$4M would pay for a lot of mediocre human composers to produce a lot of mediocre music.
I understand the value proposition here and it's neat technology, I just think it might still be a little bit early for millions to be pouring into the field. Every example I've heard of "AI music" has been some level of failure. But, maybe they just need a few million dollars to throw at the problem, and it'll go away.
The real niche IMO is in game music. If you can get the music to respond in real time to players actions without becoming repetitive that's probably worth 100+M across the industry even if it's never 'great'.
I dunno, I've heard Al Biles and his GenJam[1] autonomous accompaniment play, and it's actually pretty good. On par with most small time performers, at least. (Though maybe the synth could use some higher quality samples.) Also probably not technically AI, I imagine, but since a NN is just a generic function approximator, its conceivable a musician AI could imitate this improvisation.
I think this company faces some serious challenges, but to me the main challenge will be to create such incredible value from the interface between customers and the intended product that it is wholly better to use their software than just search "piano stock music" somewhere and pay a couple bucks (or pay significantly more bucks for a producer's labor).
From what the article says, it looks like this company is looking to provide their software as a tool for companies to generate stock music with minimal effort. They claim that stock music currently available for purchase is both overpriced and generally insufficient for commercial projects with specific demands. Perhaps it's also implied that paying a producer would similarly be too steep of a cost. My main issue with Amper's concept is not that the creation of music with AI is too farfetched- that is not a new concept. My issue is that the value of Amper's software depends on their ability to optimize the process of creating music according to strict specifications- something that I think is bottlenecked by the speed of human conversation. Like any producer, AI software can get the ball rolling, but ultimately creative decisions have to be made intelligently somewhere, and there must be an exchange of feedback which is more efficient and stronger than regular human conversation for Amper to succeed. Not only must the AI be very good, but the user interface itself must be so good that someone without a music production skillset and only a vague understanding of what's needed can do the work. I think creating this kind of interface for a layperson is the bigger challenge here.
I was also extremely unimpressed with the demo. The sound quality was very poor for commercial music and there really isn't a lot of range shown in the tracks. And it isn't because of soundcloud, there are plenty of even amateur composers with more professionally produced music. I know it's supposed to be stock music, but isn't this software supposed to be capable of generating excellent, detailed stock music?
This looks promising. Having spent many years editing stock music for video we'd spend considerable time not only finding the right mood/genre but also editing the track to make it seem like it was "composed" specifically for the video. So, I think this solution only addresses half the problem.
What would be better is if you could download the individual tracks to import into a DAW where you could further edit/manipulate them to get that more custom sound. I went through the "Pro" tutorial and this wasn't an option.
so would you like some kind of tool within the timeline of the video in order to express things? Like some way to specify the a value corresponding to a dimension of the mood? What else would be useful? Perhaps a way to place a crescendo or a bridge? Would an option to do a MIDI export be useful?
So you'd want them in STEMs essentially right? I'm not a huge fan of the format package that DJs have sometimes used, but do understand what the legacy term means. Having one track for drums, keys, ambience, etc. That could be helpful from a tweaking / editing / fitting angle maybe?
We can't automate filling out Excel spreadsheets, but investing 4 million dollars in automating music creation. Meanwhile there are plenty of CC-licensed tracks and plenty of aspiring composers/producers who would gladly work on creating tracks for "commercials or short online videos" for very reasonable price.
I think the question is of quality. Many people can make music, and anyone can automate making noise, but automating making music, good music, is something that's hard to do, and has a huge market.
One can hope that one day we'll be autogenerating music based on individual preferences in a way that every piece you hear is a unique arrangement.
I wouldn't want that any more than I'd never want to eat a favorite dish again. Novelty can be nice but it's not the only thing that makes an experience good.
Those are pretty handy features, but they are not nearly enough. I know plenty of people who spend several hours a day entering, validating and fixing data in giant Excel spreadsheets because they need to work with multiple systems that don't have real integration points. That work is mind-numbingly boring and completely avoidable. (Besides, system integration is also rote work, only one level up. I know plenty of engineers whose only job is to constantly write adapters between various web services.)
need for filling out excel sheets automatically is not greater or lesser than music automation. One field doesn't have to wait for the other to be solved imo.
Also, there are plenty of able workers ready to fill Excel sheet for a very reasonable price but that is not the point here either.
Music, entertainment, and branding cannot be separated in the mainstream. Let me put it this way - a large reason I listen to The Weeknd is because of the interesting lifestyle and excess that he portrays through his branding, his music videos, and his whole persona/vibe.
It's never just about the music, which is why you see a ton of amazingly talented SoundCloud producers that no one has ever heard of. They understand the music but lack the branding and persona development that gets people to actually stick to their music.
AI produced music will be like computer generated art - too distant for most people unless there is another component on top of the music.
That's my initial take. Curious to see what people think of my opinion.
> Let me put it this way - a large reason I listen to The Weeknd is because of the interesting lifestyle and excess that he portrays through his branding, his music videos, and his whole persona/vibe.
A large part is the evolution of the persona too. Following his growth as an individual has been amazing (although House of Balloons is still my favorite project).
While you're correct, AI musicians can easily have all the features of a real musician. Check out Hatsune Miku. Although the music and lyrics are written by humans, the singer is a completely fictional character who sings (generates) the music. Fans love her (it?). So, we've already solved the persona problem. Just need to generate good lyrics and melody as well now.
Not really. Hatsune Miku is more like an instrument. Like piano or violin, but synthesized human voice (not completely synthesized even, they used voice samples from the real human). Miku is one of many vocaloids. You can think of vocaloid as family of instruments, like violin family in traditional music with violin, viola, cello etc. In my opinion, she is not even the best vocaloid, but most widely known for sure. And, as with all musical instruments, it all comes to how composer uses her. Most songs you hear only because composer is already fairly popular in doujin circles.
AI generated music is completely different field, because AI is the composer there. This is why in the original article they talk about how they want to create symbiosis of AI and human composer, essentially creating another tool aka instrument, but this time human needs to do even less in terms of actual music creation.
> too distant for most people unless there is another component on top of the music
That's exactly the point of this tool. This tool is meant to produce background music for commercials that's cheaper than what's currently available. The "another component" is the crap that music is in the background of.
I totally agree, search for Vocaloid or Hatsune Miku for a similar thing. I even went to a live performance in Japan (probably as the only foreigner from what I saw) and it was amazing.
What it really needs is the ability to add cues, similar to how ML painting programs are now giving you the ability to define areas of a photo/painting (this is sky, this is water, etc).
If you could say I need 30 seconds of uplifting music for my trailer but here are the 3 places where things are fast paced and here is the point that is shocking, the AI could in theory accommodate that into one piece that aligns with your already assembled visuals.
I can save people a lot of time: It's not about the information on the medium, the .wav file the algorithm produces. It's about the context in which the information is presented.
AI-generated music will go along with generic food and other products. Listen already to the generic soundtracks on most Kickstarters. It works because they're paired conspicuous consumption. Just use the phone app "SoundBot" to make a generic chord progression with 1/8th note rhythms.
"Compelling" (a la Magenta) art and music will have to understand the context, have a perspective and understand what is ineffable and attempt to convey it. That's an unbelievably large machine learning task, having essentially no boundaries and with very little data to source from.
Nevertheless, as a company, Amper may create some useful musical man/machine collaboration tools.
Interesting coincidence. Just today I was thinking about when AI would be able to generate music for in Clubs (which are mixes of other songs / sound-effects).
I wonder how this will turn out. I did like the sound sample at the bottom.
As a 'traditional musician' (i.e. I play a couple of instruments, read music, jam and write songs with real humans etc), I used to think this way, however actual 'good' club music takes a lot of time to produce in terms of selection and mixing and mastering of sounds. A friend of mine, who is a talented performer and producer, deconstructed some of these tracks for me, and pointed out just how much effort and talent it takes to produce good pop / dance / club sounds. So while on the surface it would seem possible to do something passable 'automatically', I am actually not too sure how good it would sounds, especially to audiences used to hearing the best that most talented humans can come up with. Allow me to coin the phrase 'club muzak' for what early AI 'composers' will produce (although I bet then that these 2017 'club muzak' will become all the rage again across the solar system in the late 2050s')!
I used to be a rave DJ and am friends with a couple of producers. The core of a dance music track is pretty well dictated by the sub-genre.
You can do a lot with a few wel chosen loops and generative music algorithms. What really sets off the best techno tracks is engineering and mixing, and I actually think that might be harder to get right with AI.
Kidding aside, this looks useful for a specific audience and in a price point consideration that hopefully gets traction. I've been fond of royalty-free stuff via purchase / service for a long time, it's a great platform. Will be a good one to keep an eye on!
Thanks for asking. The entire composition, performance, and production is done by Amper. The music on Soundcloud is Amper's direct output without any human intervention.
Where's the AI? At best, it looks like they use a pre-trained algorithm to generate output based on a few inputs. Similar to the deep dream visual effects Google was generating a while back.
Amazon's recommender systems adjust dynamically, based on what users buy. Google's handwriting recognition is learning every time you select the correct choice and continue with your next scribble. How does this learn anything?
We were discussing dark vs light themes at the office and the general conclusion is that dark is in fashion, no logical reason.
I use a light theme because I get a lot of light in the afternoon (even with curtains closed) and the glossy mac monitor turns into a mirror with a dark theme.
Could you share your snippet? I use stylish just for turning gray/dark theme websites(domestic ones) back to colorful again since the King's passed away :/
I agree with non-distracting in in-the-other-hand perspective. Dark theme makes me too focus on the work. In long period, it behaves like a magnifier in sunny day, the focal point burns thing so quickly.
Skeuomorphism is also completely out of control in the music tools business (not in this particular example, but, it's super common in VSTs and the like). Which has really dumb effects when used on a 4k laptop display, since it's rarely subject to the OS scaling, because it's got its own UI lib.
I don't know why dark and ugly is the norm in media manipulation tools. But, I find it annoying (mostly just hard to read, due to too-low contrast), and often switch to a lighter theme if it's available in the tools I use.
I think there's a subtle evocation of a "studio" context in dark UIs: Professionals tend to associate dim lighting with a control room, an editing suite, or something along those lines. I don't know why these spaces are traditionally dim, though I'd assume it's something to do with reducing glare on equipment displays.
The technology is definitely there (and it can be done in many ways); I recently came across researchers/musicians using the self-organizing properties of slime molds to create music: https://vimeo.com/163427284
I understand the value proposition here and it's neat technology, I just think it might still be a little bit early for millions to be pouring into the field. Every example I've heard of "AI music" has been some level of failure. But, maybe they just need a few million dollars to throw at the problem, and it'll go away.