mlsubgen: Fully Local Subtitles in 45 Languages With Dual ASR and LLM Reconciliation
Dodgy_Past · reddit · 2026-10-03
A developer open-sourced mlsubgen, a local-only tool generating subtitles in 45 languages for your video library (Linux + NVIDIA).
- Unlike typical Whisper wrappers: per-segment language detection for mixed-language content; two speech recognizers run on everything with a local LLM reconciling them; prefers existing human work — embedded tracks and even OCR of bitmap PGS tracks from Blu-ray remuxes — before listening to audio.
- Every feature has measured accuracy in the README rather than claims.
- Runs on a 24 GB RTX A5000; profiles for 16/12/8 GB cards, 16–32 GB RAM, 30–65 GB of models.
- Accuracy over speed: roughly real-time per target language when listening; fast when subtitles already exist.
- Origin: Thai subtitles for Shin-chan for friends in Thailand.
More from coding & agent
- VoidZero founders on TypeScript.fm: type-aware linting goes stable, Cloudflare deal and Vite's future — cnakazawa · 2026-10-03
- The default AI coding stack: OpenAI+Anthropic models on Stripe, Supabase, Vercel infra — alexmacgregor__ · 2026-10-03
- Reddit MCP Server gives AI assistants key-free access to Reddit search, posts and user data — modelcontextprotocol · 2026-10-03
- Frenchie MCP connector adds OCR, transcription, file extraction and image gen to agents — modelcontextprotocol · 2026-10-03
- Stripe's upcoming agentic identity API may let trusted agents claim free trials for users — jeff_weinstein · 2026-10-03
- repligate demos memory system spanning 72.9M tokens, letting a model retrieve raw past experiences — repligate · 2026-10-03