Regarding the fast vs high vs extreme:
Does Extreme provide better separation to the expense of introducing more artefacts ? Or is it just better in any way ?
I’m thinking here about some algorhithm from other software, where sometimes 100% is not always the best choice (vs 70% for exemple), if you know what I mean.
In other words, is sometime “high” better than “extreme” ?
With AI and audio processing there’s always some statistical uncertainty, but Extreme is very likely (like 90% more likely) to provide better results than High. Extreme is basically High using several passes, and averaging the end results.
Still, you may always find some rare case where Fast might give better results than Extreme on vocals, because they use different AI models - so if Extreme doesn’t sound quite right for some reason, it’s always worth the shot to try the other quality settings. But again, Extreme is usually a safe bet
I am a bit confused here. Let’s say, for unmix song, drums, what is best: SL 10 Best or SL 11 Extreme? SL 10 Best or SL 11 High?
I tried unmixing song, Drums and Bass, yesterday on the Extreme setting using ARA in Cubase Pro 13, and had to Ctrl/Alt/Del force close Cubase since the unmix progress just stood still at 0%…
Admittedly, this was a 20 minute wav. But in SL 10 on the Best setting, this used to take about 4 minutes, using the gpu acceleration (Geforce RTX 3060/12). Will this kind of performance be restored in SL 11? If not, it is just not usable on my setup.
@Magnus_N Yes patch 1 coming this monday brings full GPU acceleration to Unmix Song (as long as you have a GPU with at least 8GB of VRAM, which is your case). You’ll find some numbers here:
And regarding quality, SL11 will always be better than SL10:
Thank you, Robin, for confirming this - and for the quick patch! Beautiful work, as always. And I still cannot quite believe the generous “early bird” discount.
There’s actually a chance that it will, but I prefer not to create false hopes here. You’ll find out on monday
That being said, I’m thinking of a way for patch 2 to offload some of the GPU processing to CPU in case there’s not enough memory in the GPU. So 4GB GPU could still benefit from the acceleration (just less than a 8GB GPU).
Just wondering, how scalable is the gpu acceleration in Spectralayers though? @Robin_Lobel I have a RTX3060 with 12gb gddr6 on my workstation but also have a 2nd workstation which has a RTX 3090 with 24gb gddr6. WIll the faster with more gddr graphics card give better/faster overall result or is there a limitation of sorts?
Anyway, SL11 is amazing. This update for me is soooo substantial considering I was using SL 9 for. Thank you Robin:)
Yes, it’ll definitely be faster. That being said, above 8GB VRAM the speed increase is purely due to the computation capacities of the GPU (such as number of CUDA cores and clock speed)
Slightly off-topic, but I think over the coming years we can expect great improvements in efficiencies, both through the techniques used as well as the hardware. I’m hoping for something like an affordable dedicated PCIe accelerator card based on FPGAs.
The Unmix Song hanging issue is fixed in 11.0.10 and the speed improvement now that GPU acceleration is working on the PC is huge, although curiously it only uses 7% CPU and 20% GPU even when set to Extreme. Here are some comparative benchmarks using a 13m:41s 96KHz stereo file.
Mac Pro (16 core, 384GB RAM, 32GB Radeon Pro Vega II)
Fast - 3:48, High - 6:25, Extreme - 25:31
PC (13900KF, 64GB RAM, 24GB RTX 4090)
Fast - 0:38, High - 0:51, Extreme - 3:05
Yup, gpu acceleration working like a charm now. I unmixed drums and bass on a 24 min 44.1 kHz/24 bit stereo wav file using the extreme setting. Before the update, this didn’t work at all. Now, it took ca 20 mins.
I did this from within Cubase Pro 13. Only issue is the HUGE cpr file that results: 2.5 GB in my case. Removing the SL extension on processed files and resaving the Cubase project a couple of times, brought cpr size down to normal again. Is there an easier way to do this? What happens if I make extension permanent? Will that reduce the cpr size back to normal?
Anyway, a big big thank you for this quick update!
… or why not use standalone SpectraLayers to do this? It’s easy to drag and drop events directly from Cubase into into SpectraLayers, and to drag and drop one or more layers from SpectraLayers 11 directly into the project window in Cubase. Works fine and avoids all messing around with ARA complications within Cubase.
Yes, thanks, what a good idea. That never even occurred to me. And I didn’t know that you could drop events between the different apps! Cool, will try.
Unmixing on SL10 is twice as fast on my M2 Macbook as SL11.
While it’s nice to have an unmix crowd option, it doesn’t extract the applause particularly well. On SL10 I could achieve a better result unmixing song on a room mic feed to “drums” and “other”, then unmixing the drums down to “snare”. Combining “snare”+”other” then gave decent crowd noise and I could sidechain duck the actual snare from the it in the DAW.
I like that SL11 can batch process, which semi-offsets the new speed issues, but I’ll still have to use SL10 to extract clapping, or when I need something unmixed in a hurry. Either that or save up to purchase a dedicated PC to run SL11, and that’s an expense I could do without
As said in the other topic, if you uncheck “Sax & Brass” both SL10 Fast mode and SL11 Fast perform equally fast. Checking Sax & Brass requires an additional AI model which slows things down.
Regarding unmixing options, nothing is removed in SL11 from SL10, so if another workflow or module works better for you, it’s still available in SL11, no need to use SL10