I listened with interest to the recent SonicState interview with Jessica Powell of AudioShake.ai (link below).

With my curiosity piqued, i tried it out and found it did a much better job of preserving transients than RX rebalance.

What i wasn't able (or more accurately willing) to do was to pay 60$ (absurd?) to extract a whole song to 4 WAVs so i could do a fair comparison.

Specifically, i want to know if the summed stems = the original mix. Because RX stems will sum back perfectly, which i think is important for (the future of) mastering. It's a shame Nick didn't ask her that in the interview.

I suspect that it's easier to extract a good sounding solo'd stem, if you don't require all stems to sum back to the mix. Perhaps a future (intermediate?) compromise with this tech will be to let you prioritize certain stems. So for example, if you demand perfect re-sum, you'll accept crappier sounding other instruments in return for better separated drums.

Obviously with their current pricing structure, AudioShake isn't something we could realistically use day to day, although i imagine the price will come way down in the not too distant future.

Thoughts?

Was curious too see this and learn more.

Sadly I feel that AI will take my job, who cares if if I have decades of experience with top artists and studios.

However if the AI solution ends up being more expensive than me, I still might have a chance.

Also serious question: has an actual audio engineer ever worked on developing these AI products? Sometimes I think they miss some of the subtly in certain productions.

This one will run and run...

Hopefully this doesn't become a re-run of recent well-worn "will A.I. take our jobs" threads!

Bear in mind that AudioShake are only (for now) offering a tool to help flesh and bone engineers do their jobs.

I'll say one thing about the jobs issue, then hopefully we can focus more on the tech. This is one example where A.I. is actually creating a new job, which is the job of mastering engineer who can do waaaay more than was possible in the past. To the point where we might sometimes impinge on the function of the traditional mix engineer. Although of course at the same time the mixing engineers are using A.I. to lessen the need for pro mastering! Mather-fackers!

Watching some of the video while typing.

A couple of ideas struck me:

1/ Take a mix you have the multitrack for and then compare whatever the AI spits out with the originals. This would be more useful than a tune you did not what the multitrack sounds like.

2/ What happens with more complex music e.g. Classical? Can it cope with 122 players at full bore?

Also that lady looks just like my Ex, which is slightly weird.
And yes I agree, we do not need to moan about AI.

Embrace it and make it work for you! Its only a tool.

These days RX is lagging well behind the state of the art. It is based on the relatively ancient Spleeter software - which is also available for free in various forms. RX doesn't even include Spleeter's 5 stem model which can extract piano as well as the other 4 tracks.

Demucs was looking promising up until a few months ago but I think that one of the key team members has moved on so I don't know if it is currently being developed. They brought out a 6 stem experimental model about a year ago which was very effective on drums, bass, vocals and guitar (probably at least as good as Audioshake) but less effective on piano than Spleeter. I was hoping that they would refine this and maybe add more instruments.

These tools are extremely useful for some of the work that I do.

I am not sure how AI function when is acoutics itself the one who sums. Like an in a string section.

Alrighty! So which stem splitter do you recommend today?

For me, workflow is very important. So when someone sends me a revised mix, i need to quickly be able to split and export. RX's workflow is reasonably good and it's only a few mouse click's work.

Never thought of acoustics as a summing engine before, good point.

If you listen to the interview it's very clear that AudioShake cannot separate things like individual string instruments in an ensemble.

Nowadays I run Demucs in the Ultimate Vocal Remover.

https://ultimatevocalremover.com/

Just select your input file(s) and output location and then click go. You can also choose which algorithm you want to use but, as I mainly seem to deal with guitar based music, the Demucs 6 stem model usually works best for me - but I'm still learning... (I only just noticed this morning that it will de-reverb the vocal track).

Alright!

And do the stems sum back to a perfect copy of the source mix?

No - but that isn't important for this type of work. If summing back to a perfect copy is important to you then there is an easy way to achieve that - but it is a totally academic exercise with no use in the real world when you need to use source separation.

I don't understand what you mean, if you're using stem separation in mastering, it's absolutely vital that they sum back perfectly to the mix.

If you are using stem separation then it probably means that there is something wrong with the mix so summing back to the original mix is the last thing you want. If you are using stem separation then you are effectively re-mixing in my opinion. Any stem separation algorithm can be made to null by deleting the "other" stem and then re-creating it by subtracting the original from the sum of the remaining stems.

Spleeter does pass the null test but, by using it, you throw away all the advantages (ie: better sounding individual stems) of using the alternatives.

Sorry, again, i'm not getting you. To say stem separation should only be used in mastering on bad mixes is very much just your opinion (you state it as fact).

Why would perfect re-sum be the last thing you want? Surely it's the first thing you want? Maybe you're only doing 1dB of de-essing on the vocal. You want the re-summed stems to sound arbitrarily different? This makes no sense.

And i don't agree that "Any stem separation algorithm can be made to null", this is obviously false, because an algo that's optimized to sound good when solo'ing the stems is clearly less likely to perfectly sum back again, unless they've made some breakthrough in stem splitting. I believe AudioShake is designed to sound good when solo'ing the stems, because that's most important to their industry clients.

"State of the art" right now for me seems to be UVR and some of those models like Kim 2, de-reverb and de-echo. I mainly use it to generate acapellas I can remix, and it's better than the commercial $300 RipX program I paid for a year ago (although that still has its uses). Plus it's 100% free. Unless there's something even better I don't know about? It's crazy to me how well these programs work, although they're not perfect. I would have killed for this superpower 25 years ago.

Think about the maths.

Most source separation systems include an "Other" stem which contains all the instruments that the algorithm can't identify. You can create your own Other stem rather than accept the stem created by the algorithm by subtracting the sum of the bass, drums, vocals etc stems from the original audio file. So when you combine the bass, drums, vocals etc stems with the "Other" stem that you have created, it will null with the original file.

That's my whole point. If the vox, drums and bass sound really good solo'd, then (with the present tech) they're unlikely to sum back to the original mix, regardless of whether you use your own generated (like you suggest) other stem or the one spat out by the algorithm.

That's what i would like to know about AudioShake, if they have made an advance over Spleeter in terms of splitting stems yet retaining the original whole when summed back.

Had a look into this subject over the weekend.

Very competitive market, and I suspect the big money is not going to be coming from us music types.

Excellent point on acoustic summing earlier, I don't think that would be so much of an issue on a modern multitrack recording.

It is interesting this tech, and in theory can only achieve perfection in the future.

Hi, late to the party.......

Just wondering what "UVR" is?

Thanks

Guessing...Ultimate Vocal Remover?

Thanks

Seems to be the cleanest separation I've used so far, and free. Pretty amazing.

Still looking for software that will create separate stems for lead and background vocals. In other words, give me a lead vocal only stem and a backing vocals only stem.

I looked into Ultimate Vocal Remover but it doesn't seem to have this feature.

Yes, but does it re-sum and null with the original?

I suspect not, and if not, it's useless for mastering.

EDIT: oops, is was referring to Audioshake.

Does Ultimate Vocal Remover null after re-sum?