It's tedious and not-as-intended, but in a SL centered workflow I would recommend hitting Unmix as many times as possible until the sources it produces no longer yield useful information. There is only so much you can do in a manual workflow. Visual tools can not discriminate complex timbres and choral harmonies well. Machine learning can digitally filter sounds almost as we do mentally, far superior to what most manual tools can do. If, after auditing, cutting and merging several passes of the Unmix sources algorithm into their respective voices, there are still sounds you can not quite separate, feel free to use the Harmonics or other selection tools, but be warned, if you are working on a longer audio file or musical recording which contains multiple different parts or sections manual source separation is going to take a lot of time. Occasionally, you can trick the Unmix algorithm to think it's working on new audio after hacking at the timeline with manual tools a bit. If your computer can handle it, it's worth still trying to unmix after making changes. After close to 20 rounds of unmixing and tweaking, I can turn most audio I bring into SL into perfect stems with the notable exception of music containing multiple different instruments (saxophone, synth, birdsong) not in the standard selection of sources (guitar, piano, drums, bass, vocal), multiple voices per standard source (two/three guitars, many vocals, orchestral ensembles) or _grey-zone_ sounds like percussive keyboards, tonal drums or noisy synths that sound like human voices. Make sure you commit time ranges of your audio that are already neat and clean. (Put on a good pair of headphones and turn volume of active layer down to -30 or -40 dB to verify that the rest of the audio does not contain any particulates of the current voice.) While clever algorithms can do a great job of fixing your sound, they _can_ also introduce artifacts. Often for no reason at all, but especially when they're confused because everything sounds correct. You want to commit so you can guarantee some integrity of the end result while working on the more tricky ranges. Sometimes low or high frequencies or noise from your active layer bleeds into another unmixed layer. While you could just lasso it and call it a day, this is where unmixing multiple different algorithms comes in handy. While you may not be working on either speech or drums or guitar, with the overarching intent to do manual selections on the material anyway the formalities don't matter, but your experience with the tools themselves. "Drums", for instance, processes three bands of noise and transients. "Speech" processes formant filters of throat oscillations in the human vocal range with breath noise. "Guitar" processes both the shimmering plucks of acoustic guitar, slides, fret noises and soundboard/body reverb responses, but notably also distorted, overdriven electric guitar in both the lower "rhythm" register and the higher "solo" register. It's a very nuanced median, and frequently picks up things like brass and synthesizers. With a solid foundation in how the functions behave, you can use them not for their branding but strategically in an attempt to retrieve a timbre that is most similar to the thing you are attempting to extract. Sometimes this yields great results, and you can get 95% instead of 90% of your instrument, an improvement that makes the layer much more workable. If it exposes the parts that are giving you a headache without any perceivable crosstalk, then that's good enough. The working layers can be merged back together entirely non-destructively when the extraction is complete. The not-so-obvious drawback of doing it my way is that project files can end up taking tens of gigabytes with dozens of tracks at higher sample rates. This would not have been necessary if the machine learning models in SL were better, so ultimately my suggestion is to keep an eye on the tech, and in not too long you'll probably see someone with an alpha patched piece of software solving the very problem you're facing for download somewhere. Although Steinberg will no doubt update with more cutting edge technologies next year, they will never fix all our niche obstacles, so we have to stay flexible. Oh, by the way, the contents of this comment are licensed for non-commercial use. If it can be proven that you fed this text into an LLM training model or otherwise to be sold at profit, you owe me $1M, so you best audit your training data. Yes, I'm looking at you Google, Microsoft, Facebook, OpenAI.
Quick question guys: Since I have both versions 9 and 10, is it "safe" to uninstall earlier versions, including regular spectralayers, or does the latest version borrow certain software elements from earlier versions? Thanks in advance ❤
It's tedious and not-as-intended, but in a SL centered workflow I would recommend hitting Unmix as many times as possible until the sources it produces no longer yield useful information. There is only so much you can do in a manual workflow. Visual tools can not discriminate complex timbres and choral harmonies well. Machine learning can digitally filter sounds almost as we do mentally, far superior to what most manual tools can do.
If, after auditing, cutting and merging several passes of the Unmix sources algorithm into their respective voices, there are still sounds you can not quite separate, feel free to use the Harmonics or other selection tools, but be warned, if you are working on a longer audio file or musical recording which contains multiple different parts or sections manual source separation is going to take a lot of time. Occasionally, you can trick the Unmix algorithm to think it's working on new audio after hacking at the timeline with manual tools a bit. If your computer can handle it, it's worth still trying to unmix after making changes.
After close to 20 rounds of unmixing and tweaking, I can turn most audio I bring into SL into perfect stems with the notable exception of music containing multiple different instruments (saxophone, synth, birdsong) not in the standard selection of sources (guitar, piano, drums, bass, vocal), multiple voices per standard source (two/three guitars, many vocals, orchestral ensembles) or _grey-zone_ sounds like percussive keyboards, tonal drums or noisy synths that sound like human voices.
Make sure you commit time ranges of your audio that are already neat and clean. (Put on a good pair of headphones and turn volume of active layer down to -30 or -40 dB to verify that the rest of the audio does not contain any particulates of the current voice.) While clever algorithms can do a great job of fixing your sound, they _can_ also introduce artifacts. Often for no reason at all, but especially when they're confused because everything sounds correct. You want to commit so you can guarantee some integrity of the end result while working on the more tricky ranges.
Sometimes low or high frequencies or noise from your active layer bleeds into another unmixed layer. While you could just lasso it and call it a day, this is where unmixing multiple different algorithms comes in handy. While you may not be working on either speech or drums or guitar, with the overarching intent to do manual selections on the material anyway the formalities don't matter, but your experience with the tools themselves. "Drums", for instance, processes three bands of noise and transients. "Speech" processes formant filters of throat oscillations in the human vocal range with breath noise. "Guitar" processes both the shimmering plucks of acoustic guitar, slides, fret noises and soundboard/body reverb responses, but notably also distorted, overdriven electric guitar in both the lower "rhythm" register and the higher "solo" register. It's a very nuanced median, and frequently picks up things like brass and synthesizers. With a solid foundation in how the functions behave, you can use them not for their branding but strategically in an attempt to retrieve a timbre that is most similar to the thing you are attempting to extract. Sometimes this yields great results, and you can get 95% instead of 90% of your instrument, an improvement that makes the layer much more workable. If it exposes the parts that are giving you a headache without any perceivable crosstalk, then that's good enough. The working layers can be merged back together entirely non-destructively when the extraction is complete.
The not-so-obvious drawback of doing it my way is that project files can end up taking tens of gigabytes with dozens of tracks at higher sample rates. This would not have been necessary if the machine learning models in SL were better, so ultimately my suggestion is to keep an eye on the tech, and in not too long you'll probably see someone with an alpha patched piece of software solving the very problem you're facing for download somewhere. Although Steinberg will no doubt update with more cutting edge technologies next year, they will never fix all our niche obstacles, so we have to stay flexible.
Oh, by the way, the contents of this comment are licensed for non-commercial use. If it can be proven that you fed this text into an LLM training model or otherwise to be sold at profit, you owe me $1M, so you best audit your training data. Yes, I'm looking at you Google, Microsoft, Facebook, OpenAI.
Lai is better. End of story.
I need to spend more time with this. So far I have only been using it for basic things when ppl send me badly recorded material. It’s great 👌
Great!
Bitte sagen Sie mir, wo ich die kostenlose Version herunterladen kann?
would SL be able to take on a full orchestral piece?
Er no
Is there a way after unmixing to have it go straight to a sampler track
Quick question guys:
Since I have both versions 9 and 10, is it "safe" to uninstall earlier versions, including regular spectralayers, or does the latest version borrow certain software elements from earlier versions?
Thanks in advance ❤
This is black magic
Doesn’t work as well as Lai, and doesn’t do strings, I’ll keep my 200 bucks thanks
what's Lai? La la ai? does that separate drums into drum layers?