Be very wary about thinking of "AI" as a selling point, since it is only as intelligent as the programmer. Nice term to throw out, but in reality it is more likely an analysis tool followed by a preset choice. Beyond that point it is unknown, but if a human never touches it, that is all there is.

That's not accurate - Landr uses Machine learning. If you look at the theory, machine learning is as accurate as the DATA its given rather than as intelligent as the programmer. Unlike conventional programming where you write an algorithm to solve a problem, machine learning is a branch of data science where the data is used to solve the problem. One definition is "Machine learning is a field of study that gives computers the ability to learn without being explicitly programmed" There is a brief overview here.
 
Machine learning is being used widely for all kinds of applications today including DSP, studying the brain, medicine and finance. As I mentioned earlier, our Vocal sync feature was designed using machine learning to understand how to analyze and extract data from a huge subset of vocals samples.
I'm not sure why there is so much resistance and doubt about applications that use modern science to solve problems. It sounds like a repeat of fears people had about computers taking over when I first started out in programming 35 years ago :) Did that happen - of course not, people just use them to work more efficiently now. 

skitch_84
Here's what Steven Slate had to say about Landr and this kind of technology. Spoiler: he was actually very optimistic about it. 
https://www.gearslutz.com/board/10438197-post77.html
Followed are plenty of people arguing against his statements, so have fun reading!

That's actually a great thread. I hadn't seen it - its great that Steven Slate is so open minded and gets this technology. Some great quotes from that thread that resonated with me:
 
"Ultimately, if someone uses LANDR or any similar service and LIKES THE RESULT, then they got their money's worth. Perhaps they could have gotten a better result from a mastering engineer... perhaps that would have cost them money they didn't have... and... perhaps they would have gotten WORSE results from a mastering engineer. Human Mastering doesn't mean quality either. I've had some mastering engineers butcher my mixes, others have made them sound amazing."
 
"Abstraction and innovation is to often confused with "oversimplification". But this increase in efficiency is what drove humanity since the beginning. This efficiency is what secures your rent, cares for your health and increases your living standard (which is much higher than what we had just 20 years ago, isn't it?)."
 

I tried it out yesterday. I was disappointed that you only get a 30 second preview to decide if you want to make the master. I, too, feel like the price for quality ratio is kind of poor, I would prefer a different business model, at least less expensive. I also use Ozone and feel like their presets and my ears do a pretty good job, but I'm not at all opposed to the LANDR concept, just the execution. It also took quite a while to load my 2:30 length song, which surprised me, my internet is very fast. But I'm going to play around with it a while, try a few different use cases (such as outlined in the ezine, which I agree is a must-read) before I decide if I wish to purchase a subscription or pay by the song for a finished mastered version. 

I just uninstalled it.  For my uses, I don't need or want it.

Noel Borthwick [Cakewalk]
That's not accurate - Landr uses Machine learning. If you look at the theory, machine learning is as accurate as the DATA its given rather than as intelligent as the programmer....our Vocal sync feature was designed using machine learning to understand how to analyze and extract data from a huge subset of vocals samples.

In the Vocal sync case I can see the mechanism behind analyzing many vocal examples to find out common characteristics in these files.
 
But with LANDR, what is it learning? Who is giving it feedback to say, "That sounds better, what you did this time is right"?
 
It wouldn't surprise me if it turns out all it does is figure out genre based on tempo, harmonic density, etc. and then applies a preset, possibly even on a commercially available mastering tool.

I for one have loved using Cakewalk since Sonar 5. I know some of you go way back.  I loved Sonar for its simplicity and what it did best, record great tracks.  Pretty decent effects and to run smooth with no pop's, hiss or down time.  Lately, all the added gizmo's and third party collaborations, lite version this and lite version that, all added with the intention of upgrading to the full version. I can do without all that. Just give me a good simple product and I can buy whichever add on I like and choose.  I love the forward thinking from Cakewalk but don't make this great product too complicated.  Just my two cents.

Soop

Noel Borthwick [Cakewalk]
That's not accurate - Landr uses Machine learning. If you look at the theory, machine learning is as accurate as the DATA its given rather than as intelligent as the programmer. Unlike conventional programming where you write an algorithm to solve a problem, machine learning is a branch of data science where the data is used to solve the problem. One definition is "Machine learning is a field of study that gives computers the ability to learn without being explicitly programmed" There is a brief overview here.
 
Machine learning is being used widely for all kinds of applications today including DSP, studying the brain, medicine and finance. As I mentioned earlier, our Vocal sync feature was designed using machine learning to understand how to analyze and extract data from a huge subset of vocals samples.
I'm not sure why there is so much resistance and doubt about applications that use modern science to solve problems. It sounds like a repeat of fears people had about computers taking over when I first started out in programming 35 years ago :) Did that happen - of course not, people just use them to work more efficiently now. 

Noel, I don't think people doubt the science, at least I don't, but there is a difference between a clear view and a short distance.

By this I mean that just because you apply decision trees or Bayesian or an attribute importance algorithm, doesn't mean the machine will solve the problem reliably. There are models and programmers involved not just data fed into a perfect machine. Lots of places for flaws.

We can get there if we try is not the same as we have arrived.

I am hopeful, just careful.

cparmerlee

Anderton
Whether that's worthless or not depends on your gig.

Would you give a 192 kb MP3 to a client?  I wouldn't.

 
I guess you didn't follow the link, so here's the pertinent part:
 
Well if there's one thing I've learned from all the back-and-forth here, it's that apparently there aren't a lot of professional mastering or mix engineers who frequent this forum. I guess they're off making money and doing projects.
 
MP3s are the lingua franca for giving quick demos to clients. They don't want uncompressed WAV files to put on their iPhones to listen to mixes. They don't want something with no EQ or multiband limiting, because they're going to be listening in context with other material. They want a ballpark approximation of a finished product while they listen to rough mixes to figure out what to do next.
 
A lot of mix engineers do not consider themselves mastering engineers, with good reason; they're different skill sets. The client can either do the supremely stupid thing of paying a mastering engineer several hundred dollars to master a mix that will never be released so they can listen to it on their smart phone, pay the mix engineer to do a mastering job at the usual rates, or with LANDR, get a ballpark approximation for very little $$ in a couple minutes while the rest of the band is fidgeting and waiting to go home.
 
 

jpetersen

Noel Borthwick [Cakewalk]
That's not accurate - Landr uses Machine learning. If you look at the theory, machine learning is as accurate as the DATA its given rather than as intelligent as the programmer....our Vocal sync feature was designed using machine learning to understand how to analyze and extract data from a huge subset of vocals samples.

In the Vocal sync case I can see the mechanism behind analyzing many vocal examples to find out common characteristics in these files.
 
But with LANDR, what is it learning? Who is giving it feedback to say, "That sounds better, what you did this time is right"?
 
It wouldn't surprise me if it turns out all it does is figure out genre based on tempo, harmonic density, etc. and then applies a preset, possibly even on a commercially available mastering tool.

 
This is a gross oversimplification but machine learning, like classical numerical analysis takes an input, parametrizes it and then applies a transformation function to produce a new output.
 
I don't know the specifics of what Landr does obviously, but its quite likely the transform does a spectral analysis of the audio and then using that data maps the input parameters to multiple plugin parameters that reside in the mastering chain. I expect that the parameter mapping would be done dynamically using automation so as to react to changes in the audio content on the timeline. Its not simply applying a preset but dynamically changing the DSP to produce the final result, much like a human mastering engineer would do by listening to the audio and making decisions to correct the frequency balance or loudness. 
 
The machine learning is what generates the mapping from the source data to the final parameters. Machine learning essentially builds a data driven algorithm. In the learning process you feed it thousands or more samples of human curated data sets of input and output data.  The software learns from that data and generates the transform that predicts the mapping. The more data you feed it the closer to reality it gets. 
 
It's exactly the same approach we used for vocal sync just applied to solve a different problem. If you aren't familiar with machine learning, its quite astonishing how accurate the results get over time especially as you give it more exception cases and boundary conditions.