It is time for Steinberg to add MLX / Core ML support in Spectralayers, for Macs

SpectraLayers
1

Although, I’ve ejoyed using SL with my Windows PC for a long time. The experience I had running it on my first Mac (Macbook Pro M5), was incredibly frustrating. Things like unmixing take very long to complete since the app has no optimization for Apple Silicon at all.

I understand that this might not be an easy task, but a tool aimed at professionals cannot be out of the loop when it comes to technologies that speed up workflow. And considering that Intel Macs will be soon out of support there’s not reason why this move shouldn’t be done.

That’s my thought at least… Thanks!

2

Core ML is fully supported, and already used on modules where it’s possible to use it (such as Unmix Components, Reconstruct, Unmix Drums, Unmix Chorus, Unmix Crowd, Voice DeClip, DeBleed, DeClip, DeReverb).

But it’s actually a very limited technology when it comes to audio processing because it has a hard buffer size limitation (16384 samples in any dimension, where several advanced audio processing such as music unmixing usually needs much more than that).

Potential technology replacements are being considered to move on from CoreML to MPS, such as an ONNX derivate using MPS that was just started recently (still experimental), and ExecuTorch, a new inference framework by PyTorch with support for MPS that just reached 1.x recently as well.

So it looks like this should eventually be doable, it’s rather a matter of when these technologies will be able to fully support all AI models architecture. They are tested at each major SL release cycle though.

3

Thank you for your reply Robin.

Can I ask one more question please… Is CoreML supported in the version 13? Because in SL12 processing times were very dissapointing (like a CPU - unless there is a setting to enable CoreML, which I haven’t seen)

4

Core ML support has been enabled for a couple years now. What module that I listed above did you find slow in SL12 ?

5

That’s odd.. Anyway! The two I’ve used so far are “Unmix Song” and “Unmix Soundtrack”. Both seem to run slower even than the CPU mode in Windows. I haven’t found any relative setting to switch between CPU and CoreML (something similar to switching from CPU/GPU in the Windows version). Am I missing something?

6

There’s no setting, it’s always activated for all modules that supports it.
However as I said in my first message, Unmix Song cannot, by nature, be accelerated by CoreML. Same goes for Unmix Soundtrack, for the same reasons: