Phonon-2: a 164 MB transcription model outperforms rivals four times its size

Fermion Research has released an English speech recognition model that takes up 164 MB, has an error rate of 5.21 per cent per word and transcribes an hour of audio in around 20 seconds on a MacBook Air.

The Fermion Research laboratory has put Phonon-2 online, a speech recognition model for English that weighs only 164 MB to download. To reach this size, the team distilled NVIDIA's Parakeet TDT 0.6B v3 model (2.5 GB) and then quantized the encoder to the extreme: each weight takes only one of five learned values, or about 2.1 bits per parameter.

Scores at the level of much larger models

Of the seven English datasets in the Open ASR Leaderboard, Phonon-2 has an average error rate of 5.21%, compared to 5.69% for Parakeet Redux (178 MB) and Canary 180M Flash (737 MB). It drops to 1.72% on LibriSpeech Clean but rises to 9.37% on AMI's recorded meetings, the most difficult corpus. On VoxPopuli's parliamentary debates, it even slightly exceeds the model from which it is based.

In terms of speed, Fermion claims 174 times the real-time on the GPU of an M5 MacBook, 143 times on eight x86 cores on Linux and up to 3,614 times in 128-stream batch processing on an A100. The weights are released under a CC-BY-4.0 license, inherited from Parakeet, on Hugging Face and via PyPI; the model also powers Detta, the publisher's offline dictation application for Mac.