Moises AI Mastering

[K-dub]

I don't know if folks have heard about Moises, but it is an AI site for musicians to access such tools. I signed up because I heard that they could do stem separation of various AI identification of instruments. It's like what they did w/ the Beatles recordings to modernize the mixes. It worked, but not flawlessly. Still, the technology is impressive. I had a couple of old mixes where the project files for my DAW got corrupted and the program allowed me to extract individual tracks for a remix. It's not perfect, but it is still impressive.

Another thing offered is AI mastering. Weirdly, I not use Izotope's AI mastering plug in my mixes, so I wondered if this was a superfluous offering I didn't need. Still, curious, I uploaded a file to Moises and let it do its thing. The results are below. Which of the two mixes appear more polished and professional sounding to you?

 

[Slouching Raymond]

There's not much in it. It was well done to start with.

 

[K-dub]

There's not much in it. It was well done to start with.

Thanks Ray! I do agree, in general, but when I experiment with such things, I ask myself a very simple question: Did it improve it or not?

Personally I was, at first, skeptical, but though the difference is slight (as noted), it IS better. What changed?

It's eq'd a little brighter. There's ever so slightly better definition in the bottom end. It's more gently compressed, so details like reverb tails stick out better. It's a smidge wider and deeper on the sonic stage as well.

SOME of that is, IMO, taste. I'm not sure I'd have made it more bright overall. But the other parts (wider/deeper/more detailed) are a slight step up from where I had it.

I will cede improvements where I hear them, even if slight.

But what the AI does that I really like is it can adhere easily to loudness standards. This example is mixed to -14 dbfs, which Soundcloud, Youtube et al set their digital preference to. -16dbfs is what AES recommends for general purpose loudness (broadcast spec). Apple Music and Spotify also want -16dbfs levels as well. To my ear, -14 is borderline overload. -16dbfs sounds more natural.

I had both levels printed so I can submit the right level out of the gate. If I'm attaching to a video that's going to Youtube, I can put their preferred audio format right in the product I upload to their site.

 

[VomitHatSteve]

I would be cautious of generative AI mastering. Unless it's using AI to generate an effects chain and applying that to your mix, it's actually generating an entirely new copy of the recording from scratch.

And the copyright precedent isn't firmly established yet. It's entirely possible that you are inadvertently signing your IP over to a) Moises themselves b) the public domain or c) whichever large music publisher most successfully sues the AI companies for illegally using their library for training

 

[K-dub]

I would be cautious of generative AI mastering. Unless it's using AI to generate an effects chain and applying that to your mix, it's actually generating an entirely new copy of the recording from scratch.

And the copyright precedent isn't firmly established yet. It's entirely possible that you are inadvertently signing your IP over to a) Moises themselves b) the public domain or c) whichever large music publisher most successfully sues the AI companies for illegally using their library for training

That's an interesting point, and frankly I have no idea what the answer would be. I'll have to look it up. Most companies prevent suit against themselves for usage, but I've never seen where the copyright is turned over. That would be a death knell for any business that engaged in that practice. No one sane would do business with them.

 

[Bulls Hit]

I honestly can't tell any difference between them on these ears and monitors. Some fairly intricate progressions and melodies in there, songwriting chops on display!

 

[PitchTempoAndriy]

One detail in the loudness comparison: -14 LUFS is an average loudness measurement, not a -14 dBFS peak ceiling. Spotify's Normal playback setting is -14 LUFS; its Quiet and Loud options use different levels. Their explanation is here: https://support.spotify.com/au/artists/article/loudness-normalization/

For your A/B, put the two downloaded masters on separate tracks in the same session, line them up, and lower the louder one until their perceived levels match. Then switch between the same passage, checking the reverb tails and low end you mentioned. Equal peak readings alone won't make that a fair loudness comparison. That should help separate an EQ/dynamics improvement from simply preferring the louder version.

 

[mark skinner]

First of all this is a well recorded and Really Nice powerful song.
My first impression of the original , was the piano sounded a little weak and flat sounding. It was wide enough but lacked any depth and was not immersive enough. The Moises version immediately corrected that and still left room in the middle to let the vox cut thru Nicely. I think a lot of it was some good EQ choices with fuller sounding low end and depth. Some of the percussion sounded a little hyped in the highs on both versions to me. I think the vox was a litte more powerful and pristine on the Moises version and with some "air" coming thru.
I believe this "might" be a nice tool to use for "reference" before making a final print on your own masters. In this case it would have directed you to the piano first for some changes if you liked the results. ms PS- I really wish -16 Lufs was the standard.

 

[K-dub]

One detail in the loudness comparison: -14 LUFS is an average loudness measurement, not a -14 dBFS peak ceiling. Spotify's Normal playback setting is -14 LUFS; its Quiet and Loud options use different levels. Their explanation is here: https://support.spotify.com/au/artists/article/loudness-normalization/

For your A/B, put the two downloaded masters on separate tracks in the same session, line them up, and lower the louder one until their perceived levels match. Then switch between the same passage, checking the reverb tails and low end you mentioned. Equal peak readings alone won't make that a fair loudness comparison. That should help separate an EQ/dynamics improvement from simply preferring the louder version.

Ah - my bad. I did conflate the two. Thank you for correcting me. I don't know that it's louder, though. They're pretty close. What the AI mix has got is a a slightly different EQ than the original and it appears ever so slightly more compressed.

It's one of the few times I've ever like the result out of an AI algorithm that I felt was an improvement - however slight - from my own ears.

 

[K-dub]

I honestly can't tell any difference between them on these ears and monitors. Some fairly intricate progressions and melodies in there, songwriting chops on display!

It is not a huge difference, K, but I hear it. It's mostly a slightly sharper clarity in the eq and a wider and deeper soundstage.

 

[K-dub]

I would be cautious of generative AI mastering. Unless it's using AI to generate an effects chain and applying that to your mix, it's actually generating an entirely new copy of the recording from scratch.

And the copyright precedent isn't firmly established yet. It's entirely possible that you are inadvertently signing your IP over to a) Moises themselves b) the public domain or c) whichever large music publisher most successfully sues the AI companies for illegally using their library for training

I did get around to checking. Their mastering is signal chain application only matched against their own reference tracks. You can load your own reference track if you want, but they are NOT employing generative AI w/ their mastering.

 

[K-dub]

First of all this is a well recorded and Really Nice powerful song.
My first impression of the original , was the piano sounded a little weak and flat sounding. It was wide enough but lacked any depth and was not immersive enough. The Moises version immediately corrected that and still left room in the middle to let the vox cut thru Nicely. I think a lot of it was some good EQ choices with fuller sounding low end and depth. Some of the percussion sounded a little hyped in the highs on both versions to me. I think the vox was a litte more powerful and pristine on the Moises version and with some "air" coming thru.
I believe this "might" be a nice tool to use for "reference" before making a final print on your own masters. In this case it would have directed you to the piano first for some changes if you liked the results. ms PS- I really wish -16 Lufs was the standard.

Wouldn't it be great if they let you know what the difference in the treatment actually was? That could be a terrific teaching tool. ALL of your observations were spot on, bud. I've gone in COMPLETELY the other direction from my youthful "crank it up" approach to mixing. You lose all the air around the sound.

 

[dobro]

I listened through earbuds, and the AI version sounded less earbud-tinny and had a more satisfying low end. Which is interesting, because that propbably means that the AI mixer is training itself on the stuff that prople listen to through earbuds. Earbuds, bud.

 

[K-dub]

I listened through earbuds, and the AI version sounded less earbud-tinny and had a more satisfying low end. Which is interesting, because that propbably means that the AI mixer is training itself on the stuff that prople listen to through earbuds. Earbuds, bud.

I do reference ear buds eventually,P. Same w/ cars. I don't do it while I'm mixing, but I always eventually listen on both systems. Truth told, the AI mastering sounds incredible on ear buds, bud.

If I were to lodge a complaint, and it is a minor one, the source material determines the treatment application. I've found that on softer material, the AI treatment is more ham handed. It's too loud for the application. I have to pull the treatment back and go into manual mode. I can still get it to work, but it takes vastly more massaging.

Inversely, the louder more complex content works better w/ the AI treatment. Both still need monitoring by human ears to know what's right, though.

 

[lraveton]

Agree with the point above about level-matching first, and you can do it by the numbers instead of by ear. Measure the integrated loudness of both downloads. If the Moises master comes out even 1 LU louder, turn it down by that much in your session and A/B again. A small level difference is enough to make one version sound brighter, wider and more detailed, which is a lot of what you described.

While you're at it, check the true peak of your -14 print. Spotify asks for -1 dBTP or lower at -14 LUFS, and a limiter that stops at -1 dB sample peak can still go over that between samples, or once the file is encoded to AAC or MP3. Your -16 print lines up with Apple Music, which normalises to about -16 LUFS.

I made a free LUFS meter on my site that measures several files at once in the browser, it reports integrated loudness, true peak and loudness range and nothing is uploaded: https://modwheel.net/tools/lufs-meter

 

[Maya Clars]

I don't know if folks have heard about Moises, but it is an AI site for musicians to access such tools. I signed up because I heard that they could do stem separation of various AI identification of instruments. It's like what they did w/ the Beatles recordings to modernize the mixes. It worked, but not flawlessly. Still, the technology is impressive. I had a couple of old mixes where the project files for my DAW got corrupted and the program allowed me to extract individual tracks for a remix. It's not perfect, but it is still impressive.

Another thing offered is AI mastering. Weirdly, I not use Izotope's AI mastering plug in my mixes, so I wondered if this was a superfluous offering I didn't need. Still, curious, I uploaded a file to Moises and let it do its thing. The results are below. Which of the two mixes appear more polished and professional sounding to you?

I think the source mix has so much life, the mastered one doesn't sound like anything to me, it doesn't evoke any emotion. I'm curious what makes you think it sounds impressive? Maybe I'm not paying attention to the right thing?

 

[VomitHatSteve]

I did get around to checking. Their mastering is signal chain application only matched against their own reference tracks. You can load your own reference track if you want, but they are NOT employing generative AI w/ their mastering.

Ah, that should be pretty safe then. You'd only be bound by their ToS. And as you pointed out, any company that tried to claim their user's songs like that would pretty quickly lose their entire customer base