Granary-v2 open speech model expands to 48 languages with open data and pipeline
rdesh26 · x · 2026-10-29
nithinraok explains the motivation behind Granary-v2: the same recording may need different outputs — verbatim transcripts, clean formatted text, injected context like names or technical terms, translation, or Q&A about the audio — so users should be able to ask for what they need. Key points:
- Languages expanded from 25 to 48
- Label quality improved via better ASR labeling, transcript refinement, and quality filtering
- Alongside the model, labels/instruct data and the full data pipeline are open
The attached meme (escalating Vince McMahon) captures the excitement: open weights < on HF < detailed technical report < open data AND pipeline.
More from Multimodal
- Magnific Move ships with a library of motion templates, tutorial thread wraps up — techhalla · 2026-10-29
- Character consistency trick: one base portrait, then swap only the mood word per prompt in Magnific One — techhalla · 2026-10-29
- Magnific Move drops: tutorial shows AI music video editing 'stopped being the hard part' — techhalla · 2026-10-29
- Reve publishes technical report on training infra that scaled to 10K GPUs with 10x fewer chips — cantrell · 2026-10-29
- Kling 4.0 video model lands on Cloudflare AI Gateway at $0.157/sec for 720p — ritakozlov · 2026-10-29
- AI-generated Star Wars-style game demo shows a flyable desert world with TIE fighters — jasonkneen · 2026-10-29