Speech models

Phonon-2 is the current speech recognition model, run as phonon-2 with fermion transcribe, listen and serve, or through the phonon command. It transcribes English from 16 kHz mono audio, with each encoder weight stored in about 2.1 bits. The aliases phonon-2, phonon2, phonon, speech, stt, and asr all resolve to it. Figures are word error rate on the full test sets with the Open ASR Leaderboard scorer; lower is better.

It transcribes a single stream at 174 times realtime on an M5 MacBook Air through MLX; on the CPU engine, 143 times on eight Zen 5 cores (16 vCPU), 52 times on Google Axion, and 40 times on the CPU of an M5 MacBook Air. The weights are released under CC-BY-4.0.

The Phonon-2 model page carries the full comparison, and Introducing Phonon-2 the release report.

The first Phonon release, run as phonon-1 with fermion transcribe, listen and serve, or through the phonon-1 command. It transcribes English from 16 kHz mono audio and was trained at 2.4 bits per weight from the start. Benchmark figures are word error rate; lower is better.

The smallest speech model. It trades accuracy for a smaller download and is selected with micro.