Phonon-2
Copyright 2026 Fermion Research

This product includes software and models developed by Fermion Research.

================================================================================
BASE MODEL
================================================================================

This model is derived from NVIDIA's Parakeet TDT 0.6B v3
(nvidia/parakeet-tdt-0.6b-v3), (c) NVIDIA Corporation, licensed under the
Creative Commons Attribution 4.0 International licence (CC-BY-4.0).

    https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
    https://creativecommons.org/licenses/by/4.0/

Changes: the encoder's linear weights were re-trained and quantised to a
five-value alphabet by Fermion Research; the remaining parameters were
quantised to 6-bit integer tables. The FastConformer encoder architecture,
the TDT decoder, the tokenizer and the feature-extractor configuration derive
from that model and remain under CC-BY-4.0.

These weights are therefore distributed under CC-BY-4.0. The Fermion code in
this repository remains under the Apache License, Version 2.0.

================================================================================
TRAINING DATA
================================================================================

Training audio includes CHiME-6, distributed under the Creative Commons
Attribution-ShareAlike 4.0 International licence (CC-BY-SA-4.0)

    https://chimechallenge.github.io/chime6/
    https://creativecommons.org/licenses/by-sa/4.0/

and SPGISpeech, used under Kensho's public terms

    https://datasets.kensho.com/datasets/spgispeech

================================================================================
RUNTIME DEPENDENCIES
================================================================================

MLX ......................... MIT License ......... https://github.com/ml-explore/mlx
mlx-audio ................... MIT License ......... https://github.com/Blaizzy/mlx-audio
    The Parakeet model graph (feature extraction, FastConformer encoder,
    TDT decoder and greedy search) is mlx-audio's implementation.
NumPy ....................... BSD-3-Clause ........ https://numpy.org
soundfile ................... BSD-3-Clause ........ https://github.com/bastibe/python-soundfile
    soundfile links libsndfile (LGPL-2.1), dynamically and unmodified, as
    distributed in the standard soundfile wheel.

================================================================================
REFERENCE RUNTIME (optional)
================================================================================

PyTorch ..................... BSD-3-Clause ........ https://pytorch.org
transformers ................ Apache-2.0 .......... https://github.com/huggingface/transformers

================================================================================
COREML CONVERSION (Vowen)
================================================================================

This directory is a CoreML conversion of FermionResearch/Phonon-2
(https://huggingface.co/FermionResearch/Phonon-2, container sha256
4b6bfa3a12cc3c4e0a54f2ab3ec4ca7a842b09e5c7ecfc8e7ca0ac6cc8c11468) for use with
FluidAudio on macOS. Changes: the weights were expanded exactly from the
five-value / int6 container into the stock Parakeet TDT 0.6B v3 graph and
exported to CoreML; encoder weights are stored as lossless per-row lookup
tables. Preprocessor.mlmodelc and parakeet_vocab.json are FluidInference's
Parakeet TDT 0.6B v3 CoreML files (CC-BY-4.0), which Phonon-2 shares
unchanged. Weights remain under CC-BY-4.0.
