Phonon-1¶

Phonon-1 is Fermion Research's English speech-recognition family based on the Qwen3-ASR 0.6B architecture. Its decoder is trained with a five-value weight representation and runs through native MLX quantized matrix multiplication.

Available models¶

Model	Download	LibriSpeech clean / other	Notes
`FermionResearch/Phonon-1-Micro`	285 MB	3.002 / 6.511	Smallest build
`FermionResearch/Phonon-1`	415 MB	2.640 / 5.699	Default accuracy/size balance
`FermionResearch/Phonon-1-Big`	581 MB	2.667 / 5.722	Full-precision audio tower and lowest latency

Word-error rates are the provider's published full-test-set measurements.

Installation¶

Install the STT extra so the first-run archive materializer is available:

pip install 'mlx-audio[stt]'

The first load downloads the selected Hugging Face archive, verifies its published SHA-256, reconstructs the original safetensors, and verifies every materialized file against the archive manifest. Later loads reuse the local cache. Set MLX_AUDIO_CACHE_DIR to relocate that cache.

Usage¶

PythonCLI

from mlx_audio.stt import load

model = load("FermionResearch/Phonon-1")
result = model.generate(
    "audio.wav",
    language="English",
    temperature=0.0,
    repetition_penalty=1.05,
    repetition_context_size=96,
)
print(result.text)

mlx_audio.stt.generate \
  --model FermionResearch/Phonon-1 \
  --audio audio.wav \
  --output-path transcript \
  --language English \
  --gen-kwargs '{"repetition_penalty": 1.05, "repetition_context_size": 96}'

Use stream=True in Python or --stream in the CLI for token streaming.

Boundaries¶

The released family is English-only. It returns chunk-level segments, not word timestamps, forced alignment, speaker diarization, or translation.