tylerbcrawford/subgeneratorr v3.0.0-beta.1

v3.0.0-beta.1 — Local Engine (Whisper) beta

on GitHub

🧩 v3.0.0 beta: fully-local transcription

This pre-release adds a pluggable ASR engine layer with an opt-in local Whisper engine — the whole pipeline (transcription, keyterms via Ollama, translation via Ollama) can now run offline, on your hardware, at $0, with nothing leaving your machine. Deepgram Nova-3 remains the default engine.

Try the beta

Fully-local stack (no API keys at all):

curl -LO https://raw.githubusercontent.com/tylerbcrawford/subgeneratorr/v3.0.0-beta.1/examples/docker-compose.local.example.yml
# edit MEDIA_PATH, then:
docker compose -f docker-compose.local.example.yml up -d

Or pin the opt-in images in your existing compose:

• ghcr.io/tylerbcrawford/subgeneratorr-web-local:3.0.0-beta.1
• ghcr.io/tylerbcrawford/subgeneratorr-worker-local:3.0.0-beta.1
• ghcr.io/tylerbcrawford/subgeneratorr-cli-local:3.0.0-beta.1

The small model is baked in (works offline from first boot); pick bigger models in the UI — the picker shows RAM/speed guidance. Measured on a low-power Intel N100: a 21-minute episode in ~6.5 minutes (0.3× runtime), ~89% word-level agreement with Nova-3.

Highlights

• Engine switch on the main screen — choose Cloud or Local per job; the UI hides options the selected engine can't do (via the new /api/capabilities).
• Native SRT writer with readability-focused cues (max 2×42 chars, ≤7 s, splits at sentence ends and silence gaps) + VAD anti-hallucination tuning.
• Zero-cloud deploys boot keyless: DEEPGRAM_API_KEY is optional; without it the Cloud engine is greyed out and API calls follow ASR_ENGINE.
• Live progress for local jobs — real per-file percentage while Whisper transcribes.
• Changed (affects cloud users): subtitles no longer embed [speaker N] tags by default — cues read as on-screen dialogue. Re-enable with the Speaker labels in subtitles toggle (web) or SPEAKER_LABELS=1 (CLI). Diarization itself is unchanged and still powers speaker-mapped transcripts.

Existing users are unaffected: :latest still points at v2.4.x and the lean images are byte-equivalent to before.

Full details in the CHANGELOG.

Feedback → #4