SL10 unmix, no bass extraction

I am new to SL. I have had the basic program as part of Cubase, but it was so limited, I ever used it. I often have the need to clean up audio and have always used Izotope RX for that. Now that the vocal unmix in SL 10 is competitive with the other products, I have done a crossgrade to SL10 pro.

The vocal and drum unmix seem very good, so I am happy with the product. But in doing some testing with some live recordings, I was surprised that the unmix of the bass produced almost nothing. The bass is not super high in the mix, but it is plenty audible. The piano and guitar unmix also got very little, but those are more buried in the mix (it was a jazz big band. I was just surprised that the bass extraction was weak – it is much better in Izotope.

I was also surprised to not see any controls/parameters to tweak the detection. Am I missing something? Is this really limited to basically one-button operation?

Is anybody else finding bass unmix weak?

1 Like

The datasets/libraries/dictionaries for these algorithms are pretty generic (meaning its basically a sum of all the basses of datasets and the the median is the average of those sums of basses) (meaning it’s like the concept of averaging off numbers, where you’re giving a set of data numbers and you add each set of those numbers and the output would be the average sum of those numbers). With that concept in mind, most of basses (and I cant speak for every dataset because every dataset and algorithm is different) in today’s music (so for example deep sub basses) are different then the basses in music from the 70’s, music from the 70’s more-or-so were actual bass from a bass instrument or bass guitar or string bass (not an actual sine wave). Music in 2024 is more likely to use a deep sub bass (from a sine wave) rather than an actual acoustic bass.

One technique I use to unmix a clean bass (without interfering with the phase) is to duplicate the layer and transform the layer (by frequency shifting it) down in frequency. Remember the amount you transformed it down to, because you will have to transform it back in the bi-polar polar opposite direction. Once you transform it down, apply unmix (to the transformed layer) and it should output a clean unmix result of the bass. When the bass is unmixed, transform the layer back to its original state(in reverse bi-polar of the polar opposite you original transformed it by) then turn on the phase in that layer and merge the two and you should get a clean bass unmixed.


When you say shift it down, are you talking a few cents, an octave, or what?

I honestly dont remember the sweet spot number. But try different combinations and see what happens. Try -100 and if that doesn’t work then try -170 and go down from there. It’s a trial and error process, so keep on tinkering with it until you get a clean bass unmixed. If you are working with a jazz recording then chances are its an acoustic mid-range bass (which is very different than a sine wave sub bass that is in a lot of audio recordings today in 2024), so try to lower the frequency of the bass (that’s where the algorithms mostly detect those kind of basses) and see what happens.

1 Like

Interesting. Generally, my main needs are to extract vocals or silence vocals, or to remove background noise from dialog. Those things all seem to be working nicely with SL10. Mostly, I was just surprised that I could hear the bass perfectly well, but SL10 didn’t put much of anything into the bass layer.

I had not given much thought to the sine wave versus upright jazz bass. What you mention is a fundamental attribute/limitation/fallacy of “machine learning”. What we call “learning” is really just a very complex mode of pattern matching. And with AI, the “learning” depends entirely on the various patterns that are fed into the model. Here is a snippet from my test file. It isn’t a great mix, but I can hear every note of the bass and would have no difficulty transcribing it, yet the algorithm found basically nothing. It seems like that needs more work.

I was also surprised that I didn’t see many controls over the separation process, as distinct from Izotope RX. But the SL10 algorithms for sung vocals and speech seem to work great without any tweaking of controls. Maybe there are some controls for advanced users. I have been reading the SL10 documentation and have not yet seen anything like that.

Hey man, this one was a little tricky. Along with what I mentioned earlier (about this being a jazz mid-tone acoustic bass instead of a sine deep sub bass) there is also an abundance variety of dynamic range in that bass. Most basses today are more-or-so monotone in a 1 dimensional aspect (and the basses of today hardly vary in velocity or dynamics) . So that is why the algorithm could not detect this as bass.

I manually had to unmix the drums and then used the unmix levels process (Average power with the horizontal average maxed out and the vertical average 2 px) to unmix in a more efficient manner (Because the bass is mostly vertical along the spectrum while the drums is mostly transients/horizontal).

I did the rough draft for you, see attached mp3. It’s left to you to clean up.


Cool. But I don’t really understand the quoted text. Is that all within SL? I have a feeling there is a whole level of controls I am not seeing because I am new to the program.

@cparmerlee The processes @Unmixing is refering to are:

1 Like

Thank you. I know there is a lot of depth in this program, so I need to get to work learning about it.

Big picture: for me, it wasn’t worth putting in the time in the past because it wasn’t the best solution for the things I need the most (isolating vocals and removing dialog noise). But with SL10, those things are great – and alone worth the price of the Pro version to me. But now that I have bought it, I want to understand it well enough to use it in other situations that will undoubtedly come up.

Thanks for putting so much time into that. It is greatly appreciated.



No problem.

Also I wanted to correct something I said earlier. I mistakenly said the bass is vertical and drums is vertical. I meant to say the harmonics (and or partials) of the bass amplitude is mostly horizontal along the spectrum, while the drums transients is mostly vertical.

1 Like

First of all Spectrolayers 10 is a great program and I have unmixed many of my old band live stereo recordings and improved them a great deal - but would like to get them better yet.

I am having the same issues with successfully extracting bass from a live three piece (vocal, guitar, bass and drum) recording. Very little bass is extracted and much of the bass remains in the Guitar stem. @unmixing sheds a little light on how he does this using Levels extraction but it makes little sense to me - and unsurprisingly does not work since I have no clue what parameters to set.

The links that @Robin_Lobel gives to the manual unfortunately demonstrates what is unhelpful in such manuals (Cubase and many other programs fall in to the same trap).

The manual page for Level un-mixing is technically correct but gives NO CLUE of how the level unmixing helps, under what circumstance you might want to use it and when, say, Average Power would be the better choice than Peak power etc and what the other parameters to tweak actually do to the result - horizontal versus vertical and so on. The screenshot is pretty but does not help as there is no explanation on what is going on.
A tutorial for such things would also be useful - but I have yet to find one on the use or best practices for Level Unmixing or Component or Multichannel unmixing. These are obviously advanced topics but some clue of their usage would go a long way for many of us!

Keep up the good work - but please give some thought to improving the manual to be helpful rather than just list the parameters you can see on the screen - and try to explain how they can help! - rant over…


The unmixing levels process is by far one of the best features in any spectral editing application and puts it miles ahead of any other competitors(even RX Izotope).

(As far as automated processes goes) I’ve unmixed some aggressively complex sounds using just the unmix levels process. For example I have unmixed complex synths with multiple layers of conplex arpeggios using just the unmixed levels process. Unmixed levels is underated and is a very powerful tool once you learn how to use it. When you change the fft size/resolution and fine-tune it, it takes unmixing to a whole new level. One thing I hope improves about the unmixed levels process is that it becomes hardware dgpu accelerated to where you can preview in real-time (because as soon as you change the resolution/fft and try to preview it in real-time, it cannot process it).

I dont have a problem doing a tutorial on unmixing levels and demonstrating horizontal px and vertical px and explaining peak power vs average power, however I don’t work for Steinberg.


Noted - adding advices and general comments for each process could certainly help indeed.
However, have you clicked on the Youtube videos linked at the top of the page ? They show you some practical examples.

1 Like

Thanks @Unmixing . A tutorial for unmixing a bass guitar using Unmix levels (and whatever else makes sense) would be very useful to me at the moment! You say Unmix Levels is underrated - I am not surprised since its use is pretty obscure based on current resources - but I am excited to hear you say it is one of Spectralayers best features and takes unmixing music to a whole new level! I appreciate you are not a Steinberg employee - but maybe this is an opportunity for some contract work! They need better and more detailed videos etc afterall! You could become the Dom Siglas of the Spectralayers world!

@Robin_Lobel - thanks for replying and I really appreciate the work you do. I do hope you are able to gradually enhance the manual with example, tips and best practices etc. I think it will greatly improve the popularity of such a powerful program.

I have indeed watched the videos linked in the manual page - but they did not help at all for my use case. The one on manipulating the levels of an explosion for instance does not say what his objective is and how to select the best parameters to achieve it. I am unable to translate this for me to better extract a bass guitar.
That video is rather like the manual in that he just shows how to change the parameters for unmix levels - and sure enough - the sound of the explosion changes. It is not purposeful.
I have also done exactly that for my recordings but since I have little understanding of the use of this tool in my context it has been unsuccessful.

With any sophisticated program like SL, one needs to learn not just the details of the various commands/progesses, but also the underlying philosophy. Over the years, there have been many tools that do some/many of the things that SL does. But I find that SL’s philosophy is quite different from others. Once I was able to get oriented with that philosophy, I was more comfortable.

A common problem with many complex applications is that those closest to the development tend to take many of the basics for granted, But these basics may not be obvious to new users, and they stand to lose potential customers for lack of a proper orientation.

I find that, in order to be productive with any of these tools, it is very helpful for me to write my own personal guide, basically a compilation of all the things I am likely to forget if I don’t use the program every day. I am doing that for SL. This is written for an audience of one (me), but others newcomers might find my brief orientation useful. Because the SL philosophy is quite different from other products, I wrote myself a summation of how I see that philosophy. It will probably sound very superficial to those familiar with the product.

This is a spectrum editing tool, with capabilities similar to Izotope RX. The big difference is that many of the SL operations divide a sound clip into 2 or more layers. In most cases, the layers are lossless, meaning that if you mix them together with no gain or effects changes, you end up with the original exactly.

You can then edit and process the layers individually. You can recombine them within SpectraLayers, or you can send the individual layers to the DAW to become individual tracks in the mix.

A major focus of SL is “unmixing”. This began with noise removal, separating useful material and noise into two different layers. But later versions of the product offer many different forms of unmixing. These unmix processes are heavily AI-based, and the training data is biased toward commercial pop/rock music. In those genres, the unmixing is quite impressive – less so in other genres, but still useful.

In contrast to RX and other programs, each of the unmix processes has few parameters. Nonetheless, the default results are often very good. The philosophy is that, because the splitting is normally non-destructive (all the content is somewhere in one of the layers,) you can make adjustments using manual editing, and possibly merge some of the layers, and run the unmix process again on the transformed layer.

SL operates as a stand-alone application or as an ARA-capable VSTplug-in.


cparmerlee I second your description of the usual developers bias regards the information disclosed about the program and its tools.
Then your recollection of the philosophy amplifies mostly what has been summarized in this thread and thus, is missing/misrepresenting what SL has been and actually what it is imho.
SL actually began delivering a complex set of interesting tools novel in the audio editing world. These allow for diverse and very precise selection within the spectral domain, alike what Photoshop does in the visual field.
That stated, in contrast to RX and other related programs, unmixing of all kinds in SL has plenty more parameters. For instance, users may alter the selection focus of the whole editing platform and hence interact with the chosen level, area, density of the sound to be intervened in fact going to micro-details spectral wise.

In short, Spectralayers you have at hand the widest and more varied tools to intervene the sound file, together with the most able edition and layering edition platform (10 years of cumulative development and user experience throughout 10 versions of the product), plus its unique FFT and “video” focusing controls that multiply said platform possible results.

EDIT: Plus, it has to be noted, that SL provides several specific menus prior to the application of AI processes where to change specific parameters that will then alter the results obtained by those, in fact allowing users to “tweak” or accommodate said processes to our best liking.

I don’t dispute any of that. But for new users, that may be a bit overwhelming. After all, if that was the level they were entering at, they would already have become a SL user years ago. For me, the philosophy is the important first step of understanding before getting into the deep weeds you speak of.

But everybody absorbs information in different ways. Whatever works is OK with me,

Yes, cparmerlee, just stating that if you are about to write a philosophy of Spectralayers imho you need to include the first two points I describe (SL has the widest and more varied tools to intervene the sound in the spectral world of editors and also, that it has an unique set of “video-like” focusing controls that multiplies its possible results many times over).
These two, have been the root and base of SL again in my view.
Then there is the third axis, which has been incorporated since AI processes were included two SL versions ago, that is the inclusion of user configurable points in the AI assisted tools.

As it is easy to see, SL is about providing deep user control and manageability, while workflow, efficiency, fast real-time operation and even automation have also been included in different degrees.

I am a little confused about your reference to specific parameters. Your use of the program may take you immediately into the guts of it. But my interest begins with noise removal and vocal isolation. In those cases, I see no parameters at all. In contrast, the RX noise removal dialog has a dozen things you can tweak, some of them incomprehensible to me.

When I see that, I conclude that the philosophies behind the two programs are entirely different. RX (and most other programs that have presented noise removal algorithms) has a single dialog that you must make work because that’s the only way to get the result. The philosophy of SL seems to be that each process should be minimalist, as far as parameters to tweak, and the power comes from the fact that most processes result in content being separated into additional layers, which are collectively lossless. If you end up with 10 layers, and mix them together at their original gain, you get the original mixed waveform.

That is completely different. For me, nothing else about the program makes sense until I understand that essential approach to the problem. But that’s just me. If you have a different way of wrapping your mind around the concepts, that’s OK with me.

same issue. Just purchased SL 10 and unmix has a difficult time extracting bass. RX does it better albeit with some artifacts. I performed an additional unmix targeting bass on all the originally created layers and that produced nothing. There is probably a method but I haven’t found it yet.