Subtitle generation for Plex, Jellyfin, and Emby — cloud or fully local. Transcribe with Deepgram Nova-3 or an offline Whisper engine on your own hardware, then let Claude, GPT, Gemini, or a local Ollama model handle translation into 54 languages.

Built with Claude Code. I set the requirements and the architecture, made the design decisions, and run it in production on my media server, with real users. The agent wrote most of the code under my direction.

Subgeneratorr generates language-tagged SRT subtitles for the media Bazarr can't find — obscure shows, older episodes, anything without community subs. It runs as a Docker container with a Web UI and CLI, uses keyterm prompting so proper nouns come out right, and can translate every subtitle it makes into 54 more languages with the original timing preserved.

I created it to fill the hundreds of missing-subtitle gaps in my own media library.

Since v3, transcription is a per-job switch right on the main screen:

Every AI step after transcription (keyterms, translation) is provider-agnostic too: Claude, GPT, Gemini, or a local Ollama model — pair Local Whisper with Ollama and the entire pipeline runs offline at $0, with no vendor lock-in anywhere.

Free and open-source. Not affiliated with Deepgram, Anthropic, OpenAI, or any other provider.

• 🎯 Deepgram Nova-3 speech-to-text — strong on fast dialogue, accents, and 45+ transcription languages (55+ counting regional variants; General + Medical models), including the five added in Deepgram's August 2026 expansion: Afrikaans, Armenian, Georgian, Nepali, and Punjabi
• 🏠 Local engine (v3) — opt-in -local images transcribe on your own hardware via faster-whisper: CPU-only, offline, $0, default model baked in
• 🔑 Keyterm prompting — feed character names, locations, and jargon to Nova-3 for up to ~90% better accuracy on proper nouns; generate them with one click via Claude, GPT, or Gemini
• 🌐 Subtitle translation into 54 languages — turn one transcription into many: an LLM translates the generated SRT with timing preserved and writes tagged sidecars Plex, Jellyfin, and Emby pick up; use Claude, GPT, Gemini, or a local Ollama model for free offline translation
• 🔍 Library-wide scan — find every file missing subtitles across your whole library, grouped by folder, with CSV export
• 🗣️ Speaker diarization — labeled, character-named transcripts
• 🌍 Multilingual transcription — auto-detect, multi code-switching, and regional variants across 55+ Nova-3 language codes
• 🛡️ Content control — redaction (PCI/PII), profanity filtering, find & replace, dictation
• 🧠 Audio intelligence — sentiment, summarization, topic/intent/entity detection (English)
• 🐳 Docker-based — Web UI and CLI, batch processing with parallel workers and real-time progress
• 📺 Media-server ready — language-tagged sidecars (.eng.srt, .spa.srt, .und.srt) auto-recognized by Plex, Jellyfin, and Emby

_{Browse your media library — the ☁️ Cloud / 🏠 Local engine switch lives right in the action bar.}

_{Scan the whole library for missing subtitles, grouped by folder.}

_{One-click AI keyterms — here it read the filename and pulled 22 terms from Tears of Steel on its own, for $0.004.}

_{Full control over Nova-3 settings when you want it; sensible defaults when you don't.}

_{Watch progress per file and confirm each subtitle as it lands.}

_{Turn one subtitle into many — pick target languages (54 available) and an LLM provider; timing is preserved.}

_{Demo library uses Creative Commons / public-domain titles (Blender open movies, Pioneer One, Night of the Living Dead) so nothing copyrighted appears in the shots.}

Pick your flavor:

• ☁️ Cloud (default) — Deepgram Nova-3: best accuracy, processes in ~1% of runtime, ~$0.0057/min with $200 free credits. Needs an API key.
• 🏠 Fully local — faster-whisper on your CPU: $0, offline, no account or key needed, near-real-time processing (larger image). Both engines stay selectable per job.

Requirements: Docker + Docker Compose · media files (MKV, MP4, AVI, MOV, MP3, WAV, FLAC, …) · a Deepgram API key for the cloud flavor only

git clone https://github.com/tylerbcrawford/subgeneratorr.git
cd subgeneratorr

cp .env.example .env
cp examples/docker-compose.example.yml docker-compose.yml          # ☁️ cloud (default)
# cp examples/docker-compose.local.example.yml docker-compose.yml  # 🏠 fully local

# In .env, set:
#   MEDIA_PATH=/path/to/your/media
#   DEEPGRAM_API_KEY=your_key_here   # ☁️ cloud flavor only

docker compose build
docker compose up -d          # Web UI at http://localhost:5000
# ...or run headless:
docker compose run --rm cli   # processes the whole MEDIA_PATH library

Security: DISABLE_AUTH=true is the default and is for local access only. For remote/production use, set DISABLE_AUTH=false and put an authenticating reverse proxy in front that forwards X-Auth-Request-Email or X-Forwarded-User. The CLI is synchronous and headless; the Web UI adds async batches, progress tracking, library scanning, and AI keyterm generation.

Speech models nail everyday words but mangle proper nouns — "Heisenberg" becomes "Heizenberg," "Los Pollos Hermanos" becomes gibberish. Keyterms tell Nova-3 exactly what to listen for, boosting recognition at decode time (not as post-processing). Up to ~90% accuracy improvement on prompted terms, ~20–50 terms per show.

Provide them manually as a CSV, or click Generate Keyterms in the Web UI: an LLM infers the show from the file path, researches it, and returns 20–50 names, locations, and jargon terms in 3–5 seconds for less than a penny. Gemini's free tier makes it effectively zero-cost, and the keyterms apply to every episode in the show automatically. See the model benchmarks and CSV format.

One transcription, many languages. After Nova-3 produces a subtitle, an LLM translates the cue text into the languages you pick and writes tagged sidecars (.spa.srt, .fre.srt, and so on) that Plex, Jellyfin, and Emby auto-detect. Timing is copied from the source frame for frame, so the translation never drifts, and the show's keyterms ride along as a glossary to keep names spelled consistently.

Open the Translate panel in the Web UI, pick your target languages, and choose a provider: Claude, GPT, Gemini, or a local Ollama model. Ollama runs on your own hardware over its OpenAI-compatible endpoint, so translation is free and fully offline (the cost estimate shows $0.00) — use a capable 3B+ model like qwen2.5:7b for reliable results. Existing translations are skipped unless you choose to overwrite.

The opt-in -local images swap Deepgram for faster-whisper on CPU. The small model is baked in, so it works offline from first boot; bigger models download into a cache volume, and the model picker shows RAM/speed guidance for each.

Start from examples/docker-compose.local.example.yml, which builds web, worker, and CLI from the local target. No DEEPGRAM_API_KEY needed — without one the UI greys out the Cloud engine and defaults to Local. For headless runs, the subgeneratorr-cli-local image uses the same engine via ASR_ENGINE=whisper. Add Ollama for keyterms and translation and no audio, text, or API key ever leaves your machine.

On the standard image it's the reverse: Deepgram stays the default and Local shows as unavailable. Either way the engine is a per-job switch in the Web UI, preselected from your server's ASR_ENGINE.

One change to cloud output in v3: subtitle cues no longer embed [speaker N] tags by default — re-enable them with the "Speaker labels in subtitles" toggle (or SPEAKER_LABELS=1 for the CLI).

Point it at a library of thousands of files and it tells you exactly what's missing. A two-phase scan checks sidecar files (seconds), then optionally probes embedded tracks with ffprobe (~50–100ms/file). Results come back grouped by directory, persist across page reloads, and export to CSV. A 4,662-file library scans in ~6 minutes with embedded detection on, or in seconds in sidecar-only mode.

1. Scan for missing subtitles (gear icon → "Find All Missing Subtitles")
2. Review results by directory to see which shows and seasons have gaps
3. Generate AI keyterms per show (one click, shared across all episodes)
4. Select files from the results → transcribe (keyterms auto-applied)
5. Resume anytime — scan results persist and processed files drop off the list

Two values are required; everything else has sensible defaults.

MEDIA_PATH examples: /home/you/media (Linux), /Users/you/Movies (macOS), C:/Users/You/Videos (Windows). On Linux, set PUID/PGID to match your user so generated files keep the right ownership. Full reference: Technical docs.

Nova-3 costs ~$0.0057/min of audio:

New Deepgram accounts get $200 in free credits — roughly 35,000 minutes (~585 hours).

• Technical docs — architecture, API endpoints, advanced config, speaker maps, AI model benchmarks
• Language support — the full Nova-3 transcription matrix (55+ with variants) and the 54 translation targets
• Roadmap — planned features

docker compose -f docker-compose.yml -f examples/docker-compose.hostnet.override.yml up -d --build

Issues and pull requests welcome — see CONTRIBUTING.md. For local tests you don't need Docker or API keys: make test.

MIT — see LICENSE.

Built on Deepgram (Nova-3, Python SDK, captions), faster-whisper by SYSTRAN (the local engine), and Claude Code.