper large-v3-turbo), translation (any OpenAI-compatible API), voice cloning (VoxCPM2), timing alignment, and FFmpeg subtitle burn — no manual steps between URL and final mp4. Stage-level resumability is well thought out: a failed run restarts from the last failed stage rather than from scratch, which matters when you're processing long videos with heavy GPU models. The security posture is unusually careful for a personal-project tool — Argon2id password hashing, per-session CSRF tokens, HttpOnly cookies, and strict file permission enforcement on startup. Bringing Demucs in as a source submodule rather than a pip dependency gives you the actual upstream code with reproducible behavior, and the model cache and SQLite state are fully local.
The in-process background worker means one crashing task can take down the FastAPI server; there's no process isolation or queue worker separation, so multi-user or high-volume use is fragile by design. VoxCPM2 is a ModelScope-hosted model with no Western mirror, which makes first-run setup genuinely painful outside China and introduces a hard dependency on a single vendor's hosting. The translation quality and timing alignment for long-form video is unproven beyond the author's own use case — dubbed speech that runs longer than the original is a known hard problem and the README mentions time-stretching via audiostretchy without documenting how badly it degrades for large mismatches. The requirements.txt approach with no pinned transitive dependencies means CUDA, PyTorch, Whisper, and Demucs version combinations will drift and break silently on fresh installs.