CoreML export for the open source turn detection voice model

March 9, 2025

CoreML export for the open source turn detection voice model.

PR here^[1]

Hugging Face upload here^[2]

125ms inference on an M4.

Not quantized, yet. There are issues with the export when quantizing that I think are probably bugs in the export process rather than precision issues.

Do you want to review the PR, @maxxrubin_ ?

https://t.co/Jy74cLR8kB

Max@maxxrubin_

@kwindla @ikevictor MLX isn't the right move here, but CoreML is (this way you can utilize the neural engine. 1/4 the power draw of running on gpu
I ran a performance report, check it out :)

(I'm unfamiliar with the framework and how to use it, was my first attempt at converting a model so unsure

@maxxrubin_ Also, @maxxrubin_ you may have a better way to do this export!

I wrote a script based on this guide^[3]