Phonon-2 and Phonon-2.5

Speech recognition using Fermion Research's five-value Parakeet TDT models. Phonon-2.5 supports 25 European languages; Phonon-2.5 English and Phonon-2 support English. All load directly without a Fermion runtime.

pip install 'mlx-audio[stt]'

from mlx_audio.stt import load

model = load("FermionResearch/Phonon-2.5")
result = model.generate("audio.wav", chunk_duration=30, overlap_duration=2)
print(result.text)

Use load("FermionResearch/Phonon-2.5-English") for English-only recognition. The first v2.5 load verifies and unpacks the release into a cached MLX checkpoint. Later loads reuse it. Requires the stt extra for archive decompression.

result.sentences contains sentences and timestamped token pieces. stream=True emits overlapping chunks from a recording; live-input streaming, translation, diarization, and hotwords are not supported.

Weights are CC-BY-4.0; adapted code is Apache-2.0. See the model README for input requirements and CLI usage.