Although I haven't much use for LANDR I say the more the merrier.  As above, there are home recordists w/o the time to learn mastering, or new to recording and just don't want their songs self-butchered in mastering.  Many SONAR users are more musician than engineer, so this can be a help.
 
Doesn't mean you have to use it, and I'm sure it didn't cost Cake a penny, or much anyway and some users will benefit.  I don't see a downside at all, except for the surprise of a big honking shortcut on my screen.
 
@

Noel Borthwick [Cakewalk]
If you look at the theory, machine learning is as accurate as the DATA its given rather than as intelligent as the programmer.

That only works in a closed loop system where the "AI engine" can make judgments (actually, try things randomly) and then determine QUALITATIVELY whether the outcome is better or worse.  That's the rub.  If we knew algorithmically how to judge "goodness", then we really wouldn't need an AI engine.
 
That's why 99.9999% of the AI promises made over the last 50 years have ended in failure.  In 1985,people were talking about the imminent times when medicine would be practiced by AI.  The computer would have much better knowledge than the very best doctors.  But not only do we still not have AI practicing medicine, Siri still can't carry out the most basic request with more than about 20% accuracy.
 
I would love to have a magic button to make all my mixes sound better.  So far, I'm not seeing anything very special in this LANDR stuff that I can't do in the same amount of time with Ozone.

Anderton
I guess you didn't follow the link,

Yes, I did read that.  But I asked you a very specific question.  Would you provide a client an MP3 file crunched below a 200 kb bitrate?  I would not, regardless of whether it was targeted for an iphone or not.
 
If your answer is "yes", that's OK with me.  I just don't get the point of going to all the trouble to make the music sound its best and then crunch the life out of it.

cparmerlee
That only works in a closed loop system where the "AI engine" can make judgments (actually, try things randomly) and then determine QUALITATIVELY whether the outcome is better or worse.  That's the rub.  If we knew algorithmically how to judge "goodness", then we really wouldn't need an AI engine.

 
Its definitely possible to judge "goodness" using machine learning. Its based on how much data is fed into the learning engine. By feeding the engine thousands of samples of human curated "good" masters the system learns to predict what is considered to be good. There will always be exception cases where the system fails but its surprisingly accurate. As explained even in the vocal sync case the system learned to make good predictions as it was fed more curated data.

jpetersen
But with LANDR, what is it learning? Who is giving it feedback to say, "That sounds better, what you did this time is right"?

PRECISELY.  I have some knowledge of AI systems with missile guidance.  In a system like that, instruments are able to measure the target,  They can provide instant feedback as to whether the "AI" maneuvers are helping to reach the target or working against that objective.  Over time, such a system can "learn" what maneuvers under what circumstances are likely to be most productive.
 
There is no such feedback loop in this mastering thing.

jpetersen
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.

 
Once again, PRECISELY on the mark.  I question whether there is even any real "AI engine" involved in the process.  It does seem to be a case of software analyzing the sound file and selecting a preset that fits a profile.  It is not computer learning without that feedback loop.  At best, one might stretch to call it an "expert system" because presumably there are some mastering experts involved to help fit different sound patterns to different presets.
 
If done well, that could be useful, but the examples I have seen so far don't seem to do much more than compressing and jacking up the dB level.  I think a person could get pretty much the same results simply by adding a multi-band compressor and the Concrete Limiter to the master bus in their mix.

Noel Borthwick [Cakewalk]

cparmerlee
That only works in a closed loop system where the "AI engine" can make judgments (actually, try things randomly) and then determine QUALITATIVELY whether the outcome is better or worse.  That's the rub.  If we knew algorithmically how to judge "goodness", then we really wouldn't need an AI engine.

 
Its definitely possible to judge "goodness" using machine learning. Its based on how much data is fed into the learning engine. By feeding the engine thousands of samples of human curated "good" masters the system learns to predict what is considered to be good. There will always be exception cases where the system fails but its surprisingly accurate. As explained even in the vocal sync case the system learned to make good predictions as it was fed more curated data.

I take your point.  However goodness is quite a subjective thing, and varies widely across musical genres.  In theory, with enough curation, what you say is possible.  From what I have seen so far, it ain't ready for prime time.  It may provide a quick-and-dirty improvement more times than not, which could be worth something to some people.  But again, I am not hearing results that seem to be substantially better than simply dropping a MB compressor on the master bus and then using Concrete Limiter to crank the whole thing up to max dB (which itself isn't necessarily a good thing in the first place.)  At the right price and the right speed and right quality of output files, I might use the thing on occasion.  But IMHO, this offering is wrong on all three counts.

cparmerlee

jpetersen
But with LANDR, what is it learning? Who is giving it feedback to say, "That sounds better, what you did this time is right"?

PRECISELY.  I have some knowledge of AI systems with missile guidance.  In a system like that, instruments are able to measure the target,  They can provide instant feedback as to whether the "AI" maneuvers are helping to reach the target or working against that objective.  Over time, such a system can "learn" what maneuvers under what circumstances are likely to be most productive.
 
There is no such feedback loop in this mastering thing.

 
Actually, there's at least one that I know of. There are three levels of intensity. I would never use the high level but then again, I don't master Metallica. To my ears the low level "does no harm" on pretty much anything, and the medium level may or may not work, depending on the program material. If 60% of the users choose the low one, 30% the medium one, and 10% the high one, then if the people running LANDR took that into account they would bump up the intensity of the middle one somewhat, and insert another option that splits the difference between low and medium to round out the three options.
 
Also, I highly doubt that no humans are involved in any aspect of LANDR. I assume they have a database of well-mastered material they use for comparison. I also suspect they do the "this call may be monitored for quality assurance" thing. If masters in particular genres are constantly being rejected, it would be foolish if a human didn't listen to them and try to analyze what doesn't work. Then they can program the algorithms to take a different path.
 
Here's an example. When I master, I always look for resonances that are a constant throughout a piece. This is most common with acoustic projects and live recordings due to room resonances. I then apply a notch to deal with this. If this isn't in LANDR's algorithm, and they find that acoustic and live recordings keep getting their masters rejected, someone might tell LANDR to look for resonances above a certain peak value that don't change with different notes and keys.
 
The article I linked to about "sonic signatures" is very much a part of understanding mastering. And, anyone who has asked for more analytics in SONAR, like being able to detect average levels, understands that many aspects of audio (note I didn't say music) CAN be quantified in the mastering stage. The human ear is much less sensitive to level variations than pitch variations, which is why something like Pleasurize's attempts to end the "loudness wars" are commendable. They provide a measurement tool that quantifies dynamic range, and makes recommendations on what dynamic range to aim for with particular styles of music. In conjunction with understanding the "sonic signature" of different kinds of music, this is something a machine could handle easily. It could also handle looking for frequency response anomalies, like excessive buildup in the 300 - 400 Hz range, which is a common problem with many mixes. 
 
I don't believe we're at a stage where algorithm mastering can replace a good mastering engineer, and although "never say never," I don't see that happening any time soon. But there are a lot of people out there who don't know how to master and can't afford to pay for it. Getting an acceptable mastering job is a big improvement over no mastering at all, and if they want a stellar mastering job, then can pay a couple hundred bucks for it instead of $10. 
 

cparmerlee
I think a person could get pretty much the same results simply by adding a multi-band compressor and the Concrete Limiter to the master bus in their mix.

 
Not if they didn't have at least 8 stages of EQ, and probably some imaging. But that's not the point. The point is that they'll only get the same results if they know how to set the EQ, dynamics, and imaging to create an acceptable master. That's not a given. If they know how to master, then they don't need to hire a mastering engineer, nor do they need to use LANDR.
 
If they get the results they want with a preset from Ozone, great. But I've yet to find a preset that replaces what a mastering engineer can do; the preset doesn't know to look for resonances or frequency buildups. It seems to me that LANDR sits somewhere between calling up a preset, and calling a mastering engineer.
 
 

Noel Borthwick [Cakewalk]
Its definitely possible to judge "goodness" using machine learning. Its based on how much data is fed into the learning engine. By feeding the engine thousands of samples of human curated "good" masters the system learns to predict what is considered to be good.

OK, I can accept that.
 
And it turns out genre detection is, indeed, central to what they do.
 
Quoting Justin Evans from LANDR:
"Genre detection is something I can really get enthusiastic about discussing. It's been around in a pretty robust form for a number of years....Genre recognition isn't rocket science, similar processes are driving ... facial recognition in photos, recommendation engines, etc...we are hoping to release a live system that should be able to achieve a successful level of genre recognition within 1 second of live music, if not shorter."
 
These guys at MixGenius are mix engineers by trade. With experience you take a listen to a project and pretty much automatically know, ah. One of those bands. And you pull out an old, similar project. They clearly decided to automate this process.
 
Then, detecting and fixing nasty resonances is not far from what feedback suppressors do. And making broad assumptions about scooping heavy rock whilst pulling up vocals in ballads is what I mean by "presets". Software-wise, it's possible their analysis engine generates preset files that get loaded by the mastering tools. That's how I would do it. Nowhere do they claim to have built their own mastering tools.
 
What else? Varying degrees of compression/limiting, exciters, etc. Give the user some settings to play with and the result is good enough for an "audio instagram" public.
 
I am a bit skeptical on this. I remember similar attributions of voodoo surrounding the Audio Enhancers from BBE back in the day.

Plain and simple.   Landr does a lot - way more than just compression. I've used it for multiple projects that are out on CD and clients and listeners are happy with the sound.   Again, YOUR MIX IS THE MOST IMPORTANT PART OF MASTERING.  The headroom is needed so things don't get overly crunched from the get go.   They apparently do not do any pre-master level adjustments so your mix level as they are submitted become the starting point.   Mastering is a difficult art and regardless of what many people say, I'm willing to bet LANDR does as good if not better job than most who are NOT real mastering people    One key quality I like from them is CONSISTENCY.   If I send 8 songs from a project there, they comes back sounding like they belong together.   For me, when I attempt mastering, that is a difficult result to get.   My mixes are excellent  and I am a very very good mix and recording engineer but mastering is a whole other story and it drives me crazy,   When I can't afford a real mastering person, I have no qualms using Landr instead.   Low, medium or high all work depending on the style of music.   Also, As I've said before, I pay less than half the normal rate so I get unlimited 44.1/16 wav files back as opposed to MP3s which I agree are not professional enough for many of our needs but to check you mix and get an idea how your stuff WILL sound mastered, its a great tool as well.   

cparmerlee
 
Yes, I did read that.  But I asked you a very specific question.  Would you provide a client an MP3 file crunched below a 200 kb bitrate?

 
It depends. If I was creating the MP3 myself and it was a single, I'd just go for 320 kbps out of habit, regardless of whether people can hear the difference or not. I have the Sonnox encoder and have done extensive comparisons of bit rate vs. perceived quality. Most people can't hear the difference between 160 kbps and 320 kbps, let alone 192 kbps and 320. Once the bit rate gets below 160 kbps most, but certainly not all, people can tell the difference. But people don't know that, so they assume that 320 kbps is better based on the "it goes up to 11" principle.
 
However if it was an album, that's a different issue. A typical album will take 145 MB at 320 kbps. Look at how many people complain because LANDR adds 100 MB on a desktop, probably with terabyte drives, then imagine your rock and roll buddy's iPhone with a music library, a couple movies, a bunch of apps, and because they never deleted and re-installed Tunein.com, there's 3.9 GB of playlist data files ...and another 2 GB of photos.
 
At 192 kbs, that same album would be around 73 MB. That's a big difference in size, and unless they're listening through an outboard DAC and really good headphones, they're not going to notice a difference. (If they're listening through Beats, they wouldn't tell the difference if it was 128 kbps .) And remember that we're talking about something disposable - rough mixes. I would even argue that it's desirable for someone to hear what the average consumer will hear instead of what an audiophile hears if they're doing a "reality test." If they put even a 96 kHz/24-bit song on YouTube it will end up being compressed to much worse than 192 kbps.
 
It's the same reason why people listen to mixes on Auratones. I would never demand that someone not listen on Auratones because I wanted them to hear the full quality of the mix. They can do that on the main speakers.
 
Now, also remember that some people might be wanting to test out different song orders. I do that all the time where I have several versions of the same album on my portable player. Three versions at 320 kbps is going to be be close to half GB. At 192 kbps, a quarter GB. Right now I have 1.9 GB available on my iPhone. Which version do you think I'd put on there?
 
But, this is why mastering engineers exist: they can make value judgments and tailor something to a specific situation. I think it's short-sighted to say you would "never" do something if it was against the client's best interests.
 

I just don't get the point of going to all the trouble to make the music sound its best and then crunch the life out of it.

 
Try this: get some musician friends in a room. Encode a rock song at 320 kbps and another version at 192 kbs. Do blind testing and switch back and forth. Let me know how many can identify the difference with better than random consistency.