TalkMate/fluid-speech-pro-coreml

Fluid Speech Pro (Core ML, 8-bit)

Fluid Speech Pro by FluidVoice, converted to Core ML by TalkMate for the Apple Neural Engine.

Same inputs, outputs and tokenizer as parakeet-tdt-0.6b-v3-coreml (Preprocessor, Encoder, Decoder, JointDecision), so FluidAudio loads it as a Parakeet Ultra-contract package.

Part	Weights
Encoder	int8 per channel, Pro's extra pre-Conformer block included
Decoder, JointDecision	fp16
Preprocessor	FluidInference's Parakeet TDT 0.6B v3 preprocessor

On accented English (EdAcc, 600 clips, M4): 10.5% WER, against 10.1% for the original .nemo. About 0.6 GB.

Licensed CC BY 4.0. Built on NVIDIA Parakeet TDT 0.6B v3 and Moondream Parakeet Ultra, both CC BY 4.0; the preprocessor is FluidInference's (CC BY 4.0).