DubTitlerr

Dubtitles for anime: subtitles that say what the English dub actually says.

Most anime already ships an English subtitle track, but it is a translation of the Japanese audio — different wording, sometimes different meaning. Watching a dub with it on gives you two scripts at once. DubTitlerr transcribes the dub audio itself, corrects the proper nouns, and writes the result back into the file as a subtitle track your player picks up.

It runs as one container, per episode, incrementally.

This is a beta. It is validated end to end on One Pace, Sword Art Online and Digimon. Other shows work and are also where the bugs are found. Version stamps may be invalidated without notice: pin your image tag. If you have pulled a newer image but are not ready to re-transcribe, stop the container until you are.

Start here

I want to…	Go to
Get it running on one show, start to finish	Your first show
Solve one specific problem	How-to guides
Look up a variable, a filename, a schema	Reference
Understand why it behaves the way it does	Why it works this way
Pick a transcription model for your GPU	Choosing an ASR model
Point an AI agent at the project	Agent guide

What it does, per episode

Transcribe the English dub audio with Whisper, then reflow it into readable cards — sentence-split, at most two lines, timed to the spoken onset.
Correct names against a per-show glossary that is mined from your own files and checked against the show's wiki.
Repair the low-confidence and name-suspect lines with a small language model, local or a hosted API, anchored on the existing subtitle track where you have one.
Gate the result — drop hallucinated music and silence lines, collapse runaway repetition, and flag what it is unsure about for you to read.
Merge the on-screen signs and song lyrics back in.
Mux the finished track into the file with its original fonts, so signs render in their real typeface.

Then it stops and waits for you. The lines it was unsure about are queued at http://localhost:8842, and a verdict you give there applies to that line everywhere in the show.

What it does not do

It does not translate. It transcribes English audio.
It does not proofread. The review queue holds the lines the pipeline doubted — the ones it was confident about were never shown to anyone, and some are wrong.
It does not re-mux an episode because you saved a verdict. That is a separate button, on purpose.

Requirements

An NVIDIA card with CUDA for transcription. 6 GB works; see choosing a quantisation.
Docker with the NVIDIA container runtime.
An OpenAI-compatible endpoint for repair: a small GGUF model you serve yourself, or a hosted API with a key (see Repair with a hosted model).
A library with English dub audio, laid out as Show Name (Year)/Season 01/…S01E01….

Image: ghcr.io/xenarathon/dubtitlerr

Roadmap

Arr API and poster library view - connect directly to Sonarr and Radarr APIs to present a poster grid with status bars, replacing filesystem walks and letting you set a show's wiki URL before its glossary exists.
Run a show again from the Runs page - a show that already ran is skipped. Queueing it again will ask whether you want to run it again or cancel, so you don't redo one by mistake. Run again redoes only that show and season. This helps after a new version changes the output.
Season ranges in run requests - a request on the Runs page targets one season or the whole show. It should also take a range of seasons, so queueing a show's early arcs is one request instead of one per season.
Remove a queued request - queued requests can be added but not withdrawn yet. The Runs page should let you remove one before it starts.
What the current run is doing - a status section on the Runs page showing the show, episode and stage being worked on right now.
Wiki scope for more shows - the scoped wiki lookup is built and tested live against Naruto, Pokémon and two one-season shows. What is left is trying it on each show's real library and writing the wiki_scope block for the shows people actually run.
Reference lines from a published subtitle file - repair anchors on the English subtitle track embedded in the video. Dub re-cuts made from another release carry none, so repair works from the glossary alone. The idea is to match each dub line, by its words, to a published English subtitle file for the same scene, allowing the timing offset to change partway through an episode. A first blind test on three episodes of one arc preferred the referenced output over the glossary-only output on 34 of 65 contested lines, against 7 the other way. It needs a wider test and a spot-check before it is built.
Community glossary repo - XenaRathon/DubTitlerr-glossaries is the shared repository of per-show glossaries. Planned for v1.0.0 or a later 1.x release: downloading glossaries from it and uploading yours back, from inside the app. Downloads will never overwrite your own fixes or a wiki you set. Today it is a git clone into your GLOSSARY_DIR and a git pull to update, which decisions.py already documents as the intended flow.

This repository's glossaries/ directory is a snapshot, used by the tests and by the examples in these docs. The pipeline reads whatever is mounted at GLOSSARY_DIR and never the copy in its own source tree, so the shared repository is the one to install.

Glossaries — community per-show name dictionaries, pulled into GLOSSARY_DIR.
Subtitles — finished dubtitle files for episodes that have had a human review pass, published as review completes.