Phonon-2

Phonon-2 is an open automatic speech recognition (ASR) model for English, released by Fermion Research on 29 September 2026. In a 164 MB download (about 177 MB on disk) it averages a 5.21% word error rate (WER) across the seven English test sets of the Open ASR Leaderboard.

The model is a low-bit quantized version of nvidia/parakeet-tdt-0.6b-v3 (relation: quantized). Its encoder stores each weight as one of five learned levels at about 2.1 bits per weight. As a result Phonon-2 holds the accuracy of its 2.5 GB full-precision teacher set for set, and beats it on some sets (meetings, parliamentary speech), from a file roughly 15 times smaller.

Phonon-2 runs locally: on Apple silicon (MLX and Core ML / Neural Engine), as well as on CPUs (Linux x86-64 and Arm, Windows) and NVIDIA GPUs. It transcribes an hour of audio in about 20 seconds on an M5 MacBook Air (174x realtime), at 143x on eight Zen 5 cores and at 6,680x on a single H100 in batches of 128. It is the model inside Detta, Fermion Research's dictation app for the Mac. The weights are released under CC-BY-4.0 and the package code under Apache 2.0.