FermionResearch/Phonon-2 · Hugging Face
Short summary
Phonon-2 is a 164 MB speech recognition model that matches the accuracy of a much larger teacher model while being 15 times smaller. It averages 5.21% word error across seven English test sets and runs extremely fast—transcribing an hour of audio in 20 seconds on a MacBook Air.
- Achieves 5.21% average word error rate across Open ASR Leaderboard benchmarks, beating larger models including Whisper large-v3-turbo (6.58%) and matching its 2.5 GB teacher model's accuracy despite being 1/15th the size
- Extreme speed: processes audio 174x faster than realtime on M5 MacBook Air, 143x on Zen 5 CPU, and 6,680x on H100 GPU in batches
- Quantized to 2.1 bits per weight using five learned levels, enabling the tiny download size while maintaining teacher-model performance
Full summary
Phonon-2 is a compressed speech recognition model that achieves surprisingly good accuracy despite being only 164 MB — about 15 times smaller than its teacher model. It averages 5.21% word error across seven English test sets on the Open ASR Leaderboard, which beats several much larger models. The trick is aggressive quantization: each weight in the encoder sits at one of just five learned levels, using roughly 2.1 bits per weight instead of the standard 32. On a MacBook Air M5, it transcribes an hour of audio in 20 seconds (174x faster than realtime), and on an H100 GPU with batch processing it reaches 6,680x speedup.
The benchmarks show where it lands relative to competitors. Phonon-2 edges out NVIDIA's Parakeet TDT teacher model (2.5 GB) on some datasets like parliamentary speech and meetings, despite being 15 times smaller. It beats Parakeet Redux and ties Canary 180M Flash on average error rate, while staying well ahead of Whisper large-v3-turbo and other larger alternatives. The only model that beats it — Parakeet TDT itself — weighs 15 times more. Phonon-1, the previous version at 415 MB, had noticeably worse accuracy at 6.56% average error.
You can run this through several paths: the MLX library on Apple silicon, a command-line tool called `fermion transcribe`, or Docker containers for CPU and GPU setups. It's already shipping inside Detta, a dictation app for Mac. The code is open source (Apache 2.0 for the command line and repository, CC-BY-4.0 for the model itself), and it's based on NVIDIA's Parakeet model with modifications tracked in a notice file.