_ (_) _______ _ _ _ | ____ | | | | (_) | |___| | | _____ ____ __| | ____ _ _ ___ _ ____ | ___ | | (____ | _ \ / _ | | \| | | |/___) |/ ___) | | | | | / ___ | | | ( (_| | | | | | |_| |___ | ( (___ |_| |_|_| \_____|_| |_|\____| |_|_|_|____/(___/|_|\____) _ _ ________ _______ __ __ | | (_) | | | | | ____ _____| | _ _ ____ ____ _ __| | | |__| | \(____ | |_/ ) | _ \ / _ | (_) || || | | | | / ___ | _ (| | | | ( (_| | _ ||___________||__| |_|_|_\_____|_| \_)_|_| |_|\___ | (_) | ___________ | | (_____| ____ || | ||__| _ || || | | | ___| | ___ | |__| _____ ____| |__ _____ _____ ____ | | | | | (____ | / ___) _ \| ___ (____ | _ \ __|__ |_____| __|__| / ___ | ( (___| | | | ____/ ___ | |_| | | | | | | | \_____| \____)_| |_|_____)_____| __/ __|__|___|___|__|__| |_| | | | | | | ___ _ _ |___|__|__| / __) | _ (_) ____ _____ _| |__| | _____ ____ _| |_ _ ___ ____ / ___) ___ (_ __) || ___ |/ ___|_ _) |/ _ \| _ \ | | | ____| | | | || ____( (___ | |_| | |_| | | | | |_| |_____) |_| \_)_____)\____) \__)_|\___/|_| |_| AI and music making : A cheap reflection, borderline ranting, on what makes a track sound good to us puny humans, by our special reporter BoeufStroganoff AI... AI... ! AI is the new information superhighway, the new "in the year 2000". You hear about it in every start-up company pitch and hi-tech product description. But now even your coffee machine is equipped with AI. Of course, it is not used for making better coffee, but for snitching on you for the benefit of the manufacturer. He probably makes more money selling your day- to-day routine data than he does selling the coffee machine in the first place. We music aficionados are not spared by this AI craze. As would flashy click- baits title have you know, it is the future of the music industry, the eardrum revolution, a new era of computer generated music content. Wait a minute, hold on... that sounds more like a big pile of bovine discharge to me. Or is it ? It is important to make a distinction between two concepts involving AI in music production. There are studio tools, a natural evolution of previous VST's and plugins, allowing you to do impressive stuff like intelligent noise cancelling, instrument isolation, auto-autotune, auto-mixing tools and so on. No problem there, they are designed to make your life easier. But then you have AI as the composer of the music itself. The latter, in combination with the former can produce an entirely new track out of a dataset and some initial parameters like genre, length, tempo, you name it. OK cool yeah whatever... But how does it sounds and is it really something that we want ? Of course not all music in the world has to be an amazing transcendental experience. It's quite the opposite in fact. Most of the music produced today is more of the functional type, just fulfilling its task of filling a silent void that would otherwise be perceived as uncomfortable. You can't really play intricate clever banging junglecore in an elderly home elevator (or can you ?) We need proper elevator music. Similarily, that background deep house music they play in hipster hotel lobbies nowadays isn't that complicated and it might very well be cheaper to generate it on the fly with AI. And that sort of music should be appreciated just for that, regardless of how it was produced. It is a segment of the music industry that can totally be taken over by AI because this type of music is only used as a reassuring background soundscape. And let's not talk about the millions of scripted Youtube videos in need of canned music that were uploaded as you read these lines. This plethora of cheapo music will have an impact on the "price" of music : in the same way that the value of quality pictures decreased constantly since humans are producing images, the once-reserved-for-kings privilege of ordering music to an artist will continue to go down until it will reach zero, Enron style. So far, all the examples I've seen of online services that provide music generated by AI fall in that category : something you listen to without paying much attention. Decide some parameters, click generate and PRESTO, you have a piece of music. But is it any good ? Listening to it seems ok for a few seconds and then re- PRESTO : it starts to make you feel a bit weird, something is off, you feel the industrial-processed taste of it. Fake butter doesn't taste like butter, and fake music doesn't feel like music. But can you use it or is it all bad ? The analogy to photography imposes itself here : in this pro- Instagrammability era of cheap imagery, anybody can be a good photographer. The recipe is simple : take hundreds of pictures and only select the top 3 nicest ones. Then, here we go, everybody thinks you are an amazing legit photographer. The experience with AI music can be quite similar. Just generate tons of tracks, take the best one and delete the others. Then, suddenly, AI is OK at making music. Of course, we could train another AI to select that 1 good track out of 10,000... seems like we can russian-doll the process here, but how would you train that AI to process these tracks ? We hear very often that the creative act is based on accidents. Sure, it is true that you can come up with amazing melodies that you didn't originally envisioned. Has anybody here ever programmed a MC-202 ? Make a mistake while recording the midi track of your next supersaw lead or something and BAM, illusion of talent ! We could also "mutate" whatever an AI is giving us, but this is where it becomes interesting : accidents are indeed important, but what is more important is the ability to judge those accidents and decide if they should be rejected, assimilated or modified further for even better results. And THAT is the thing that is a bit more tricky for a computer : having taste. The musician's judgement about composition and sound design is a very complex process, relying on the entire life, taste and background of the artist. Maybe the color of the wallpaper in the toilet of your grandmother's bathroom might have a small impact on all aspects of your choices as a musician, taking your track on an alternative path. Taste is a constant and deliberate state of mind with which the artist applies a set of rules defined by his past life experiences at every level of the composition process. It's a bit like taking a hike on a mapped trail, slipping, and instead of returning onto the trail, you decide that it looks more beautiful down here. Chances are you will discover a new valley filled with cool rock formations that you've never seen. You might as well camp there for the night and explore. AI of today do not have a life, they do not have experience, they do not have... consciousness. And this is why they will not replace musicians anytime soon. Without trying to channel good old Nostradamus here, I think this whole AI business will become extremely interesting when it will be complex enough to develop some kind of weird cyber-brains, capable of cognitive thoughts, thus opening the Pandora box. But our definition of what sounds good might be different. How will those silicon cyber-beings feel the passage of time for example ? Time, or rythme, is the most defining concept of any kind of music, and I'm sure the vibe must be different when you can perform 57248 MIPS. Maybe they will produce legendary tracks that only last 5 microseconds. A great piece of art for them, but just a faint "click" for us. I'm quite curious to hear music motivated by a computer "Vague ‡ l'ame". A good AI would also integrate the fact that sometimes you need to listen to a song several times, in different contexts, in order to suddenly unlock it's magical appeal. And what about that track that sucks for the first 5 min and then suddenly becomes amazing by contrast ? Some tracks are impossible to listen to during a rainy commute, but play it loudly with the right amount of ethanol in your blood and you will take off to space. It would take a concious AI to understand those concepts and AI stuff generated today only tend to provide music that are designed to be liked right away. Boring ! Today's AI can be very good at taking care of some parameters of your tracks though. Making sure the technical aspect of your piece of music follow some pre-established rules, like not clipping for example. But I'm afraid of this kind of conformity, and if your AI avoids, on a technical standpoint, to make any mistakes, then you're missing out on quite a large landscape of these happy accidents. As a way of analogy, using Lens flares in cinema was considered a technical no-no, a proof of your amateurism as a director of photography, and any lab technician would report it if he sees one. Well... JJ Abrams anyone ? It seems like the use of AI in human business is quite successful when applied to specific repetitive tasks, like image recognition and manipulation, pattern recognition within datasets and so on. But for the creation of an genuine original artistic piece, AI seems to lack the framework possesed by humans to be able to create something ex-nihilo. You can ask an AI to recreate an old Flemish painting if it has all the other existing Flemish paintings available to it, but can you give an AI a blank canvas and ask it to create a new style of painting without copying whatever was done before ? So what makes a good "human" track ? well I can't help but somehow comparing it to sex. Everybody has their special taste, a little kink that suddenly talks to you. It can be the position of your arm, the angle of the neck of your partner, a sound... Something would just suddenly click in your head because of what you see or feel, drives you, and makes you come. To some weird extend, I think the fact that our brain suddenly decides to trip on a track is quite similar. But as for sex, it is a very personal experience, a sort of individual signature, a set of rules that only applies for you. You can easily feel that tipping-point moment when good musicians jam together. It's a bit sketchy at first, then their brains synchronize and the energies add up. Suddenly something happens, a pleasure nerve is activated in each of their head and the public feels it in a resonant fashion. Now the pleasure of the audience is proportional to the one of the musician, because a bridge is being build between the two minds. This is the moment where my shy-self would suddenly allow my body to let go of its inhibitions and dance like nobody is watching. Music and the pleasure of listening to it is about setting a standard for yourself. People might call you an elitist for not liking pop stuff, for listening to "obscure" electronic stuff. But what you listen to is no one else's business. You like what you like because of the pleasure it generates in you, not because you want to impress the others. You like what you like because you see a different twist, an originality, a colourful soul in it. You want to explore your limits and push them further into new genres and artists. The love of Music is mainly a matter of pleasure but also the love of being challenged intellectually by it a bit. If you think AI is here to supplant humans, just give a listen to a Prokofiev concerto to convince yourself otherwise. AI will not replace a lifetime of emotions, talent and hard work. It just provides a new musical product using a checklist of gimmicks, rules and stereotypes that make a track sound like the real deal but it is not the real deal. AI today only adds up to the digital background noise that we are generating more and more in our society. It's going to pollute our cyberways by drowning authentic human-generated content in a gigantic sea of automatically generated fake stuff. Do you remember how charming it was when all the webpages of the internet where designed and written by humans ? They all had a special charm and differed from one another. Now human generated content has become a rarity. Small, niche groups, like our synth community, are still maintaining some of them, dedicated to a brand of samplers or compiling anything there is to know about a certain synthesizer for example. But as their owners die or forget to pay the bills, they are slowly being replaced by the dreadful 404 page. Since the very essence of non-functional music is to communicate emotions, and since we know that an AI does not feel any at its current stage of evolution, we therefore can't fall into the trap. AI for any artistically relevant music is a scam, a well adjusted algorithm designed to trick us... an Ode to Fakery. The late James Stinson mentionned that he isolated himself from every influence to create his music, that it came from within, untouched, clean from the corruption of external magnetism. AI's very nature can't do that. But what if we start cross-feeding it with other datasets, in fact with all the other datasets available in the world. Suddenly it can learn about ancient Japanese clay pot design, fashion from the 20's, shapes of tree leaves around the world, and literature from Ouzbekistan. Perhaps, with access to all this complexity, AI will slowly be able to touch us, to pass the eardrum Turing test, and produce some good shit. But one important question remains : can an A-grade fake be considered to be as good as the original ?