Is Jukedeck looking into real-time composition?
A friend of mine did the music/soundscape and worked with the sound-engineers to create layer upon layers of discreet musical loops and sound effects which they then blended together in many different ways.
It's a good idea that makes sense for a lot of games. Just pointing out a counterexample.
Personally, when I see a team of people who are passionate about what they are doing create something new and interesting, I like to cheer them on.
I think the technology is interesting and cool. Even if it isn't marketable in current form (what 1.0 product is), I wish them the best in iterating on it until it is.
The investment model is set up to let founders "burn through" money while they explore new approaches to old industries. The investors don't really care that the money is lost. To be an investor, you have to assume 9 out of 10 of your investments will be write-offs.
So if you're not defending the investors' money, and if the founders are happy doing this, then why are you intentionally being harsh? Let them do their thing. Yeah they might fail, but so what? It's the only path to success.
It's not really productive to try to save people or companies via internet comments. You're more likely to demoralize them than to change their minds. Unfortunately, demoralization is often someone's hidden motive.
Moreover there is great ideas* in this thread that can actually help the company.
As a developer, occasional FPS gamer, and musician, I'd like to see them tackle adaptive generative music that is actually convincing. I want music that takes cues from the gaming environment without obvious loop splicing points and without feeling mechanical.
1. I upload a video
2. Jukedeck finds me the perfect audio for it.
Be respectful. Anyone sharing work is making a contribution, however modest.
Ask questions out of curiosity. Don't cross-examine.
Instead of "you're doing it wrong", suggest alternatives. When someone is learning, help them learn more.
When something isn't good, you needn't pretend that it is. But don't be gratuitously negative.
Calling someone's startup "headed for the deadpool" is the definition of gratuitously negative.
I wish people were more frank about what they think of start-ups, and I wish there was an acceptable way of doing that without insulting people.
(1) Convenience - There's limited market demand for royalty-free music that you have to manually edit and sort through yourself. By virtue of making it simple enough for consumers (vs. professionals) to use, the market got a lot bigger. By way of comparison, you could have said that there was limited market demand for mobile apps prior to the App Store. But the convenience the App Store offered with respect to finding and installing apps made all the difference.
(2) Licensing - This is proof of concept for the overall technology. You could view Jukedeck as a play to simply generate the sort of interest you need to get larger players interested in licensing tech, rather than an end-game in and of itself. E.g. you could imagine Apple licensing the tech for use in iMovie.
(3) Legal risk - It's one thing to say some random music track on the Internet is royalty free. It's quite another to prove it. There's no easy way to guarantee that the supposed original author of a track didn't copy it from someone else. This might seem like an edge case, but per the above points, if Jukedeck licensed their tech to companies like Apple, those legal issues become much bigger deals. Algorithmic music simplifies the legal question quite a bit.
I think the project is great, maybe they haven't found the right niche just yet. Hint: http://lmgtfy.com/?q=midi+loops
There's some basic algo-comp in Logic and Ableton already. But generally, musicians really hate the idea of having a machine writing all of their music for them.
This even applies to musicians who only work with loops, and to DJs who only work with complete tracks.
I think there are business models for algo-comp, but this isn't the best one. I'm not sure JukeDeck is necessarily heading for the deadpool, but it's maybe 50:50 at best.
Also, the production values on good stock music are very high - higher than here. Try https://www.shockwave-sound.com/ for some examples.
I think JukeDeck isn't quite up to the lowest quality tracks on there, and it's a very long way short of the better tracks.
Then why at least 3 musicians in this thread have requested MIDI export?
> Musicians really hate the idea of having a machine writing all of their music for them.
You don't get the point. I'm not saying that all your songs should be composed by a NN. This is valuable as a starting point (phrases, chords, chord progressions, melodies, etc). Then you build on top of that (or transpose accordingly). You know, there are those days as a musician when things just don't sound great and you need a spark.
I've been producing since I was 12 y.o. in MS-DOS. I know the market pretty well and I have producer friends that own recording studios and write songs for professional bands (rock, metal, etc). And I'm 100% sure they'd buy the concept because they've previously asked me for related tooling.
The technology is the key here, and could be used as another tool.
A "real" musician probably takes cues from other music, other sounds in the world. Why discriminate against algorithmically created sounds?
Secondly, there's a clear distinction between taking cues and reusing a harmony or a melody from another source. People have always done it and will do it more than ever, if they have the right tools.
Of course, it only matters if we still think about music as more of an art form than a production of non-material goods. Hope we still do.
Similar ways of diminishing other people's work are used everywhere, but are typically a result of being more or less elitist. "Not a real X if you didn't do it in way Y or use tool Z."
For some, the process is more important than the outcome (or just as), but I think it's wrong to impose such opinions on everybody else. Just my opinion, of course.
... unless they develop the algo that writes the music. E.g., think Autechre and their Max/MSP patches?
I think this is the perfect example of "failure of imagination"
GoPro is a $2.5B company based around low cost equipment for amateur film makers. They produce film editors that let you edit your films around per-licensed sound tracks.
Let me say that again - you make your film around the sound track they supply.
The opportunity for Jukedeck isn't competing with the thousands of individual composers 1-on-1. It is building a platform that generates sound tracks automatically, and selling the platform as software both to end users and sub-licensing it to companies like GoPro.
At the moment the per-unit cost of per film is constant. If software "composes" the music then the per-unit cost drops with each additional user.
That's the radical thing.
But it's not. There's considerable spread in the price of music depending on the quality, specificity, popularity, and licensability* of what you want to buy.
* up to a particular number of copies made, or with restrictions on sublicensing or other contractual factors.
You make a new film (which includes the entire range of things from Hollywood films, to pro-Youtubers to films you share with a couple of people).
For music you can
1) Get custom music written specifically for your film. This is usually expensive (unless you do it yourself, or know someone to do it for you) but has the advantage (if done well) of matching exactly what you want for your film
2) Find an existing piece of music and license it. This has variable costs (or maybe free), but won't be written specifically for your film and possibly will also be used elsewhere.
3) Use "template" music included in whatever software tool you are using.
AI-generated music has the advantage of (1) but the cost structure of (3).
It'll be a long time before it is useful for Hollywood films. But for Youtubers it is a completely different story.
Plus, the idea that I've paid for legit copyright use and I still have to deal with takedown BS is doubly frustrating.
Being able to construct a track with the right feel without worrying about copyright would be amazing for a simple video editor or on something like Instagram.
why didn't you go for one of those sources?
honest question, as gf is having the same problem sourcing music for her channel
Maybe it's not good enough for the final product but it could be good enough for testing and development.
But this falls kind of in a gray area. If the AI created the tracks, why does the company own the copyright (and thus, have the right to sell it)? In December 2014, the United States Copyright Office stated that works created by a non-human are not subject to U.S. copyright (see: https://en.wikipedia.org/wiki/Monkey_selfie#Copyright_issues ). So, in theory, AI could also own copyright.
Moreover, do they actually check every newly generated track to make sure its not too similar to previously-sold tracks?
Interestingly, I think New Zealand law does have this case covered. There's special mention of "computer-generated" works under Section 5 of the Copyright Act which states (as I read it) that the person who wrote the computer program to generate the audio is the author of the audio.
EDIT: Had another read and thought that actually it could also be the person that inputted their choices into the computer program that undertook "the arrangements necessary for the creation of the work"... Copyright is a murky section of law.
"Bits don't have Colour; computer scientists, like computers, are Colour-blind. That is not a mistake or deficiency on our part: rather, we have worked hard to become so. Colour-blindness on the part of computer scientists helps us understand the fact that computers are also Colour-blind, and we need to be intimately familiar with that fact in order to do our jobs.
The trouble is, human beings are not in general Colour-blind. The law is not Colour-blind. It makes a difference not only what bits you have, but where they came from. [...] The law sees Colour.
Suppose you publish an article that happens to contain a sentence identical to one from this article, like "The law sees Colour." That's just four words, all of them common, and it might well occur by random chance. Maybe you were thinking about similar ideas to mine and happened to put the words together in a similar way. If so, fine. But maybe you wrote "your" article by cutting and pasting from "mine" - in that case, the words have the Colour that obligates you to follow quotation procedures and worry about "derivative work" status under copyright law and so on. Exactly the same words - represented on a computer by the same bits - can vary in Colour and have differing consequences. When you use those words without quotation marks, either you're an author or a plagiarist depending on where you got them, even though they are the same words. It matters where the bits came from." - from http://ansuz.sooke.bc.ca/entry/23
Basically, non-recorded metadata about how a sequence of bits was created matters too. Not just the bits themselves.
cat /dev/random > /dev/eth0
Isn't the world a wonderfully messy place?
Sure, but (and I'm not a lawyer), intent of law is just as important as the literal meaning of law. Cases play out this intent and add to the corpus of knowledge as case law. So if you did write a program to generate all the data in the world, I'd imagine people would look at your intent, rather than just what you literally did.
Within the realm of audio production, there's an even closer equivalence: soundfonts, which are also little nearly-Turing-complete programs. Playing a MIDI file "through" a soundfont? No problem. Putting that track on a CD and selling it? Nope, need to license the font.
Really, you could think of procedural music generation as a really complex and "stateful" soundfont program that has a nonlinear relationship with its input.
I don't think there's anything legally stopping either behavior; it's just that neither GNU nor Adobe are in a market position where anyone would put up with that sort of behavior.
This was followed by the famous Bridgeman vs. Corel, which established that taking a picture of a work doesn't create a new copyright. Thus, pictures of public domain works are public domain. Despite much huffing and puffing and disinformation by the museum community, that's now settled law. Nobody has gone to court to try to overturn it. (The National Portrait Gallery (UK) threatened to sue Wikipedia, then backed down once they realized they would lose.) There's also Meshwerks vs Toyota; a 3D scan of a physical object doesn't create a new copyright in the 3D scan. That's an appellate decision and reaffirmed Bridgeman.
So, for copyright in the US, there must be an Author. (This can be a corporation, but that comes under the law on work-for-hire; the individuals involved are the initial authors but the rights accrue to the employer.) You can make a strong argument that under US law, works created by computers are not copyrightable.
> You can make a strong argument that under US law, works
> created by computers are not copyrightable.
I don't know exactly how Jukedeck works, but it seems self evident that it is the product of a huge number of creative decisions - which sound samples to include or generate, which sounds can be combined in a pleasing way, which melodic patterns are appropriate for various genres, how melodic patterns are modulated over time, etc. etc.
It's not as if the Jukedeck team created a general purpose AI and said "go make some music". They designed a system that can generate a limited (though large) range of music based on their own sense of creativity, taste and style. I argue that the music generated by such a system is clearly creative expression and thus subject to copyright to the same degree as traditional creative expression.
I do not think "sweat of the brow" decisions are relevant here. The lack of "sweat of the brow" copyright means that simply performing labor without a creative element does not qualify the product of that labor for copyright protection. There must be a creative spark present in the generation or transformation of the work.
The fact that, with Jukedeck, the creative spark happens at the time of producing and editing the code rather than at the time of the code generating the music is not relevant in my opinion.
Kind of like a studio letting you use their equipment under a prior agreement that anything you generate they own until you purchase it.
The same reason artists like Autechre deserve to own the copyright to music they created using algorithms. They wrote the software that made the song, it's just somewhat more abstracted version of composing isn't it?
Uh, do you think there's a sentient being here? Its just a relatively simple algorithm here (at least compared to sci-fi style AI), the same kind of thing that decides how to autofix colors in photos or do a transform in photoshop.
Dynamic music generation has been a thing since at least the 80s. I think I had an Apple// program that did this.
>works created by a non-human are not subject to U.S. copyright
I'd like to see you prove some silly music generator is on the equal footing of a living and intelligent animal to any court. I can't imagine you not being laughed out of courtroom.
E.g. the recently infamous selfie by a monkey was ruled public domain, having been produced by a non-human.
Every picture you've ever taken, the Universal Slideshow already contains that picture. The picture of your birth, every memorable moment of your life, and every possible variation of your death.
It's already there it just needs to be found. That's the entire point!
>(There is an extended Library that contains entire novels to be searched - the web version is a bit smaller/limited to pages.)
The creator has variations of the code that aren't online. For example, lower-case alphabetical letters was a choice to "remain true to the original concept" but a base64 variation that allows for capital letters is possible. Furthermore - searching for entire books rather than being limited to 3,200 characters is also possible. Also other languages would also be possible but require further variations of the code.
It's just having this work on the web to a website that sees 30-40k daily visitors because the algorithm isn't fast enough to meet those demands.
Rate limited in posts. My reply to the below is as follows:
Your claims only hold for the web version. I was not speaking of the web version. So you are explicitly wrong. There is a version of the library that is not limited to those characters, that is not limited to lowercase Latin alphabet, and is not limited to being searched 3,200 characters at a time. So what are you contending by making such statements?
Yes. The web version is held to those limits. Which is why I explicitly mentioned I was not referring to the web version. Your claims only hold against the web version. So I'm not seeing the point you are trying to make here.
I hate how HN rate limits like this. :)
I'll concede that. Does ASCII art of a Chinese character represent the same information as the Chinese character? That's stretching things so I'm not making that as a counterpoint - but more of a thought experiment.
For example: https://libraryofbabel.info/bookmark.cgi?ygofnzdyjijlsc299
Edit: Reply to your reply:
One inherent limitation with the current code is that it has to assume some encoding. Currently, there is no encoding that contains all known glyphs (e.g the Prince symbol, uncommon kanji, etc) so there will be texts that can not currently be generated, regardless of how much you increase the character set.
Edit2: Even if one were to allow ASCII art to represent a character, you then have the problem of how to distinguish between ASCII art substitution and actual ASCII art. Consider "Densha Otoko," which basically consists of message board posts that often contain ASCII art.
Why doesn't then Adobe own result of whatever "autofix colors" produces?
The goal for this technology is to beat the two guys who write most popular music. Another few years.
I would very much like to try it with MIDI output, as I personally enjoy tweaking synths and effects more than the compositional tasks of harmonizing melodies etc.
It's crazy how good the tracks sound.
One thing that would be nice is the ability (like some others have stated) to tweak the generated sound. For example, there might be a lull at 0:24 but given my video it would be best for the lull to be at 0:52 instead. Would be cool if the layers/pieces were moveable a bit to make it perfectly fit the content it's paired with.
Also I noticed in long tracks, like 5min+, that it gets a little repetitive, like the algorithm is trying to pad it out. At least in the few I created, could be a fluke.
Totally agreed - it would be great to be able to tweak the generated sound. This is something we want to open up as a possibility asap.
Awesome idea and execution, quite jealous I didn't think of this!
I'm specifically curious how much variation there is between tracks -- perhaps we can get more HNers to post?
Meditative percussion'd ambient yoga music with a fast tempo.
I'm an ambient electronic music maker (and software maker too). I saw a lot of similar projects since Brian Eno's generative music project. I have been also interested in making algorithmic music (using some AI/artificial neural network schemas), using amazing Supercollider, for a while.
I never achieved interesting results, in terms of "deep energy" instead "embedded" when music is made by "special" humans (so called "artists").
Experimenting soundscapes CREATION, in recent years I was back to what I call (along with Steve Roach): "analogic approach": sonic seeds as analogic waves (electro-acoustic) -> (digital) elaborations made by human(s) artist. No MIDI. No "samples" usage as-is. No presets.
I could call the musical secret as a case of human intelligence: "search and discovery" of unknown.
Because, this is the point, music is discoveryng of mistery.
BTW, I do not want to enter in the loyality-free / real-time composition topic. So long discussed for so many years among electronic music communities!
Good music, is like science inventions: come from "singularities".
That said, an interesting point, for me, is the fact Artificial Intelligence could help musicians to make music. Ok, but this is another story, another vision of what music is,
Instead I got a 5 minute long song. For my needs, I'll stick with soundcloud "Dj sets" which seem to go on for a few hours each.
I would pay in the order of thousands of $ (pounds in your case...) for this system running locally in my network, and in what I would call 'continuous mode'. By which I mean, I give it initial conditions and the track just evolves continuously forever.
Ideally, I could then send additional configuration messages and have it evolve further. e.g. Start with ambient/sparse. After a few minutes, I send a command to transition to ambient/sci-fi. Then later, to electronic/aggressive, with a seamless transition between. Even better if I can have a command that is essentially "react to this message", and it does something like a cymbal clash or whatever is appropriate for the settings.
I've done improvised theatre shows before with a jazz band that did all this for us, and it was AMAZING! Being able to do it with an automated system would be a game-changer for my business. I've been considering writing it myself, but it would cost at least tens of thousands of dollars worth of my time, and our current solution is sufficiently adequate that I'm not ready to do so yet.
 Ironically, given your current business model, I'd actually have almost zero interest in retaining the music after it's been played.)
On a positive note, I really enjoyed all the tracks I did make!
The start time for a clip is much more important than the end time, so you should adjust the cutoff time for the clip by shrinking it from the end (never expanding it, as clips are often truncated when the footage becomes unusable, i.e. when someone walked in front of the camera, etc.)
I can imagine a Premiere plugin with a stored credit card, and you guys doing a brisk business that way. You could even then become a marketplace for royalty music, have a store where people can upload their music tagged with moods, and video creators could see royalty-free options from Jukedeck or royalty options from third parties.
Places you can add unique IP are: the ability to extract a mood from the video clip itself (most likely through color grading, the speed of pans, the frequency of cuts, and whether it's tripod or handheld), and the ability to automatically extract a mood from 3rd-party music, so I can automatically see the pop songs that are best fits for my video.
Find a way to do this without having to upload the actual video file, because that will kill the experience. Work-in-progress video files are often multi-gigabyte.
On a side note, please minify your app.js file. It's 4mb and took forever to load. I almost left the site due to that.
A lot of people are mentioning video games, but another interesting use case for real-time streaming would be for running. There's tons of data on my smartphone that can help inform where the music might want to go next. Would definitely be cool to have my own personal band synced with my body's running rhythm.
Here are some:
The market for a tool like this is decent sized, specially for Electronic Music. You could also bake the technology in a VSTi or AU (although the webapp a la beatport might be much better), or even license the technology to big players in the music production space, like Apple, Yamaha, Sony, Steinberg, etc
Tip: ask here for opinions https://www.gearslutz.com/board/
I find an interesting parallel between 1/2/4/8 repeating beats/parts in songs, and binary...
tl;dr: Frequencies that are related by some simple ratios are defined as "intervals" which are organized into scales. A handful of those make up the vast majority of popular western music and "sound good together".
I'm finding the tunes the least successful element here. They sound like a fairly naive Markov sequence/chord spine mashup, with none of the character of really good tunes.
The sounds/production are much better. The drum parts are midway between the two.
>I find an interesting parallel between 1/2/4/8 repeating beats/parts in songs, and binary...
There's more to it than that.
Good algo-comp is PhD-level work. Kemal Ebcioglu, who made one of the best efforts to date, had to add a lot of hand-rolled heuristics to an expert system that was trying to produce classical counterpoint. He was last seen working for IBM research on compilers and architectures for parallelism - a very smart guy.
It's easy to underestimate how hard the problem is.
In building Jukedeck, we're applying our own theories about what makes good music. And, like any musical composition, it's an experiment!
Have you thought about applying also common patterns found in the best selling hits? Hooktheory has a nice DB for this:
re: what determines what sounds good, in some sense you can read up on music theory for this (scales / keys). in a more basic sense ("why is this sad and this happy?") it's actually an open and actively researched question in the physics of music. it's anyone's guess, really, why certain melodies are pleasant. is it a physical/structural/formal feature of the sound itself? is it a feature of our psychology or neurological makeup? who knows.
Being as every second is precious, to a first-time user of your site. And as we know, "Any sufficiently advanced technology is indistinguishable from a rigged demo."
My Folk song sounded pretty cool, and couldn't help but imagine some video game action that would fit the music.