matt-2012/parakeet-primeline-de-coreml

Parakeet TDT 0.6B v3 (German) — Core ML

A Core ML export of primeline/parakeet-primeline, laid out for FluidAudio's .v3 ASR path on Apple Silicon.

This repository is a redistribution with modifications of a CC BY 4.0 model. It contains no new training. See Changes made for the exact list.

Attribution chain

Layer	Source	License
Base architecture and weights	nvidia/parakeet-tdt-0.6b-v3	CC BY 4.0
German fine-tune (.nemo)	primeline/parakeet-primeline	CC BY 4.0
Core ML conversion tooling and vocabulary	FluidInference/parakeet-tdt-0.6b-v3-coreml	CC BY 4.0
This export	you are here	CC BY 4.0

NVIDIA's card states: "GOVERNING TERMS: Use of this model is governed by the CC-BY-4.0 license." No additional restrictions are imposed here, and none may be added downstream.

Changes made

Relative to primeline/parakeet-primeline:

1. Exported from the .nemo checkpoint to Core ML with FluidInference/mobius' convert-parakeet.py.
2. Patched the preprocessor input shape. The upstream converter emits a RangeDim(1, max_samples) flexible input; this export uses a fixed shape instead.
3. Compiled .mlpackage → .mlmodelc. Only the compiled form is shipped.
4. Renamed to FluidAudio's expectations. In particular parakeet_joint_decision_single_step → JointDecisionv3.mlmodelc — the 5-output v3 head. FluidAudio's .v3 path will not load the 3-output parakeet_joint_decision.
5. Kept the encoder at fp16. Encoder.mlmodelc/weights/weight.bin is 1,185,860,032 bytes (~1.19 GB). This is deliberately not the int8 encoder FluidInference publishes (445,187,200 bytes) — the fp16 encoder is what this export was evaluated with.
6. parakeet_vocab.json is byte-identical to FluidInference's (sha256:7ec60e05…98735); it is included so the bundle is self-contained.

Layout — the directory name is load-bearing

FluidAudio's AsrModels.load(from:) resolves its repository root as url.deletingLastPathComponent()/parakeet-tdt-0.6b-v3. These files must sit in a directory named exactly parakeet-tdt-0.6b-v3 or the load fails.

parakeet-tdt-0.6b-v3/
├── Preprocessor.mlmodelc/
├── Encoder.mlmodelc/
├── Decoder.mlmodelc/
├── JointDecisionv3.mlmodelc/
└── parakeet_vocab.json

17 files, 1,224,182,759 bytes total.

Download

hf download matt-2012/parakeet-primeline-de-coreml --local-dir ./parakeet-tdt-0.6b-v3

Set ModelHub.offlineMode = true before loading. Without it, a missing file is a cue for FluidAudio to fetch the stock English NVIDIA v3 into the same directory; its cache check is filename-existence only, with no checksums.

Verifying what you got

The SHA-256 of every file. Hugging Face's LFS oid is the SHA-256, so the table can be checked against the tree API without downloading:

File	Bytes	SHA-256
Decoder.mlmodelc/analytics/coremldata.bin	243	a2acb22d3901041daff2bc09434b228b0838f0333ebe1ec92edd039ca263125e
Decoder.mlmodelc/coremldata.bin	554	cb1a4041bce8cb5ce66bb340fbad2edae291383b34ecd960d110b9f42f87f6ad
Decoder.mlmodelc/model.mil	13,108	9d9beb4c19ce5ecd9cf656bf80534c6e2d7054c24108225e06d4f3e9cbe9471c
Decoder.mlmodelc/weights/weight.bin	23,604,992	d92ce68e4dcdb650b0446a390017e1a7470fd8c2b457585cd19c5909f8c3e5c5
Encoder.mlmodelc/analytics/coremldata.bin	243	e896501dcdcd5bc48a5f978ee585518fe1b9926df0b67092da44be007871402b
Encoder.mlmodelc/coremldata.bin	478	1a8d3cfa00be12fb007ed6f6dbdfa1e762b21a081ff3561987ef4c7f2b69818a
Encoder.mlmodelc/model.mil	892,025	fb84017adfb8a5b1019203c6eb7b0092bc9990fc964e82301bd51aa7c4398c4c
Encoder.mlmodelc/weights/weight.bin	1,185,860,032	2594e5225447ac4d951685406c1f8329a7bf61a9df0ed015f9c6eab0d864fa66
JointDecisionv3.mlmodelc/analytics/coremldata.bin	243	006c170b4cb28628fa835cb555bf0fd7db26eca8ee516927523af2a03246c430
JointDecisionv3.mlmodelc/coremldata.bin	591	c0e611380ef092c5804a6459e9a9f4ea0ffd41c2ca0f3f54cacb69d532b2dd4e
JointDecisionv3.mlmodelc/model.mil	11,775	be60732943389a047175111a83f8839f3eb39d4803adafa828a0871b2f39818d
JointDecisionv3.mlmodelc/weights/weight.bin	12,642,764	a0c74f8b291468b11dd40c812a57df6bb880c084e8971ade72a45374f294b2d2
Preprocessor.mlmodelc/analytics/coremldata.bin	243	900bf9c0b13759e22bcadb3dca62e7e716b586c642e07208a10fc8ccad91f25b
Preprocessor.mlmodelc/coremldata.bin	485	a126be17938b9039f1a2fb3e6af4a0c27f4052441c4cb16619347ef274c70912
Preprocessor.mlmodelc/model.mil	27,157	4da2091a9c8c252b5b83b8d9504827f0a5e21779471f42bb72f9318a54b4afe7
Preprocessor.mlmodelc/weights/weight.bin	976,704	f3c3d600f5fc6311ceff6406c10aeb4285a7f0fda4de6ac86b5bee3e76f80e32
parakeet_vocab.json	151,122	7ec60e05f1b24480736ec0eed40900f4626bce1fa9a60fd700ec7e2a59198735

Stability

main is never force-pushed. Consumers should pin a commit SHA; ueberblick pins the commit tagged v1 and verifies every file against the table above before installing it.

Citation

@misc{parakeet-tdt-0.6b-v3,
  title  = {Parakeet TDT 0.6B v3},
  author = {{NVIDIA}},
  year   = {2025},
  url    = {https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3}
}