PG Music Home Forums General Discussion Off-Topic State-of-the-art Stem Separation for Bass

I’m seeking a stem separation software program that can finally meet my needs. I’m assuming the state-of-the-art is there; but I could be wrong and another year or two may be needed for this technology to mature.

My experience with SongMaster Pro (ver 2.1) has so far, been lackluster in this regard. When separating the bass I have found the resulting audio often sounds warbly and squishy and even has regions of no music at all when I know music should be there.

Here are some applications in no particular order that I understand can do at least a limited level of separation.

SongMaster Pro
Voice.ai Stem Splitter
Gaudio Studio
AuDimix
Ultimate Vocal Remover
Splitter AI
LALAL.AI
iZotope RX
RipX Daw
Spectralayers
Melodyne
Acon Digital Remix
Moises
Logic Pro (macOs Only)

Eventually, my final goal here is to produce quality bass tabs of songs I want to learn. Bass separation would be step #1 followed by (a different?) software to produce a tab from the stem. For me, a great solution would be if my DAW (Studio One) had this capability built-in but I'm not holding my breath.

Has anyone found an application that can extract bass stems you are actually happy with across at least 2 or 3 different genres? FWIW, I'd prefer a stand-alone program as opposed to a website for this.

If no, is anyone interested in doing a semi-deep dive on this subject offline?

You forgot Rip X Daw

You can watch some of examples from this Video How Harmonies work He is using Rip X in this video.

Thanks, I added it to my list.
Have you used it to separate bass?

I do not need to 'use' the stems in any further form; I only use Song Master Pro to hear them isolated when there is some question while I transcribe a song.

I cannot speak about all those other programs you listed, but I thought they all use the Spleeter technology, so I would be surprised if they produce markedly different results. That's just a guess, however.

I have a number of Spleeter based apps including Ozone 11 Advanced. Spleeter doesn't do a good job separating the bass from the harmonics in the other instruments. Nothing does, really. Not hearing a bass line well enough to transcribe has never been my problem, however.

Steinberg's Spectralayers Pro does a better job but nothing is great on bass.

I used Spectralayers on a live recording I did about 25 years ago for a new recording of the same song. I was able to extract the voice of a friend who had passed away in 2008 to blend him into the chorus. On listening, I realized that my upright bass was at a similar tempo and, if isolated, would be nice to include it. My handicap precludes me from ever playing a string bass of any kind again.

I had to give up after a few hours — just couldn't get the track clean enough to re-use. Easier to play it again using VIs.

So far, FWIW, the consensus seems to be that the technology simply isn't there yet.

No doubt, to do this right is a herculean challenge. Imagine a song with a heavy kick drum, low-end piano and keyboard, maybe some tuba or trombone plus bass guitar; all hovering around the same frequency band. It really is asking a lot for an AI code to find and isolate the bass guitar "needle" in a giant low-end and time-changing "haystack".

That said, I believe it's just a matter of time before it's done and done well. The basis of my optimism is how LLMs have taken the computer community by storm; truly amazing, especially in medical applications.

I'm guessing that this will require a minimum 2-step process. Step #1 to isolate (as best it can) the bass instruments, in part by it's frequency content to produce a "pre-stem". Then Step #2 (the hard part) is to take that pre-stem and apply a musical knowledge to remove everything that isn't bass guitar. What remains would be the desired bass stem.

I believe that David Cuny first said that PGMusic would be well-positioned to do this kind of work because BiaB could be used to create training data sets for such an AI program.. Each data set would be made up of "the problem" (a song containing a bunch of overlapping low-frequency instruments) and "the solution" (the crystal-clear bass track). But you couldn't do this manually. You'd need to create a program to automatically create these sets by the thousands if not millions across a handful of instrument lists, genres, keys and tempos.

Thanks for mentioning Spectralayers, I added it to the list above.

Steve, I agree with your conclusion but since it is a technique in its relative infancy, I hope we will see improvement. It may never get to the point where it gets everything right, and you gave one good example.

While the isolated bass line may sound rather poor in Song Master Pro, there can be a benefit in having it muted if you want to hear the rest of the mix more easily, or use the track as a music-minus-one for bass.

Hearing the bass line certainly isn't any problem for me either, Mike, but if a tool exists to make it easier, why not use it?

Not just bass no. I separate the entire song to hear all the parts of a song I want to learn. The program has a couple of versions I believe and is quite powerful. I don't think it is perfect by any means but considering this was considered impossible some time ago. It was like how to un-bake a cake. For Mac users with Logic Pro 11 this feature is built into the software if you have a M chip.
I only use this program to dissect a songs parts to learn what it is doing. I do not use any of the parts itself.

I hear ya Matt, I frequently "mute" the bass in Studio One using Pro EQ when I want to record bass over popular songs I like; basically high-pass filtering. This has served me well for years even back to my early Audacity days. Of course, no stem extraction is being applied in doing this but it does work well for its intended purpose.

I wasn't directing that at you. Are you trying to make bass TAB?

Never touch the stuff.

I am surprised that no one has mentioned Melodyne yet. I use it when transcribing bass lines. I really don't have any use for isolated bass stems. I want to get an accurate transcription of the bass line, and then practice to the complete song, playing along with the original bass player. I have Melodyne Studio, so I can't tell you what the minimum version needs to be to get the different audio sources separated.

Melodyne is great for determining what is a bass note and what is a note from another instrument. If there are two notes close to each other in the piano roll view, just click on each blob to hear it and identify it. Melodyne is especially useful when dealing with microtonal sounds, like you could get from a double bass note played between "fret" positions.

For example, identifying the first two notes from Michael Arnopol's superb bass line In Patricia Barber's "Too Rich For My Blood" drove me nuts until I used Melodyne. I just couldn't get the right sound for those two notes on any my fretted basses. It turned out that the first note was 50 cents sharp, the second note was 25 cents sharp and the remainder of the 8 minute bass line was in a spot on F major. I came close with a fretted bass using a 1/2 bend for the first note and a 1/4 bend for the second, but it wasn't a smooth transition.

Being able to see the bass line blobs in Melodyne on the piano roll makes creating the bass line notation for a given song easy. Creating the tab after that takes some time playing through it and figuring out where you want to be on the neck for any given section.

It may be possible sometime in the future to have software generate an accurate notation of a single instrument from a band or orchestra performance, but automatic generation of the best tab from the notation won't happen any time soon, IMO.

Hey Maartian, thanks for your very thoughtful reply; I added Melodyne to the list above so to be visible in the coming days and weeks for anyone else interested in this subject. You are way further down the musical road than I am.

Am I correct that your method is to view the entire mix (all instruments) in "blob format" and upon playback you're basically slowly and manually moving from blob to blob to identify bass notes? And then once identified you decide where on the fretboard to place that note on your tab? If yes, I'd say that's Melodyne-assisted manual transcription, a skill that I'd have to develop.

FWIW, on many songs I want to learn I'll print the chord sheet and play only the root to verify my chord sheet is accurate; this gives me an idea of where on the fretboard I'm operating. If required, I'll start this at 10 or 20 BPM slower than the original. Then I'll slowly increase the complexity as required by adding the 5th, then arpeggios, then small fills/runs until I'm somewhat satisfied to claim "victory". I suppose this too is manual transcription, but not nearly as advanced as what you do. Of course, my basslines will always pale compared to the pro in the band. And so I'm looking for a better, minimal effort and maximal accuracy method to up my game.

I also overlooked one: Acon Digital Remix. It's a plugin for VST, VST3, or ProTools. It's made by the same folks who make Acoustica 7, the stereo editor I use now.

I totally forgot I had a license for this. It's in the Tools tab of Acoustica Premium. Thanks for the reminder!

Yes, that's basically correct. In most cases, it's pretty obvious which blobs are the bass notes. It's only when the bassist is playing near the bridge on the D or G strings that notes from another instrument like a guitar or piano can make it confusing. That's the only time I need to audition the blobs. Also, I make no attempt to go straight to tab. The first step is to transcribe the bass line to notation. That is unambiguous. An A5 is an A5. But there are several A5's on the fretboard. I start working through the bass line notation and figure out where I want to be on the neck and then develop the tab from there (if I even want/need to create at tab).

I never heard of this one, thanks.
We have quite a list. Assuming these applications represent different software teams and that these teams continue to maintain their products, there's a bunch of people working this problem which tells me that major progress is quite likely.

Also, for you 6-string guitarists and maybe banjo players?, in principal, what we're talking about here could benefit you too.

FWIW, I recently became aware of another player in this space; Moises.
So I tested it out on a song and the bass separation is roughly the same as SongMaster Pro . . . lackluster.

All of the Spleeter based apps perform about the same. Spleeter is open-sourced freeware if you want to try your hand at it.

Logic Pro 11 has an impressive splitter. I’ve tried the bass removal on songs I played
on live with our band on stage and home
recordings. But all my bass was upright within the context of acoustic bands. I may try it at some point on a few recordings with electric bass. Whenever I get to a need something to do point. 😀

Bud

Hey Bud, it would take an "act of congress" for me to migrate from Windows to macOS but for grins if you need a song to split, You've Got Your Troubles is a good test case; it has lots of low-end brass.

SongMaster Pro will separate the bass but it's wobbly and full of static. This, maybe, could be cleaned-up with some EQ attention but it loses its mind around 0:42 and produces no bass at all, which I see no way to fix.

I find Some Master Pro quite good isolating bass with most styles, of course, the better the source is, the better results you'll get.

Moises works also very well, IMO, not 100% sure, but I think both (and many others) are based on the (open source) Spleeter code.

I use stem separation mostly to remove the original bass line, so it works pretty well for me. I guess you know SMP can export Midi from stems, a couple of versions ago that feature didn't work really well, but the last version seems to be considerably better on this regard (I've only experimented a little bit with that feature). Did you try it?

A classic tune. I can image it would take an enormous effort to pick a clean bass line from that.

Separately, I wonder what AI could do? Maybe it could intelligently(?) determine the bass notes based on multiple other factors within the song?

Absolutely! I'm telling ya, it's all about the newspaper ink

From Wikipedia: The Fortunes are an English harmony beat group. Formed in Birmingham, . . .

Why would anyone need stem separation, AI or anything else to hear the bass in You've Got Your Troubles, I've Got Mine? It's quite easy to hear every note in that bass line.

Yes, it is easy to hear the bass line.
But you are missing the entire point. This is about
1. Progressing towards software that can separate bass stems and then produce tab. Did you not read the initial post? And
2. Finding an effective stem separation tool for all songs not just You've Got Your Troubles, and
3. Establishing a common song to use as a test case. Without a common song, it's difficult to compare apples-to-apples.

These are not hard concepts to understand.
Do you get it now or do you still need teaching?

I’m intrigued by AudioTrack’s question.

Forgetting the exact song for a moment, at what point will software try to use AI to make a guess at what the bass was playing, when the stem separator isn’t sure?

Educated guess,,, That's what I would do...

I am not really sure how and if AI could achieve anything. However, presumably it could be used to interpret many other melodic components of the music, including the surrounding song structure, and from that establish what would potentially work along with the stem separation that had been interpreted, including suggested corrections to stem separation where there is a lower degree of confidence. In essence, could it smarten up / value add to the stem separation process?

(Disclaimer: I have no idea if this is really feasible, I was just throwing ideas around, but I think the concept has some merit.)

My answer is "yes".

Before the pandemic really took hold, my wifey and I enrolled in a music theory course at a local college. Three other couples were also middle-aged; everyone else was a twenty-something. The class was subsequently cancelled but we had already purchased the textbook; a massive, densely-written tome of some 800+ pages; Music Theory Remixed by Kevin Holm-Hudson.

My point is, a music AI could be trained on this book and every other music related book ever written. Those that are formally trained in music can appreciate just how much knowledge this would represent. Combine this with another (or the same) AI that has been trained on thousands or millions of songs and the result will meet every bass separation and tab generation task I could dream up. One down-side however, is that authors of music theory books may no longer be required. Simultaneous progress and anti-progress?

In the meantime, I'll use bass tabs that happen to be readily available and construct my own bass lines when they aren't.

My take on this is a little less optomistic, though I'm pretty sure it will improve.

The reason I'm less optimistic is that what the software is doing is identifying all the little nuances of the sound and attributing those nuances to a particular sound source, e.g., a bass guitar. How well it isolates depends on how well it does that attribution. It can attibute by frequency, phase, time and articulations. But if any other source supplies nuances that fit close-enough to the bass, nuances will likely be wrongly attributed. A solid electric bass is probably a fairly predictable sound, so relatively easy, but even then, slaps and percussive effects may sometimes be difficult to attribute to bass, rather than percussion. I presently hear quite noticable 'fails' between vocals and strings and synths and whilst I anticipate that will improve, I do wonder how much it realistically can improve, particularly when the audio processing perhaps interferes, e.g., by ducking, chorus and so on.

The way I'm thinking about using AI is that it might be able to work much more closely to how a human thinks and acts.

A skilled human could more easily hear through the surrounding noise clutter and determine that the bass note was obviously (for example) an Eb. Whereas using software or hardware to identify the required separation will be often clouded by all of those adjacent articulations and intonations, including the undesirable sonic variances that are present in lower resolution recordings.

But, bring AI into the picture, and you've introduced a more superior 'human-like' intelligent thinking into interpreting exactly what that note is, the same way that a skilled musician's ear could achieve a better result than some software separation applications might.

If a skilled musician could determine that the note is definitely an Eb, then I'm convinced that AI could deliver the same or better.

Just my thoughts on other ways to achieve a successful result. Perhaps premature at this stage, but give it time... it will happen.

It seems that AI powered Stem Separation will be available directly in the new release of Studio One Pro 7 as released today.

I have not bought it or tested it yet.

https://www.presonus.com/en-US/studio-one-pro-features.html

It's so refreshing to read the perspectives of intelligent people who can actually comprehend this thread and converse at an adult level. You and AudioTrack are among those that have a strong set of healthy brain cells, keep using them

Yes, percussive slaps and other musical elements (including musical "liberties" taken by the musicians) can certainly make it challenging for an AI to "properly" separate the bass stem. But remember, it doesn't have to be perfect, just good enough.

Google DeepMind founder Demis Hassabis has won a joint Nobel Prize for Chemistry for using AI to predict the structures of proteins. The potential impact of this research is enormous. Proteins are fundamental to life, but understanding what they do involves figuring out their structure—a very hard puzzle that once took months or years to crack for each type of protein.

So if the difficult protein-structure problem has largely been cracked, how hard can it be to extract the bass stem from a song by The Fortunes, especially if it has a knowledge of how good bass players play bass?

That sounds more like extracting the notes, e.g., as MIDI, rather than isolating the stems, which is what BassThumper was asking. For me, "the stems" means the actual sound of the bass extracted from the audio. I suppose one way to get a clean stem might be to identify all of the notes and articulations and convert that to MIDI and identify a suitable-sounding instrument and creat a MIDI+VST track that does a good emulation of the stem. I.e., creating a new stem that sounds very like the one from the song. One would certainly get a clean output.

How far it could go with other sound sources is another question. If one extracts the voice, translates that to MIDI and allies that to a vocal VSTi, how close can one get to a clone of the original. I guess if the original singer has good voice samples out there, possibly pretty close.

... and rebuilding would satisfy that.

When does an extracted stem become a straight forgery?

Suppressing the original bass in the remainder of the recording remains an issue due to artifacts from the bass that 'look' to the software like articafts of some other sound.

+1

Hmmm, now there's some unique thinking. My thought on this from the beginning was to keep everything in audio format, like how SongMaster does it. But if the separated stem were to be pushed into MIDI, now you have a much cleaner, note-by-note representation of the bass . . . added complexity in doing it but a much cleaner output.

Lately I've been thinking about a more hybrid approach. Rather than have an AI do everything from beginning to end, why not have a human in the loop? As I mentioned, SongMaster produced a bass stem with a gaping hole in the middle of the song. What if there was a checkbox: Are there gaps in the music? If "no", then it would know to prevent any gaps from occuring in the stem. If "yes", you'd be prompted to specify where in the timeline the gaps ocurr. This approach could be extended to deal with other areas of ambiguity thereby guiding it to do its job.

I wonder if such human guidance was built into the protein-structure code.

As user schlind just mentioned, I received my email notice to upgrade to Studio One 7, with AI-assisted stem separation. Four tracks:

Vocals
Drums
Bass
Other

Well Jimmny Cricket and Jumpin' Bullfrogs!!!
Look at my original post . . . ask and ye shall recieve

The fact that it has the same 4 stem categories as SongMaster kind of tells me that it uses similar underlying AI code which will result in similar limitations, but hey, it's a start and they can only build on this to hopefully reach my final vision:

Audio -> Bass Stem Separation -> Bass Tab

You must have special-early-bird-backstage-access to their products.
Thanks for sharing.

Perhaps the Song Master Pro version differs, because I have SIX stems:

Vocal
Drums
Guitar
Bass
Piano
The Rest

How well does your version do on The Fortunes Song?

Yes it does; those are omitted from the free version.

Ahhh, the old saying, "you only get what you pay for" <ha ha>
Thanks for the clarification.

I wonder how worthwhile the experiment would be, since this is a compressed file on YouTube. What would recordings from the 60s have had for full frequency response even to start with?

I have on occasion run several different stem-splitter applications in the hope of better results. This includes SMP (full version of course), Moises (free) and Ultamite Vocal Remover - UVR. In one case I did find that UVR gave me more favorable results for one of the tracks I was working one. Other than that one instance, I have concluded all these application, based on Spleeter code, are equivalent. None are Perfect if you really want to get the the note for note of any specific track. At the same time, all are miraculus at doing, what up till a decade ago was not even imaginable.

So I ran your "Fortunes" thru UVR - NO Bass track was extracted. The bass appeared in "the rest" with the rest. But are you telling me that no one has used the full paid version of SMP on this song?

Here is the first 32 bars illustrating all six stems. In regards to the bass, while it is not perfect, I am confident it is good enough for SMP (or melodyne) to render it to midi and than allow you to transfer the midi to Guitar Pro where it can be "cleaned up" and viewed as TAB. This is a quantum leap from a decade ago but if you can wait another decade it will separate each track perfectly and will use the stems to create a new song for you using the same notes and also send the notes to your bass so your bass will be playing along. But don't worry if you are not there, it will do it without you.

Out of curiosity, has anyone tried running one of these tools on the separated output of a previous run and typically does it improve, degrade or do nothing?

AudioTrack, I do have the paid version.
What version do you have?

Wheather Compressed/Not Compressed or 60s/No 60s I don't understand your question.

It's a software test case and a relevant one.

Matt brings up an excellent point. Has anyone tried to use a full frequency song on any of those stem programs? I would suspect the end results would be different.

Now here's a guy that's a "bird of a feather". Rather than just talk about the test case, he actually has done something physical with the test case.

As mentioned above, my results with SMP did produce a bass stem, but muddled and more troubling, a gaping hole was produced.

So my conclusion based on this datapoint is that SMP is superior to UVR on this test case . . . valuable info for sure.

I'm certain the folks that produce SMP and UVR are well aware of their limitations and probably hard at work to produce improvements, as well as everyone else in this business space.

I'd be more suspicious of the effects of psychoacoustic compression than of frequency response.
Even 60s half-decent recordings almost certainly had above 10kHz, likely 16kHz.
Grundig's 3-3/4ips recorders from the 50s claimed ~10kHz.

The very old 78rpm discs were typically significantly poorer, but by the 50s vinyl was taking over.

I fail to see how Matt's point is excellent.
Are we to limit ourselves to only "modern" uncompressed songs???

My answer is "No"; I want a full-fledged, full-bodied stem separator. I'm not asking for perfection, but I am asking for acceptable quality.

A software team, somewhere will deliver on this. I know this because AI is powerful.

My point is not intended to be excellent. Your test is reasonable. I am merely asking if the results might be different if we used a source file that didn’t start with the life already compressed out of it.

Was watching an Intro to Studio One Version 7 video last night and it has a built in "stem separation" function that will extract individual parts from a stem. The discussion on this feature starts at about 9:00 on this video. The example includes a bass extraction.

FWIW, I'm on Studio One Version 5.5 and don't plan to update...

Matt, I have no idea if the results would be different or not. And I don't have the experience to know what to do if the results are different. I'm just searching for a full-bodied solution to this problem.

If you can propose an uncompressed song to serve as Test File #2, perfect. You could extract the bass stem from the compressed file (Test File #1, supplied above) and an uncompressed file (Test File #2), with your version of SMP. I don't know how to identify an uncompressed file or how to measure its level of "compressionness". So that would be a rabbit hole for me. The majority of songs I work with are YouTube videos converted to MP3 format. I know this is not ideal from an audio quality standpoint but the number of songs available is literally staggering.

But no matter what the results are, I (and thousands of others) will have to deal with compressed files in the real world.
Are you proposing that an added step of uncompressing compressed files would be required? Is that even possible? If so, the nature of that would need to be established so to maintain apples-apples comparisons.

PS> Shouldn't an AI trained on compressed files be able to handle compressed files?

Uncompressing a compressed file makes no improvement to the audio.

BIAB provides a utility to do exactly that with compressed WMA RealTracks, along with a caution not to expect any difference in audio quality. It did provide slightly faster operation in the early days of ReslTracks.

I asked about uncompressed files because I wonder if the absence of the original dynamic range, or the absence of the upper harmonics, make any difference in the algorithm’s ability to distinguish between, say, a guitar and a bass or a bass and a tuba.

Yes, I see that ver 7 has built-in stem separation plus a busload of other improvements.
My experience with Presonus is that when releasing integer revisions they pack in a lot of goodies; I plan to upgrade.

But even if Studio One ver 7 doesn't solve this problem, my engineering instincts tell me someone will crack it. And shortly, with no restrictions regarding compressed files or tubas in the mix.

PS> Any tuba players out there, we love ya

Maybe a CD with the old song originally recorded with audio equipment and originally released on vinyl and later re-released on CD as it is i.e. without any remix etc, could be seen as the "uncompressed" version of an old song (but I don't know).

It could be ripped as a WAV file with Windows Media Player (legacy).
To make a compressed MP3 version to use as comparison maybe one could use Audacity and select a suitable compression level (unless already available on YT

All good ideas. Here's a thought:

If sufficient interest exists we could research and test this and related topics with the goal of producing practical "best practice" reports in PDF format with associated audio files. Based on how much we are all struggling here, I think I can guarantee that we'd all learn and grow. And perhaps in the bargain this could be the spark needed to extend this forum from being the center of the BiaB universe to a center for the much larger subject of audio processing and sound engineering.

If done right, it could be an attractant to techno-musicians in younger generations to this forum and hence BiaB. I'm not experienced enough to lead such an effort, but I'm not too shabby at documenting our results. There could be all kinds of subjects we could cover, subjects way beyond simply producing bass stems. Of course, we'd need smart, experienced, "worker bees" that are willing to stretch themselves.

I've been accused of "thinking too big" in the past, so if this idea is considered too grandiose for folks here, I get it.

FWIW, I think that in the near future AI will not only be separating stems with frequency but also with tone. Thus it could separate an acoustic guitar from an electric guitar, a bass guitar from a tuba, etc.

Hey Steve, I certainly never suggested that you didn't have the full version. In fact, with your significant involvement, I would be surprised if you had mentioned otherwise.

I don't have any version. I've never had a requirement to use it. My input here was really only to think about solutions that might give improved results, hence the discussion about using AI.

I apologize if my 'tongue-in-cheek' comment was taken the wrong way about finding that a paid version of anything usually had more features. It was only a generalization that can be applied to many things, and certainly not directed at anyone here.

Try Ripx DAW ( Ripx DAW PRO ) - Ripx DAW PRO has extra -
means of cleaning recordings. Ripx DAW can separate stems into more than four tracks and has its own algorithm for this, which is different from Spleeter. He separates the drums into three feet. If the bass is mixed with the guitar, it can be returned from the guitar track to the bass track manually. Notes as blobs can be tapped to sound. Both types of Ripx can be tried for free for 21 days.
https://buy.hitnmix.com/

I think the present software already does that to a very large extent, though perhaps not quite in those terms. I think it would be impossible to achieve the separation they're managing today without a good recognition of which bits of the total sound apply to which instruments. The envelopes will tell a lot, but certainly not everything.

If all instruments produced just harmonics of the fundamental, I suspect isolation would be relatively(!) easy, but when most instruments, including voice, produce overtones and sympathetic resonances, it must become very difficult to identify what sounds belong to which instrument. We've probably all heard times when a bass has causes a snare or cymbol to sound, similarly a piano with the dampers off.

To be honest, I wasn't quite sure how to interpret what you were saying, and so we got our wires crossed a bit.

All is well my friend. Hopefully by the time this thread has run its course we've all learned at least something regarding separation of instruments in general, not just bass.

With AI unfolding before our eyes, we are truly living in exciting times; not all that different from when we first walked on the moon.

"That's one small step for neural networks and one giant leap for human capability"

All good. I actually thought that must have been the case. Yes, AI is here to stay (like it or not), and unquestionably can bring about many improvements, and hopefully we will all continue to benefit from what it can deliver. Eventually though, to continue to evolve, AI will need to learn new things from, well, AI. That bit scares me

Ya, I got that one - it is a beast of features. Not one of my favorite. But, I agree that this does appear to be a different type of algorihym - sems to have the Melodyne blobs. I tried it on our test case. There was NO sound on the bass track (first 32 bars below) . The bass was mixed in on the Kick Drum track. And I didn't and wouldn't even try to adjust anything manually.

Many thanks Dan, you're saving folks like me from wasting time with software that doesn't perform.

Gause Zdenek, did you try RipX on our test case?

Why scared?
If an AI was trained by an AI and the result was a highly effective program to separate bass . . . I'm good.

Where I become concerned is if/when it becomes sentient.
In my mind, sentience implies life; artificial life perhaps, but still life.
And we know 2 important things about life; it fights to survive and it fights to reproduce . . . I don't think we want a "HAL" on our hands.

Sorry, but I haven't tried your test example, just trying to advise that Ripx is mostly sufficient for my needs as far as tools go. With my voice - bass position - under C2 I have a problem with Ripx, but also with Melodyne. But with Ripx it shows me two tones instead of one and they are higher. A note sounds at the correct pitch if both are sounded at the same time. If I mute one of them, it sounds at the pitch it is displayed at. I know the reason, because I sing under D2 (I can't do it otherwise) with the "Growl" technique and it is not continuous, like the bass tone (the one is continuous). These "Growl" tones could be replaced by perhaps a Tuba. I'll let here the "Growl" technique sound from professional Eric Hollaway (ThatBassVoice), I'm just only an amateur at the singing level. I have not tested this song on Ripx. Sorry for getting off topic. Here is the link:

Well gentlemen (and any ladies watching), I think we have success!

I upgraded to Studio One ver 7 and asked it to separate the bass on the above song by The Fortunes. This test-case is no mamby-pamby test case, it’s a stress-test; intentionally chosen for the low-end brass in contains. We might call it a “tuba test”.

There is so much that can be said here. The software team at Presonas knocked this out of the park. It isn’t perfect but is definitely something musicians can work with. I hear no brass, the bass that is retained makes melodic and rhythmic sense to me, there are no erroneous gaps and I can even hear the slides that he’s doing on his neck.

I exported this out in WAV format but that’s a moot point because SoundCloud will certainly step on it in its processing. FWIW, I hear no meaningful difference between the original WAV and what is up on SoundCloud.

I’d be interested in the perspectives of any audiophiles or bass players. To me, it sounds like he’s using a pick, which maybe he is on the original recording or maybe it’s an artifact of the separation process. There also seems to be a bunch of reverb present and it's a bit wobbly. But these are nits, it extracted a usable bass line from within a cloud of other instruments competing in the same frequency range and that just blows me away.

Note that Presonas’ solution didn’t care one bit about this being a 60s song or that it was compressed to keep the YouTube people happy. Most likely they trained it on “messy”, compressed, real-world training sets and that, in part, is why it works. If I were on the team, that's what I would push for; train on the lowest common denominator.

This thread is titled "State-of-the-art Stem Separation for Bass", based on one attempt with Studion One and imho, Presonas has achieved the state-of the-art and I'm guessing they're hard at work on 7.1 and beyond.

Very cool stuff!

Separated Bass Stem

Yes,
it seems to be a significant improvement.
I just loaded the YT mp3 version into SMP and extracted the bass stem and compared it with your SOP7 version on SC (above).
It is a significant improvement for the bass stem of that song.
It would be interesting to know how the other stems are.

I’m glad to read this result. I just installed Studio One Pro 7 but I haven’t had time to conduct any experiments for this thread. [ I’m immersed in writing Christmas concert music. ]

Congratulations, this is a great effort. If Studio one can do it, others will follow for us non-Studio One users. I just sent a note to our SMP support guys asking if future work is planned liked this.

Certainly no congrats are owed to me, I'm just a Joe-user. But I'm guessing they're popping champagne bottles at PreSonus over this release.

If I put my project manager business hat on, I'd be thinking about selling/leasing their stem separation code to people like RipX, SongMaster and DAWs that haven't yet matured this technology. But only their middle-shelf version; you never want to share your top-shelf stuff. It could be a meaningful revenue stream if done right.

I also like to think about and find info on what degree they collaborated with other entities to achieve this. I haven't found any press releases but my AI assistant says this:
PreSonus has a history of both in-house development and collaboration with external partners for various projects. While specific details about the development of the stem separation feature in Studio One 7 haven’t been publicly disclosed, it’s common in the industry to leverage both internal expertise and external collaborations to achieve complex technological advancements.

My guess is that they worked with at least a University, AI consulting firm, or a heavyweitght like Microsoft or OpenAI. It would seem to be a lengthy, expensive project to grow the required AI talent in-house. Kudos to PG if they collaborated with PreSonus, as David has said they are well positioned to do.

If anyone finds info on this, do share.