Kea-B 0.1 for Core ML
Kea-B 0.1 converted to run on a Mac: a Core ML encoder for the Neural Engine or the CPU, and the decoder as plain arrays for native code. It is the same model with the same weights. The results, the training notes and the model card are in that repository.
What is here
| Folder | Runs on | Encoder written for | Size |
|---|---|---|---|
macos26/ |
macOS 26 or later | Core ML, iOS 26 target | 138,084,560 bytes |
Builds for other systems will be added as folders beside it, each complete on its own.
Files
| File | What |
|---|---|
bundle.json |
The manifest: the SHA-256 of every other file, the sample rate and the input length. |
Encoder.mlmodelc/ |
The compiled Core ML encoder. It takes 15 seconds of 16 kHz audio. |
frontend.json, frontend.f32bin |
The front end's constants. |
decoder_joint.json, decoder_joint.f32bin |
The prediction and joint networks in FP32. |
vocabulary.json |
The token table. |
This repository holds weights only. Running them takes code that computes the features, calls the encoder and runs the TDT decoder, and that cuts longer audio into 15-second pieces. Sorello, Audeo Labs' dictation app for the Mac, is that code. These are the files it downloads, and each manifest names it.
Download and check
hf download audeolabs/kea-b-0.1-coreml --include "macos26/*" --local-dir kea-b-0.1-coreml
shasum -a 256 kea-b-0.1-coreml/macos26/bundle.json
| Manifest | SHA-256 |
|---|---|
macos26/bundle.json |
84d9aea06797cb2d08c77c1514edca0885f179a48a5d5e16ee72d47d1d4c870c |
It was converted with coremltools 9.0 from compat/export.safetensors in
audeolabs/kea-b-0.1, SHA-256
ec05526ed7aa7c742ee970c09724b1b8ccfaef6eb10f2ff1a11b031e9dfc4995.
Licence
The weights and documents are licensed CC-BY-4.0, like Kea-B 0.1
and the model it modifies, NVIDIA's
nvidia/parakeet-tdt-0.6b-v2. NOTICE lists the changes,
the training data and the tools, with their licences.