The No.1 Website for Pro Audio
 Search This Thread  Search This Forum  Search Reviews  Search Gear Database  Search Gear for sale  Search Gearslutz Go Advanced
There's no AI in auto-mastering services Synchronisers/Clock Generators
Old 4 weeks ago
  #1
Lives for gear
There's no AI in auto-mastering services

Do prove me wrong but first demo this; Stereo Tool - Download

This is a broadcast processor; It has auto eq, all the dynamic processing you need, audio cleaning (de-clip, humm etc.), stereo widening...likely way more capable than any current auto-mastering algo's.

If the website gives you options for your mastering request (Genre, loudness etc.) basically you're the AI here cause all you're doing is selecting a preset for the software

One gets his/her music mastered to withstand a processor like this, doing this double (if your music is successful) is basically torturing your mix. In a nutshell this is why auto-mastering is plain wrong.
Old 4 weeks ago
  #2
Gear Maniac
I think anyone who considers themselves "pro" or "prosumer" is smart enough to realize this. If you're making enough money and are serious about your music, you should master your own music, or pay a human to do it. I only would give budget hobby people full carte blanche to use auto-mastering services like Landr without a shade of ridicule. They deserve a cheaper alternative, albeit one of lesser quality.

If you're serious about your music and have some extra money, you shouldn't ever use an auto-mastering service. I use Doctor Mix, a human-based mastering service and they only charge $41 a song for mastering. It's budget mastering for those who still want a human involved, and they're proven (to me anyways). I found them the traditional way, I looked at the credits of a vinyl record I was listening to, and saw they were listed as the mastering service. I even tweeted the band asking about the service and got a response. When I get a great master from them, I try to match it myself and learn a little more about mastering. Some songs I do master myself, but the ones I really want to sound good I pay someone for that.
Old 4 weeks ago
  #3
Lives for gear
 
robert82's Avatar
For a hobbyist like me, Izotope Ozone 7 is enough. You start with one their presets, then you go through each of the modules and tweak them until you get what sounds good to you. I would say that having a good monitoring setup is more important, so that what you are tweaking ends up sounding good elsewhere. I've messed with LANDR a little, and I'm never just thrilled with what I get back from them.

If I ever were to put a song/album out for release, I would absolutely hire a real ME.
Old 4 weeks ago
  #4
Lives for gear
 
Ben F's Avatar
 

Verified Member
Quote:
Originally Posted by robert82 View Post
For a hobbyist like me, Izotope Ozone 7 is enough. You start with one their presets, then you go through each of the modules and tweak them until you get what sounds good to you. I would say that having a good monitoring setup is more important, so that what you are tweaking ends up sounding good elsewhere. I've messed with LANDR a little, and I'm never just thrilled with what I get back from them.

If I ever were to put a song/album out for release, I would absolutely hire a real ME.
The mastering assistant in Ozone 8 is arguably better than an automated service.
Old 4 weeks ago
  #5
Gear Maniac
 
B Elgin's Avatar
 

Quote:
Originally Posted by Ben F View Post
The mastering assistant in Ozone 8 is arguably better than an automated service.
I'm with you on this. AI is an amazing tool for assistance and suggestions as long as a human operator has the final call on whether or not they help and decides how far to go in any given direction (assuming we're talking about art and music making contexts).

Sometimes the suggestions will surprise you and you learn something or try something useful that's way outside of your customary habits.

Logic based scripting can't really be considered AI but it's also a useful tool for doing repetitive tasks and organizational things.

Both of these things are truly useful and hold a lot of potential - unlike the auto-mastering services of today.
Old 4 weeks ago
  #6
Gear Addict
 

Quote:
Originally Posted by robert82 View Post
For a hobbyist like me, Izotope Ozone 7 is enough. You start with one their presets, then you go through each of the modules and tweak them until you get what sounds good to you. I would say that having a good monitoring setup is more important, so that what you are tweaking ends up sounding good elsewhere. I've messed with LANDR a little, and I'm never just thrilled with what I get back from them.

If I ever were to put a song/album out for release, I would absolutely hire a real ME.
Sounds like the right approach to me.
Old 4 weeks ago
  #7
Lives for gear
The mastering assistant in Ozone 8 is still very far from being ideal, even as a starting point.

Any dynamics processing it has made will be affected by what is added before it+Anything added after will mess with the limiters auto-adjustment (and as we know typically the limiter goes last). So unlike how its marketed, whats left is leaving all the work to ozone 8...Then we go back to the original point of this topic, I see it no different than the broadcast processor.

The "machine" doesn't hear anything, only see's what you see on the analyzer and reacts accordingly. For example, it can't tell if a vocal is buried or too upfront to your liking in a mastering context etc.

As being one of them, I do encourage any hobbyist doing their best with the tools they prefer but not rely on them completely. It takes time and patience but choosing not having to select a preset is something every audio enthusiast must experience.
Old 3 weeks ago
  #8
Lives for gear
 

The title of this thread, "There's no AI in auto-mastering services" is just flat out wrong.

These services are not using PRESETS. If there are style/flavor/loudness options provided that is for REFERENCE the same way a human ME uses reference tracks.

The "AI" part analyzes the dynamics, spectral balance, average levels, stereo width, etc., and adjusts all those elements to closely match the reference.

Wow.
Old 3 weeks ago
  #9
Gear Maniac
 
Magnus Lindberg's Avatar
 

Quote:
Originally Posted by Nonlinear View Post
The title of this thread, "There's no AI in auto-mastering services" is just flat out wrong.

These services are not using PRESETS. If there are style/flavor/loudness options provided that is for REFERENCE the same way a human ME uses reference tracks.

The "AI" part analyzes the dynamics, spectral balance, average levels, stereo width, etc., and adjusts all those elements to closely match the reference.

Wow.
My LANDR Test
Old 3 weeks ago
  #10
Lives for gear
Quote:
Originally Posted by Nonlinear View Post
The title of this thread, "There's no AI in auto-mastering services" is just flat out wrong.

These services are not using PRESETS. If there are style/flavor/loudness options provided that is for REFERENCE the same way a human ME uses reference tracks.

The "AI" part analyzes the dynamics, spectral balance, average levels, stereo width, etc., and adjusts all those elements to closely match the reference.

Wow.
When a human uses a reference, ideally he/she does not let the integrity of their mix affected in a negative way.

For example; if the bass in the reference is huge and cannot be adapted to their mix, they don't mud their mix for the sake of trying to be identical to the reference.

As explained from my previous example, an algo can't do this. It will choose to sacrifice the mix for the sake of the reference.
Old 3 weeks ago
  #11
Lives for gear
 
Ben F's Avatar
 

Verified Member
Quote:
Originally Posted by Nonlinear View Post
The title of this thread, "There's no AI in auto-mastering services" is just flat out wrong.

These services are not using PRESETS. If there are style/flavor/loudness options provided that is for REFERENCE the same way a human ME uses reference tracks.

The "AI" part analyzes the dynamics, spectral balance, average levels, stereo width, etc., and adjusts all those elements to closely match the reference.

Wow.
That’s not AI. It’s machine learning, and has been used in automation and broadcast processing for decades.
Old 3 weeks ago
  #12
Lives for gear
 

Quote:
Originally Posted by Ben F View Post
That’s not AI. It’s machine learning...
Took me 2 seconds on Google to find this:

"Artificial intelligence (AI) is an area of computer science that emphasizes the creation of intelligent machines that work and react like humans."
Old 3 weeks ago
  #13
Lives for gear
 

Quote:
Originally Posted by cooker View Post
When a human uses a reference, ideally he/she does not let the integrity of their mix affected in a negative way.

For example; if the bass in the reference is huge and cannot be adapted to their mix, they don't mud their mix for the sake of trying to be identical to the reference.

As explained from my previous example, an algo can't do this. It will choose to sacrifice the mix for the sake of the reference.
Point taken, but that's also partly the choice of a WRONG reference. A human ME wouldn't use an EDM track as a reference for an acoustic vocal/piano track either, would he?
Old 3 weeks ago
  #14
Lives for gear
Quote:
Originally Posted by Nonlinear View Post
Point taken, but that's also partly the choice of a WRONG reference. A human ME wouldn't use an EDM track as a reference for an acoustic vocal/piano track either, would he?
I was actually referring to more of a masking effect.

In a mix, an analyzer can detect this (Izotope neutron, though it still lets the human to fix it) but on a stereo file this can't be seen on any sort of analyzer as it doesn't show itself as a peak. But a ME can hear this and fix it or worst case request stems.

Speaking of ozone I actually tried this too; Use ozone's master assistant to analyze and process a track. Then don't disable it, open another one on the same track and analyze again while the previous instance is working (same area, same settings). The second instance will process (and wasn't so minimal either in my experience) although it shouldn't, I bet this would happen with any auto-mastering site too. So their reference, preset whatever we call it doesn't work so well even in its own limitations either.

Thats why to me the "AI" we speak of today isn't a tailor that works for the individual, but tries to fit the same dress to everyone.

This is something we'd request from a broadcast processor if we own a TV station or radio, not for the final process on our personal music we worked so hard for.
Old 3 weeks ago
  #15
Lives for gear
 
Ben F's Avatar
 

Verified Member
Quote:
Originally Posted by Nonlinear View Post
Took me 2 seconds on Google to find this:

"Artificial intelligence (AI) is an area of computer science that emphasizes the creation of intelligent machines that work and react like humans."
It’s a misunderstanding of the term.

“Most of what is being called “AI” today, particularly in the public sphere, is what has been called “Machine Learning” (ML) for the past several decades. ML is an algorithmic field that blends ideas from statistics, computer science and many other disciplines (see below) to design algorithms that process data, make predictions and help make decisions. In terms of impact on the real world, ML is the real thing, and not just recently. ”

Artificial Intelligence — The Revolution Hasn’t Happened Yet
Old 3 weeks ago
  #16
Lives for gear
 

Quote:
Originally Posted by cooker View Post
Speaking of ozone I actually tried this too; Use ozone's master assistant to analyze and process a track. Then don't disable it, open another one on the same track and analyze again while the previous instance is working (same area, same settings). The second instance will process (and wasn't so minimal either in my experience) although it shouldn't, I bet this would happen with any auto-mastering site too. So their reference, preset whatever we call it doesn't work so well even in its own limitations either.
OK, that test you did can be explained. The Ozone "Mastering Assistant" doesn't compare the track to a single reference - it compares it to multiple references and picks an average that most closely matches the source track.

In other words, it IS giving you that variability you seek. They designed it that way on purpose for the very reason you state.

The reason the second instance of the plugin came up with different settings is because you were feeding it a different VERSION of audio that had been modified by the first instance.
Old 3 weeks ago
  #17
Lives for gear
Quote:
Originally Posted by Nonlinear View Post
OK, that test you did can be explained. The Ozone "Mastering Assistant" doesn't compare the track to a single reference - it compares it to multiple references and picks an average that most closely matches the source track.

In other words, it IS giving you that variability you seek. They designed it that way on purpose for the very reason you state.

The reason the second instance of the plugin came up with different settings is because you were feeding it a different VERSION of audio that had been modified by the first instance.
Correct, but this test also reveals it will always do some level of processing to whatever its fed (cause we fed the "perfect mix" to the algo with this test).

So while it fix things, takes out the nice things too cause "less is more" doesn't apply for the auto-mastering algo.
Old 3 weeks ago
  #18
Lives for gear
 

Quote:
Originally Posted by cooker View Post
Correct, but this test also reveals it will always do some level of processing to whatever its fed (cause we fed the "perfect mix" to the algo with this test).
No, you fed it a mix, mastered to one set of references then processed that result to another set of references.

10 different human MEs would likely produce 10 different masters as well. What one thinks is the "perfect mix" is mostly a matter of opinion.

I'm not defending Robot mastering - I'm just saying as human MEs we better come up with some real good reasons to justify our price and existence because it seems like the robots are winning the war.
Old 3 weeks ago
  #19
Lives for gear
 
SmoothTone's Avatar
 

Quote:
Originally Posted by Nonlinear View Post
I'm just saying as human MEs we better come up with some real good reasons to justify our price and existence because it seems like the robots are winning the war.
Firstly, I didn't know I was in a war with the robots. And if someone feels that they are competing with automated mastering, then they are aiming for the wrong clients IMHO. Totally different markets.

Secondly, the robots have investors and huge marketing budgets. However, they probably aren't even turning a profit yet. Totally different business model to the majority of professional MEs.
Old 3 weeks ago
  #20
Lives for gear
Quote:
Originally Posted by Nonlinear View Post
No, you fed it a mix, mastered to one set of references then processed that result to another set of references.

10 different human MEs would likely produce 10 different masters as well. What one thinks is the "perfect mix" is mostly a matter of opinion.

I'm not defending Robot mastering - I'm just saying as human MEs we better come up with some real good reasons to justify our price and existence because it seems like the robots are winning the war.
To not prolonging the argument, I do agree ME's also can be an issue. Just a few days ago I posted here about my friends experience as it took 3 revisions (the final one he had to attend) to complete his master close enough to how he wanted. And it was a top-notch studio where the gear was worth probably a small house.

But the final master and his mixes analyzer average response was near (L&R identical, only noticed the differences on the side signal) identical as the ME I guess used more stereo image and saturation tricks than eq'ing. In this case robot-mastering wouldn't help either.

The cinema industry to my knowledge did standardize this (or at-least dolby). There is no mastering stage for movie music, they pass the audio through a device similar to the broadcast plug-in I shared (of-course way more advanced). If I remember correct I noticed it on one of the Batman soundtracks (orchestral theme music) as the response was very flat when analyzed.

I like the depth, balance of such tracks but most musical terms (warmth for example) we used today are lacking on the other hand. It was the reason I experimented a lot with match-eq's and finally understood it not the way to go cause the nice things were demolished too for the sake of freq. balance.
Old 3 weeks ago
  #21
Lives for gear
Quote:
Originally Posted by cooker View Post
I do agree ME's also can be an issue. Just a few days ago I posted here about my friends experience as it took 3 revisions (the final one he had to attend) to complete his master close enough to how he wanted. And it was a top-notch studio where the gear was worth probably a small house.
Just wanna say...for me, 3 rounds of revisions is very rare. But it happens. I don't think that makes ME's "an issue". For unattended sessions where all the communication is via email, it's inevitable that there are occasionally going to be times where that communication isn't 100% clear. And/or the track is a style of music the ME isn't 100% familiar with.

The changes over those 3 rounds of revisions could likely have been sorted very quickly in an attended session, and it would just be a normal part of the process. Try this, try that...you like that? Great, we'll go with that.

Online/unattended...sometimes it just takes a little more work to figure out what people are looking for.

At least 90% of the time, clients are totally happy with the first thing I send them (which still amazes me). I'm sure it's the same for the other pros here. That's a pretty good batting average. You can't expect a home run every single time.
Old 3 weeks ago
  #22
Lives for gear
 

Quote:
Originally Posted by SmoothTone View Post
Firstly, I didn't know I was in a war with the robots.
Then you better take a look around!

Quote:
Originally Posted by SmoothTone View Post
And if someone feels that they are competing with automated mastering, then they are aiming for the wrong clients IMHO. Totally different markets.
This comment actually sounds a little bit snobbish. If you have a client base with big money to spend on mastering consider yourself lucky - you have a niche market.

Quote:
Originally Posted by SmoothTone View Post
Secondly, the robots have investors and huge marketing budgets.
And that's a disadvantage? Who do you think sells more beer, Budweiser or some little guy?

Quote:
Originally Posted by SmoothTone View Post
...they probably aren't even turning a profit yet.
I believe you are wrong about that. There are more of these systems now than a couple of years ago. If it wasn't a money maker they'd have been long gone.

Quote:
Originally Posted by SmoothTone View Post
Totally different business model to the majority of professional MEs.
Exactly! And to a growing population of music MAKERS that means AFFORDABLE.

Last edited by Nonlinear; 3 weeks ago at 07:04 PM..
Old 3 weeks ago
  #23
Lives for gear
 
SmoothTone's Avatar
 

Quote:
And if someone feels that they are competing with automated mastering, then they are aiming for the wrong clients IMHO. Totally different markets.
Quote:
Originally Posted by Nonlinear View Post
This comment actually sounds a little bit snobbish. If you have a client base with big money to spend on mastering consider yourself lucky - you have a niche market.
It did sound a bit snobbish. What I meant was that if you are going to try to compete with automated mastering, you're going to lose. LandR, for example, explicitly targets users that would otherwise not use mastering. Like I said, totally different markets. What the robots offer and what professional MEs offer are very different.


Quote:
Secondly, the robots have investors and huge marketing budgets.
Quote:
And that's a disadvantage? Who do you think sells more beer, Budweiser or some little guy?
No, it's an advantage. Most mastering services - even the big ones - have nowhere near the marketing budget as the robots.

Quote:
...they probably aren't even turning a profit yet.
Quote:
I believe you are wrong about that. There are more of these systems now than a couple of years ago. If it wasn't a money maker they'd have been long gone.
LandR has raised 10M from investors (primarily Warner) since 2012. Their current estimated yearly revenue is 1.5M. You do the maths. But like I said, it's a different business model. IIRC Spotify only started turning a profit in the last couple of years. It's just how things work at this level. LandR is set to make a lot of money (they wouldn't have attracted further investment if they weren't), but they're not quite there yet.
Old 3 weeks ago
  #24
Quote:
Originally Posted by cooker View Post
Do prove me wrong but first demo this; Stereo Tool - Download

This is a broadcast processor; It has auto eq, all the dynamic processing you need, audio cleaning (de-clip, humm etc.), stereo widening...likely way more capable than any current auto-mastering algo's.

If the website gives you options for your mastering request (Genre, loudness etc.) basically you're the AI here cause all you're doing is selecting a preset for the software

One gets his/her music mastered to withstand a processor like this, doing this double (if your music is successful) is basically torturing your mix. In a nutshell this is why auto-mastering is plain wrong.

As you noted, this tool has a very specific audience and objective, and has absolutely nothing in common with "AI" buzzword bingo. It's not mentioned once on their site, is someone hallucinating?


Every classic radio or TV broadcaster on the planet uses such tools. They have very specific issues and demands to address. And they have been using exactly the same approach now for over 60 years, with immense monetary success. This is not AI, this is a broadcast chain made to drive an antenna.


"Software AI" is really just a set of tools that goes beyond primitive arithmetic and branching. Tools that try to adapt to and take into account the perceptual and intellectual capabilities of human operators, tools that make "sense" to humans. Tools that feel ergonomic and intuitive. All this smells "intelligent".

We have great difficulties understanding our own "system" (human brain and nervous system), or even defining what intelligence is. Let's maybe first complete this job before we fantasize!

In software dev, "AI" describes a category of tools. Outside of it, it's used to scare and fool the laymen.
Old 3 weeks ago
  #25
Lives for gear
 
duvalle's Avatar
 

Yes - „AI“ seems to be a buzzword in many cases.

Without going into detail … AI is statistic / machine learning.
ML is around for decades … ML never took of back in the days because of the lack of big data and big compute. In some cases companies now have both - big data & compute. In only a few cases the investment in ML/AI makes commercial sense - like with Google AD-Services and such …

So how could AI in music/mixing work?

Assuming Izotope (or others) used a AI learning method called „supervised learning“
(Supervised learning: predicting an output variable from high-dimensional observations — scikit-learn 0.20.0 documentation)
it only works when you have, simply said, both the information and an „answer“ in your dataset.
Meaning: You have a selection of songs and describe/rate the songs in some way and then let the AI learn from both informations. Like: Songs 1-100 are the best mixed dance songs. The AI now can analyze the audio in as many ways needed and try to find statistic correlations within this data and learn those. So when you feed a new song into a pre-learned AI it tries to find matching patterns and returns an assumption (like this song is to 86% perfectly mixed). It’s a good thing to remember with AI: the results ONLY reflect the quality of the datasets and the method used for learning.

Anyway - I don’t think in the case of Izotope the AI/ML part is pure snake oil …
Analyzing audio is their main business for years and all the AI tools they would need are available basically for free now.
Old 3 weeks ago
  #26
Lives for gear
I remembered a song shared in another forum where every member loved it. Terrible mix, beginner homestudio written all-over it.

They applied a heavy high-cut likely by mistake on the mastering stage, we would take it as mud but the fans I think took it as dark sounding. The song was a slow-grunge/Brit. pop sort of thing so it worked great :D

Likely a machine would try to fix it and ruin everything.
Post Reply

Welcome to the Gearslutz Pro Audio Community!

Registration benefits include:
  • The ability to reply to and create new discussions
  • Access to members-only giveaways & competitions
  • Interact with VIP industry experts in our guest Q&As
  • Access to members-only sub forum discussions
  • Access to members-only Chat Room
  • Get INSTANT ACCESS to the world's best private pro audio Classifieds for only USD $20/year
  • Promote your eBay auctions and Reverb.com listings for free
  • Remove this message!
You need an account to post a reply. Create a username and password below and an account will be created and your post entered.


 
 
Slide to join now Processing…
Thread Tools
Search this Thread
Search this Thread:

Advanced Search
Forum Jump
Forum Jump