Manan Gupta: 'Introducing Phonon-2' (164 MB open ASR model)
Manan Gupta @yoitsmanan · x · 2026-09-29 · ★★ · archived
Launch of Phonon-2, a 164 MB open-weights English speech-recognition model (~37k views).
Summary
Announces Phonon-2 from Fermion Research: a 164 MB, CC-BY-4.0 speech-recognition model claimed to beat Whisper large on average accuracy and to transcribe an hour of audio in ~20 s on a MacBook Air. The model card shows it is a ~2-bit compression of NVIDIA Parakeet TDT 0.6B v3 (model file: phonon-2).
Archived text
Introducing Phonon-2: a new standard in speech recognition per byte.
At just a tiny 164MB download, it is more accurate on average than OpenAI's Whisper large ( a model 10X its size)
Transcribe an hour of audio in just 20 seconds on a MacBook Air!
Open weights, CC-BY-4.0, today.
@fermion_ai
Media: https://pbs.twimg.com/media/HTZd-dlasAAzw60.jpg?name=orig
views 36781 · likes 1075 · reposts 73 · replies 54 (at fetch time)
Archived 2026-09-30 via fxtwitter (unofficial).
All posts · id: 2026-09-29-yoitsmanan-phonon-2