Open-source Phonon-2.5: 156MB speech model beats 20x-larger Whisper large-v3
julianweisser · x · 2026-10-24
Fermion AI open-sourced Phonon-2.5, claiming a new accuracy-to-size Pareto frontier for multilingual speech recognition. The 156MB model scores 101.8% of its full-precision base's average word accuracy across 25 languages — beating OpenAI's Whisper large-v3 despite being 20x smaller.
The team compressed a 2.5GB model to 6.2% of its original download size while improving accuracy. It transcribes an hour of multilingual audio in 20 seconds on a MacBook Air, runs fully locally, and ships under CC-BY-4.0. It's already rolling out in Detta, a Mac dictation app supporting 25 languages.
More from Multimodal
- Seedance 2.5 demo: a spinning chest-cam shot that visualizes a doctor's burnout — LudovicCreator · 2026-10-24
- MOSS-Image-Edit-HQ lands on Hugging Face for hands-on image editing — Illustrious_Row_9971 · 2026-10-24
- Nano Banana 2.1 blind-tested against 3 rivals: text nails it, no model draws 8:17 on a clock — Hot-Wrangler-3692 · 2026-10-24
- FastVideo open-sources MiniMax FastH3 Omni: 12x real-time speedup on Blackwell — TinfoilTricorn · 2026-10-24
- Peter Gostev demos 3D generation with Kimi K3.1 Pro — arena · 2026-10-24
- Matt Wolfe's Weekly AI Roundup: Claude Duet, GPT-6 Update, and 22 More Stories — Matt Wolfe · 2026-10-24