FluidInference

/

parakeet-tdt-0.6b-v3-coreml

like 35

Follow

Fluid Inference Community 86

Summary

Request to add non-quantized CoreML models (Preprocessor + Encoder) as .mlpackage in the mlpackages/ directory.

Background

The mlpackages/ folder currently contains:

• MelEncoder.mlpackage — an int8 quantized fused encoder (commit: "int-8 model")
• Decoder.mlpackage and JointDecision.mlpackage — already available

The repository also includes:

• Preprocessor.mlmodelc
• Encoder.mlmodelc

These are non-quantized compiled models used by FluidAudio’s Swift SDK and achieve ~2.1% WER, but they are not available as .mlpackage.

Problem

I’m using coremltools from Python, which can load .mlpackage but cannot load .mlmodelc.
This makes it impossible to use the non-quantized encoder pipeline outside of Swift.

Request

Please add the following to mlpackages/:

• Preprocessor.mlpackage (non-quantized)
• Encoder.mlpackage (non-quantized)

Notes

• Decoder.mlpackage and JointDecision.mlpackage are already available and working
• The only missing pieces are the non-quantized preprocessor + encoder pipeline

Why This Helps

• Enables Python usage via coremltools
• Provides access to the higher-accuracy non-quantized models
• Aligns model availability across Swift and Python workflows

P.S. - This is opened by me, a human, not a OpenClaw clanker