enzolabs/parakeet-ultra-coreml

Parakeet Ultra — Core ML (greedy)

Core ML conversion of moondream/parakeet-ultra (revision 73175eb7aeb0d82f1e2a6b53b3aabc10a90bcd0b), Moondream's fine-tune of NVIDIA Parakeet TDT 0.6B v3: 25 European languages, token-and-duration transducer, word timings. Enzo uses it for dictation, meeting transcription and live transcripts on the Mac.

Attribution and licence

• Parakeet Ultra © Moondream, licensed CC BY 4.0.
• Parakeet TDT 0.6B v3 © NVIDIA, licensed CC BY 4.0.
• Core ML graphs from FluidInference/parakeet-tdt-0.6b-v3-coreml at revision 7dd20fe6b1797d35f5e3307e8b1732d9a178edfe (Apache-2.0 conversion).

This bundle is redistributed under CC BY 4.0 (see LICENSE).

What was changed

Only weights. Parakeet Ultra's Transformers ParakeetForTDT tensors were renamed to NeMo's layout and written into FluidAudio's pinned v3 graphs, which stay byte-identical; the encoder keeps that bundle's 6-bit palettes (refit to the new weights) and FP16 tensors elsewhere. The joint then drops the unused top-K outputs ("greedy" profile: token, probability and duration only). The voice-activity head in the source checkpoint is not included. The procedure was verified by converting base v3 the same way and reproducing every FP16 tensor of the reference bundle exactly. Recipe: tools/evaluation/asr-wer/ultra-coreml/convert_hf.py in the Enzo repository; receipts in conversion.json and optimization.json.

Files

Preprocessor, Encoder, Decoder and JointDecisionv3 as .mlpackage (compile on the destination Mac; a compiled cache is specific to its OS), parakeet_vocab.json (the v3 8,192-token vocabulary), and bundle.json with each file's SHA-256.

Measured (Apple M4, macOS 15.5, word error rate, Whisper English normaliser)

Corpus	This bundle	orukeet r3 (Core ML)	Parakeet TDT v2
LibriSpeech dev-clean, 342 clips	1.91%	1.89%	2.28%
LibriSpeech dev-other, 381 clips	3.61%	3.76%	3.35%
FLEURS, 8 languages, 128 clips	7.75%	8.54%	79.27%
AMI meeting utterances, 188 clips	7.26%	7.46%	—