Phonon-2.5 Core ML โ€” GPU encoder

Core ML conversion of Fermion Research Phonon-2.5, a 25-language model derived from NVIDIA Parakeet TDT 0.6B v3. The complete directory is approximately 290 MB; this differs from the 156 MB compressed source download.

Runtime

Requires macOS 15 / iOS 18 or later for the exact five-value weight representation. Tested on Apple M5 / macOS 26.6.2 with unmodified FluidAudio 0.17.3 and 0.17.7. The encoder is the dense LUT3 GPU variant, not the sparse ANE variant. Load with:

let models = try AsrModels.loadLocal(from: directory, version: .v3,
                                     encoderComputeUnits: .cpuAndGPU)
let manager = AsrManager(config: .default)
try await manager.loadModels(models)

.v3 selects the compatible tensor/decoder contract; all three prediction components use Phonon-2.5 weights. Decoder and joint retain FluidAudio's default CPU+ANE placement. Only the standard shared preprocessor is reused.

Validation scope

The GPU encoder passes the retained single-window comparison against fp32 NeMo (cosine 0.999997, NRMSE 0.00149; thresholds >0.999 and <0.05). On 83 public clips across English, French, German and Spanish it matched 72 normalized fp32 reference transcripts, with small WER changes in both directions. Silence and a 44-second repeated-utterance fixture were checked. These are smoke results, not a full corpus or leaderboard validation, and other chips have not yet been tested.

The sparse ANE encoder is not distributed here: it fails the strict activation comparison, and upgrading FluidAudio from 0.17.3 to 0.17.7 did not resolve that. Decoder/joint fp16 outputs are not claimed to be numerically identical to fp32. Conversion scripts, full results and limitations: Mรถbius draft PR #113.

Provenance and license

Source checkpoint revision: 2673618ad8636c8dfd20fe57e94d1f67f75bdeaa, model.fermion SHA256 17cb62d08e751a41819eb779db089aa7ac4eb2bab433ba5c2991d8cbd3cfefb9. Vocabulary comes from this checkpoint. Preprocessor: FluidInference/parakeet-tdt-0.6b-v3-coreml revision 7dd20fe6b1797d35f5e3307e8b1732d9a178edfe.

Weights remain CC-BY-4.0, with the original license and NOTICE preserved and conversion modifications attributed. Conversion scripts are Apache-2.0. See checksums.json for exact distributed artifact hashes. This repository is maintained by Muesli-HQ and does not imply endorsement by Fermion Research or FluidInference.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Muesli-HQ/phonon-2.5-coreml

Finetuned
(2)
this model