parakeet-ultra-coreml

Parakeet Ultra in CoreML format: a candidate model for Diction's on-device speech-to-text on Apple hardware. Same 25 languages, tokenizer and component contract as Parakeet TDT v3.

Credit

• NVIDIA trained the original Parakeet TDT 0.6B v3.
• moondream post-trained it into Parakeet Ultra.
• FluidInference built the CoreML conversion this is based on (pinned at revision 95eaa59a39d4; int8 encoder, fp16 decoder and joint, v3 preprocessor and vocabulary). All model files are byte-identical to that revision; only this README differs.

We publish it under our own namespace so Diction's runtime downloads, via FluidAudio, pin to a build we've tested instead of tracking FluidInference's default branch directly.

Licensed CC-BY-4.0, the same as all upstreams. We add no restrictions.

Use with Diction

Point FluidAudio (0.17.3 or later, AsrModelVersion.ultra) at it with ModelRegistry.repoOverrides:

ModelRegistry.repoOverrides = ["FluidInference/parakeet-ultra-coreml": "DictionLabs/parakeet-ultra-coreml"]