Kayetic/Phonon-2.5-English-CoreML

Phonon 2.5 English, Core ML (static Neural Engine buckets)

FermionResearch/Phonon-2.5-English converted into the optimized Core ML graph of FermionResearch/Phonon-2-CoreML, for on-device English dictation on Apple silicon. It is the local Phonon model of Super, a macOS productivity app, and loads with phonon-coreml 1.1.2 unchanged (Transcriber(bundle:)).

• All learned encoder and decoder values are Phonon 2.5 English's; relative-key tables are regenerated from its projection weights. Encoder weights stay as exact five-value palettes (sparse g16, 6-bit), and the decoder's 6-bit tables are repacked exactly into the existing decoder format, with no new quantization.
• Encoder functions for 3, 5, 10, 15, 25 and 35 seconds share one weight blob; the 3, 25 and 35 s functions are static specializations of the 15 s graph, so short and medium recordings pad less.
• SHA256SUMS lists every file. Source checkpoint SHA-256: c57119c0f04756932acc1d60d90506346e3708d69f00e654084acc7fe507fe27.

Results on an Apple M5 Pro

Model	Mixed clips, median	Throughput	Public WER (360 clips, human references)
Phonon 2, Core ML	87.84 ms	463× real time	4.22%
Phonon 2.5, this conversion (15 s windows)	59.63 ms	508× real time	4.07%

Timing is inference only on 40 fixed clips (1–187 s) with the default runtime options (15 s windows past 34.95 s, 4 decoder threads, 2 encoder requests in flight). The WER difference is within the paired bootstrap uncertainty. The first load compiles all six encoder functions for the device, which takes a few minutes once; later loads take about 2 seconds.

Licence and attribution

The weights are distributed under CC-BY-4.0 (LICENSE-WEIGHTS-CC-BY-4.0.txt): Phonon 2.5 English by Fermion Research, derived from NVIDIA Parakeet TDT 0.6B v3. The graph derives from Fermion's Phonon-2-CoreML (Apache-2.0 code, LICENSE-CODE-Apache-2.0.txt). See NOTICE for the full attribution, training data and changes.