Dub Studio

Professional dubbing is slow and expensive, so most videos are only ever available in one language. Fully automatic machine translation makes mistakes that matter — changed numbers, wrong Bible references, mistranslated names, or jokes that don’t land — and video processing needs a strong GPU, which a cheap shared web host doesn’t have.

A split architecture: a lightweight PHP website on shared hosting handles login, chunked uploads, the job list, the review editor and downloads, while a Windows PC with an NVIDIA GPU does the heavy work. n8n bridges the two — forwarding start, status and cancel calls and holding the OpenAI key — so the PC is never exposed to the internet and no secrets live on the worker. Video files never pass through n8n: the worker pulls uploads from the website and pushes the finished video straight back. The translation step is its own module with a quality gate that checks numbers, glossary terms, Scripture references and subtitle match line by line; critical problems block the job until a named reviewer approves or edits the line, and only failing lines are retranslated with a stronger model to keep costs low.

Around 7,500 lines across the website, worker and tests. The translation stage is covered by a fast offline suite that includes 14 required real-world cases — numbers, Bible quotes, names, humour and very short lines — with scripted and recorded model answers, so it runs in seconds and gives the same result every time.

Target languages

15

Translation-stage tests

200+

Max upload, in 5 MB resumable chunks

4 GB

• 01Three ways in: a video link, a full playlist (each video queued as its own job), or a file upload up to 4 GB
• 0215 target languages, including Spanish, Portuguese, French, Tagalog, Cebuano and Indonesian, with male or female voices
• 03Content profiles — general, sermon, teaching, business, science, comedy — that set tone, Scripture and number handling, and humour
• 04Human-in-the-loop review: pause after translation and fix or approve every line next to the English
• 05Manual mode: download English subtitles plus a ready-made ChatGPT prompt, then upload your own translation
• 06Background audio options: keep the music without the original voice, dubbed voice only, or original quietly underneath
• 07Live progress, cancel and resume, and an in-browser player for finished videos
• 08Jobs survive the GPU PC going offline, stop after three crashes, and keep finished stages on resume
• 09Password login with CSRF protection, per-job worker keys, and automatic cleanup of old files

1. 01

   Step 1: Download

   yt-dlp fetches single videos or whole playlists; uploaded files are pulled from the website by the worker.

2. 02

   Step 2: Transcribe

   faster-whisper (large-v3 on CUDA) produces English text with word-level timing on the local GPU.

3. 03

   Step 3: Translate

   OpenAI through n8n, with Ollama and DeepL as fallbacks, behind a versioned quality gate and optional human review.

4. 04

   Step 4: Voice

   Microsoft Edge TTS generates the dubbed voice, with Piper for offline voices; Demucs removes the original voice to keep the music.

5. 05

   Step 5: Mix

   FFmpeg fits timing, mixes the tracks, normalizes loudness to -16 LUFS and uploads the finished video back to the website.