mlx-community
/

Phonon-2

@@ -14,17 +14,12 @@ tags:
 - apple-silicon
 ---
-# Phonon-2 for MLX Audio
-Preconverted MLX weights for [Fermion Research's Phonon-2](https://huggingface.co/FermionResearch/Phonon-2), an English speech recognition model derived from [NVIDIA Parakeet TDT 0.6B v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3).
-The five-value encoder is represented by two native MLX 2-bit planes with the original fp16 row palettes. Remaining tables use bf16, matching Fermion's packed MLX runtime. This is a custom five-value representation, rather than a standard uniformly quantized 2-bit model. The Safetensors file is approximately 350 MB (334 MiB).
-## Compatibility
-Requires an mlx-audio build containing the revised Phonon-2 integration prepared for [PR #1030](https://github.com/Blaizzy/mlx-audio/pull/1030). The validated revision is currently local and has not been pushed or released. Installing an older mlx-audio release alone will not add this model implementation. No remote Python code is included in this repository.
-Once that implementation is installed, load these weights directly; no local conversion or Fermion runtime is needed:
 ```python
 from mlx_audio.stt import load
@@ -32,22 +27,11 @@ from mlx_audio.stt import load
 model = load("mlx-community/Phonon-2")
 result = model.generate("audio.wav", chunk_duration=30, overlap_duration=2)
 print(result.text)
-for sentence in result.sentences:
-    for token in sentence.tokens:
-        print(token.start, token.end, token.text)
 ```
-Paths are resampled by mlx-audio. Array inputs must be mono 16 kHz waveforms. The model supplies token pieces and sentence timestamps; pieces may need joining into words. Short nonempty inputs are padded to 320 samples. `stream=True` emits overlapping chunk results from a recording; it does not provide live-input streaming. English transcription only; translation, hotword biasing, and diarization are not implemented.
-## Conversion and validation
-Converted using NumPy/MLX from official source revision `7e153e4054c0a10db6d47d6fff5c062d4c77c154`. Source archive SHA-256: `98125795b6dda72f5c6eee9ba33d19815df65dcb18b50a357bf9f73c9935309e`. The archive and extracted files were verified before conversion. Packed codes and row palettes are preserved, with no additional encoder weight quantization.
-Validated on Apple Silicon with MLX 0.32.4 against fermion-research 0.2.11 in packed mode (`FERMION_P2_FAST=tdt16`). Six clips totaling 178 seconds exactly matched text, timed tokens, and overlapping chunked/streaming results. Distinct-input, duplicate-input, and different-valid-length batched-mel cases also matched the official runtime. Batch composition affected a token timestamp relative to individual processing in both runtimes; equivalent batched execution matched exactly. These limited checks establish parity on the tested inputs, not an accuracy or performance improvement. Other MLX versions and hardware were not revalidated for this revision.
-## Attribution and licenses
-Phonon-2: Fermion Research. Base Parakeet architecture and original model: NVIDIA Corporation. Weights are licensed under [CC-BY-4.0](LICENSE-WEIGHTS-CC-BY-4.0.txt). This conversion changes storage layout and uses bf16 dense tables as described above. Retain attribution and license information when redistributing.
-The upstream [Apache-2.0 code license](LICENSE-CODE-Apache-2.0.txt) is retained for provenance; this repository distributes weights and configuration, not the Fermion runtime. See the upstream repositories and the included NOTICE for attribution.

 - apple-silicon
 ---
+# Phonon-2 for MLX
+English speech recognition with [Fermion Research's Phonon-2](https://huggingface.co/FermionResearch/Phonon-2),
+derived from [NVIDIA Parakeet TDT 0.6B v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3).
+Preconverted five-value weights; approximately 350 MB. No converter or Fermion runtime needed.
+Requires mlx-audio with Phonon-2 support.
 ```python
 from mlx_audio.stt import load
 model = load("mlx-community/Phonon-2")
 result = model.generate("audio.wav", chunk_duration=30, overlap_duration=2)
 print(result.text)
 ```
+Supports token timestamps and overlapping chunk output, not live-input streaming.
+Validated against the official packed runtime on six clips using MLX 0.32.4.
+Weights: [CC-BY-4.0](LICENSE-WEIGHTS-CC-BY-4.0.txt). Copyright Fermion Research and NVIDIA.
+Encoder codes are repacked into two MLX 2-bit planes; row palettes remain fp16 and dense tables use bf16.
+See [NOTICE](NOTICE) for attribution.