Hi @Bruno de Souza Lino, SpectraLayers developer here - to answer your questions:
-The improved Hybrid Transformer Demucs implementation in SL10 is pretty fast - actually faster than the Spleeter implementation in previous versions of SL.
On a macbook air M1, using CPU to unmix 45sec of stereo audio it takes:
-20sec on SL9
-15sec on SL10 in Fast mode (which is much higher quality than SL9)
-30sec on SL10 in Balanced mode
-45sec on SL10 in Best mode
-You were wondering about frequencies above 17k. Here's a screenshot of a song unmixed in SL10. As you can see, no issues above 17k:
-Concerning future hardware acceleration, it'll likely be based on DirectML (on Windows) which allows support for NVIDIA, AMD and Intel HD Graphics GPUs, and on CoreML on macOS. I've already done some tests with those accelerators, but did not judge them stable enough for the launch date. Hopefully the situation will be better in a couple months. I added "where possible" because I know some hardware accelerator such as CoreML can't support all kind of AI nodes, so all AIs may not benefit from CoreML acceleration. And I also had some crash with DirectML on some AI models, so let see what will be the situation in a couple months...
