cyron/fluid-speech-pro-coreml

Fluid Speech Pro (Core ML, 16-bit)

An unofficial Core ML conversion of Fluid Speech Pro by FluidVoice, made for the VoiceiQ Mac app. Weights and compute are 16-bit floats; nothing is quantized. This is not a FluidVoice release.

The package has the same inputs, outputs and vocabulary as parakeet-tdt-0.6b-v3-coreml and fluid-speech-mini-coreml (16 kHz audio, 15-second windows, 8,192-piece vocabulary, blank = 8192), so FluidAudio loads it as a Parakeet v3 model. The joint is stored as JointDecisionv3.mlmodelc, the name FluidAudio 0.17.7 looks for. Clean mode is not included.

The conversion adds Fluid Speech Pro's extra encoder block (fv_insert, after subsampling) to the traced encoder, so its output matches the NeMo model.

Source and changelog

Converted from FluidVoice-Labs/fluid-speech-pro at revision 36481f6c9d353a301df1524dd0124eaf9212713e.

• 2026-11-07: reconverted from upstream revision 36481f6, in which FluidVoice updated the decoder and joint weights. The encoder is unchanged.
• 2026-11-02: first conversion, from upstream revision 53f490b.

License and attribution

CC BY 4.0, the license of the original model.

Fluid Speech Pro by FluidVoice. Built on NVIDIA Parakeet TDT 0.6B v3 and Moondream Parakeet Ultra. Converted to Core ML for VoiceiQ; the conversion changes the file format and numeric precision only.