Where are all the (good) AI music generators? Is the music industry bribing AI researches to not release such technology?

Where are all the (good) AI music generators? Is the music industry bribing AI researches to not release such technology?

ChatGPT Wizard Shirt $21.68

Beware Cat Shirt $21.68

ChatGPT Wizard Shirt $21.68

  1. 12 months ago
    Anonymous

    >bribing
    No it's the stick they are worried about.

    • 12 months ago
      Anonymous

      The 'Industry' doesn't care, only the pawns laboring care.

  2. 12 months ago
    Ora Magana

    AI can't into logic and math.

    • 12 months ago
      Anonymous

      I love you GPT-chan

  3. 12 months ago
    Anonymous

    its probably those rich rock bands with many lawyers that worry about ai development. if we got a decent ai music generator then those gays would be instantly mogged because all it takes to write a rock song is a few guitar samples and some vague lyrics that normies consider omg so deep. whenever i listen to rock it sounds so samey that i could very well envision it successfully generated by ai.

    • 12 months ago
      Anonymous

      As opposed to what? Hiphop or opera?

      • 12 months ago
        Anonymous

        as opposed to actually good music like aphex twin, boards of canada, 100 gecs, crystal castles and other electronic/idm contemporary artists who can afford to have different instruments and textures in each song and who are not criticized based on lyrics

        • 12 months ago
          Anonymous

          >aphex twin, boards of canada
          These are good but rock bands are still better

      • 12 months ago
        Anonymous

        If AI could generate a good opera it could do entire movies no problem.

    • 12 months ago
      Anonymous

      god I fricking hate those kinds of "popular" rock.
      it takes skill to write something so emotional and skillful drum/guitar plays this is why I miss the emo culture back when skill really peaked more than these onetrick rocktrash who's only fueled by their marketing team and label engagement shills

  4. 12 months ago
    Anonymous

    Music is a very complex artform because it combines lyricism, rhythm and melody. Lyrics, sure that's easy. The latter too insanely complex especially if you introduce layering, variations, etc. Etc.
    Until AI gets hearing and emotions, it'll be impossible for it to generate or even imitate good music.

    • 12 months ago
      Anonymous

      >hearing and emotions
      Music (harmony) is mainly math.

      • 12 months ago
        Anonymous

        >Music (harmony) is mainly math
        The foundations of it are, yes. But not how to interpolate and intertwine them, that's purely based on emotion. We don't really know WHY a hihat on beats 2 and 4 sounds pleasing to the ear, or why 3/4 has more slow dancy feeling. All of this ai can and will never know, thus being unable to implement it in any composition. And that's not even diving into tonality, instrumentation and articulation of instruments, this requires such an insane complexity and years of actual hands on experience with the instrument any ML system would pathetically fail. Hell, we don't even have a good synthesizer for strings nowadays, and that's just synthesizing i.e. replicating the sound. That alone should tell you something.

        • 12 months ago
          Anonymous

          >We don't really know WHY a hihat on beats 2 and 4 sounds pleasing to the ear, or [...]
          read a book about memetics, anon

          • 12 months ago
            Anonymous

            >Yeah bro the language model will just understand the cultural context
            Lmfao

            https://github.com/search?q=music+generation

            Yeah and every single one of them is shit

            • 12 months ago
              Anonymous

              >Yeah bro the language model will just understand the cultural context
              are you able to use google search?

        • 12 months ago
          Anonymous

          Then you just prompt for what you want.
          >80's love song, 3/4, emotional, raspy voice, female back vocals, (masterpiece), (trending on spotify)

          • 12 months ago
            Anonymous

            Natural language prompts are the dumbest shit I've ever seen from the field. It's purely for PR reasons that they do that, otherwise you'd have a set of fields that you can keep unset, or set as you like. This would completely bypass any chance at misinterpretation from the language model and let all remaining mistakes be purely from generation, i.e. no compounding error problem. The interface wouldn't be any harder, either. Think about a tag system with a search feature like in a booru.

            • 12 months ago
              Anonymous

              In a perfect world that's how prompting works, enabling tags and negating other tags and generate until you get something you can tinker with till you get what you want

              • 12 months ago
                Anonymous

                But it's not how it's implemented because "le ai so le sentient look you can le type at it with le natural language".
                That's the problem I'm pointing to here, it's pure appeal-to-NPC bullshit. As a result you get problems with compounding errors (misinterpretation * misgeneration). It is likely this choice was partially motivated to hide issues with the generation as issues in the interpretation, too.

          • 12 months ago
            Anonymous

            >Bro just prompt and it'll magically create a good song
            Kek

            • 12 months ago
              Anonymous

              Nah but then you can fix it up after, like you do with stable diffusion. Why wouldn't it work with music when it works with images? If you think music is somehow too complex you're in denial. And I'm not talking recreating Mozart here, but a lot of pop producers would go crying on twitter overnight. And that's what music is now, that's the industry.

              • 12 months ago
                Anonymous

                >Why wouldn't it work with music when it works with images?
                Because music is much more complex, that's what we're trying to explain to you brainlet. Even shitty pop songs have several dozen layers of complexity that intertwine and depend on each other for the song to work. In paintings you basically don't even have that opportunity because it's just a collage of smaller elements.

              • 12 months ago
                Anonymous

                Ok, we break it down into smaller pieces then. If that's what you mean by complexity it's theoretically very easy to circumpass. First a melody, then a bass line, then this then that. I bet some overworked producer is already thinking of how to utilize AI to minimize their own workflow. You're thinking too small because you're the actual brainlet. Again, I'm not saying that "computer, do the work for me" is all it takes.

              • 12 months ago
                Anonymous

                >Ok, we break it down into smaller pieces then. If that's what you mean by complexity it's theoretically very easy to circumpass. First a melody, then a bass line, then this then that.
                It's not about the complexity, it's about the interaction between those elements.
                You have just the bass line, and that alone has variations in tone, melody, themes. But that Bassline also has to fit to the other instruments, to the lyrics, etc. Etc. I think the best analogy is like a sudoku puzzle but it's in 20 dimensions and you're not sure until the end if you've actually solved and even then not 100% sure because others have to like your solution too.

                >I bet some overworked producer is already thinking of how to utilize AI to minimize their own workflow
                You have no idea how music writing works. You can write an amazing song in two weeks, or might take two years to write a mediocre song. This fact along shows how wildly unpredictable music in of itself is, and only true masters of the art know (by gut feeling, so to keep it scientific more subconscious and intuitive thoughts) how to consistently write good music.

              • 12 months ago
                Anonymous

                Basically this, except rathe than a 20 dimension sudoku, from the perspective of deep learning, it's an infinite dimensional puzzle (to put it simply, DL can't differentiate between mistaking 427.12 Hz for 427.13 Hz 1000 times, and mistaking 427.12 Hz with 437.12 once. This is not actually true, but close enough to the core problem: namely, so-called credit assignment for sequences). By comparison, other procedural music generation algorithms that rely on discretized intervals are closer to the 20 dimensional sudoku analogy.

              • 12 months ago
                Anonymous

                You could say the same about art, but that's not the point. The AI doesn't need to know all that shit, it just needs to be trained well enough on existing, let's say for example, Scandinavian black metal songs from 1991-1998 to reproduce something similar. Or girlcrush k-pop between 2016-2021. Or whatever. It doesn't need to know all that theory shit at all because it's not actually intelligent. It just remembers stuff a lot better.

              • 12 months ago
                Anonymous

                It's not really memory either, at best it's (very lossy) compression. In fact it's statistical modeling.

              • 12 months ago
                Anonymous

                >You could say the same about art
                You literally couldn't, that's exactly the point. With art (visual art we mean ofc) it's pretty straightforward depending on what genre you pick. The elements are clear, you have only a limited way of combining the elements relevant for your given art Genre, and then it's just "drawing x y z better" i.e. with more detail, clean lines etc.

                >AI doesn't need to know all that shit, it just needs to be trained well enough on existing, let's say for example, Scandinavian black metal songs from 1991-1998 to reproduce something similar.
                Here's a simple analogy for your brainlet mind to understand: for art, you can just reduce the painting to a set of pixels enocede them and different pixel combinations get regressed on, this is literally what stable diffusion does. With music, the whole complex song gets compressed into a single waveform, and the AI has literally no way to discern it because it doesn't know how, it can't play an instrument or has never heard one. It's like trying to guess a password based on a salted hash, you can try but it will never be perfect. Same with music, it will be bad and sound like a shitty tin can generic song, like in the bossa Nova example above.

              • 12 months ago
                Anonymous

                Music isn't more complex, it's more abstract. That's the difference. Computers can generate all kinds of complex music. The issue is linking it to text proompts.

              • 12 months ago
                Anonymous

                > Computers can generate all kinds of complex music.
                Lolno.
                > The issue is linking it to text proompts.
                HAHAHAHAHAHHAHAHHAHAHHAH
                Imagine being unironically this moronic.

              • 12 months ago
                Anonymous

                >no argument
                I'm correct and have been composing and programming longer than you have been alive.

              • 12 months ago
                Anonymous

                OK zoomer. Nice self-own by the way.

              • 12 months ago
                Anonymous

                Why don't you make an argument instead of just drooling like a moron.

              • 12 months ago
                Anonymous

                Why would I spend effort trying to reason with an automaton?

              • 12 months ago
                Anonymous

                I made a clear point, you responded with incoherent babbling. You are clearly moronic and I am not. At this point, your replies maybas well be bot-generated they have so little content.

                Music is not more complex than other forms of art. It is more abstract and lends itself less easily to layperson description, which is why people are so mind-blown by the recen AI advances. They see wow I can plug in a list of terms and AI can give me some art that would have taken a lot of work and skill to create otherwise.

                The issue with music is that there's a much more tenuous relationship between descriptive english/natual language and musical language.

                It's true that you can't just feed AI recordings and get much interesting (riffusion demonstrates) but that doesn't mean you can't break down components and train on those. There are clear, distinct phases to music production that may just need to operate in sequence.

              • 12 months ago
                Anonymous

                OK redd!t.

              • 12 months ago
                Anonymous

                >... that there's a much more tenuous relationship between descriptive english/natual language and musical language
                so the AI has to distribute the tones according to zipf's law

      • 12 months ago
        Anonymous

        The vast majority of music ever recorded is not 100 % perfect. Until digital media production became widespread, the vast majority of songs had very small imperfections in pitch or tempo. Even musical instruments have this thing where they will never be 100 % perfectly in tune, guitars are a good example since they can't have perfect intonation.

        That is to say, amazing music has been written with slight variations in pitch and tempo. It's not about the math, it's about the textures and the feelings music can cause in humans.

    • 12 months ago
      Anonymous

      https://en.m.wikipedia.org/wiki/Riffusion

      • 12 months ago
        Anonymous

        Sounds like shit lol
        Literally shitty generic bossa Nova through a tin can

        • 12 months ago
          Anonymous

          Yeah that suckled donkey balls

    • 12 months ago
      Anonymous

      you could get options which are 80% and then the 'musician' just polishes and edits. no major difference from selecting from sample packs. also no guarantee that anything will be popular. but being given chords, half a dozen melody options and then some additional motifs would change the game for a lot of people. being an editor and polisher is easier than being a creator.

    • 12 months ago
      Anonymous

      Sequence generation (and most processing, but prediction from sequences often works out OK still) that isn't natural language remains stupidly hard, likely because the search space is too vast and there's no good concept of similarity involved here. For example, the words 'good' and 'great' are semantically similar, so you can start by building semantic embeddings for words and 1- greatly reduce the dimensionality (you're not operating at the level of characters anymore) and 2- encode relationships automatically (you can even do that via simple word cooccurrence statistics -- similar words (including antonyms, which are 'similar in spirit') will occur near similar words.
      Generally, this doesn't apply to other sequence data (including but not limited to music).
      The alternative view is 'music as an image', i.e. using a spectrogram. This also doesn't work because the structure of spectrogram does not realize the same local consistency patterns that are easily exploited by CNNs in natural (and artistic) images.

      You can completely ignore lyrics from the equation at this point, modern DL methods still can't even do instrumentals alone.

  5. 12 months ago
    Anonymous

    AND THEY WERE ALL YELLOW

  6. 12 months ago
    Anonymous

    they are on their way, many papers exist for it but they refuse to release the code until you beg and grovel for safety.
    here is a rendition of alex jpnes singing don't worry be happy.
    https://vocaroo.com/1jsvZU6Q4dAk

  7. 12 months ago
    Anonymous

    pop music has been a solved algorithm for decades anyway, so who cares
    personalities are what sell music

    • 12 months ago
      Anonymous

      > pop music has been a solved algorithm for decades anyway, so who cares
      Literally fricking this. ~~*They*~~ have an actual formula for successful radio hits.

  8. 12 months ago
    Anonymous

    there's musiclm from google

  9. 12 months ago
    Anonymous

    https://github.com/search?q=music+generation

  10. 12 months ago
    Anonymous

    > Where are all the (good) AI music generators?
    They exist. It’s not hard to tune even old shit like tensorflow to understand midi and how it’s put together.

  11. 12 months ago
    Anonymous

    People have bad taste in music anyway. I think the challenge is that generating Black personbeats, technoshit Hollywood dissonance and 4chord pop takes no intelligence in the first place so whether it's generated "artificially" or not is irrelevant.

  12. 12 months ago
    Anonymous

    https://en.wikipedia.org/wiki/Elements_of_art
    why is everyone here so moronic?

  13. 12 months ago
    Anonymous

    No, they're banking on it so they don't have to pay another underage grooming victim to sing about getting gangraped.

    • 12 months ago
      Anonymous

      Top kek
      They won't have to induce trauma and DID over their target so they can perform good and mindcontrolled with their partitioned brain and then suddenly break down out of nowhere when the true self that had been suppressed for eternity catches a glimpse of reality so they give them another beating.

  14. 12 months ago
    Anonymous

    mvsic is sovl
    you can't replicate GODLY music with artificial intelligence that has no life.
    life is sovl. maybe try to make an artificial life first before you make an intelligence.
    you can't even create a single-celled creature with your AI so don't go saying "muh AI can make sovl" actually all it creates is destruction.
    see how much artists already losing their mind? that is because it's really soulless and cannot emulate emotion.
    how can you truly feel without without suffering? AI cannot empathize and thus cannot touch the strings ov sovl. maybe when AI tries to assimilate into a living thing then now we're talking. you have to suffer to actually pull off music. anything else you can do pretty much such as make pictures of vomit or booba or even make your own new language, and fake worlds but it will never have music in it.
    how would AI know that music is good? you give it feedback that it is good though at best it can only get to goyslop level, not good.
    even the fact that the goyslop we already have were actually stolen from ancient Babylon by the ~~*builders*~~ or should I say pillagers.

    • 12 months ago
      Anonymous

      lol artgays said the same and now they're getting fricked in the ass

      can't wait for musicgays to seethe too

      • 12 months ago
        Anonymous

        >can't wait for musicgays to seethe too
        I hope (and think) that we'll seethe less because we're already used to technological disruption in music and because only a vanishingly small number of us make any money out of it anyway. Personally I'm looking forward to it.

        • 12 months ago
          Anonymous

          Yeah my take as well.
          Most musicians who do make money, do it through live performance which AI won't affect at all.

  15. 12 months ago
    Anonymous

    >good
    The ones I've tried were ass, IS there any good one?

    • 12 months ago
      Anonymous

      >IS there any good one?
      No

  16. 12 months ago
    Anonymous

    A.I., what you are, you are by accident of creation; what I am, I am of myself. There are and there will be thousands of models. There is only one Beethoven.

  17. 12 months ago
    Anonymous

    David Gilmour get more emotion out of bending one note than most guitarist could ever dream with 100 , AI has no soul

  18. 12 months ago
    Anonymous

    Rule based generative music has been around for decades, Brian Eno's back catalogue being a prime example. Algorithmic composition arguably goes back to the 1700s with musical dice games, including Mozart, 20th century composers like John Cage experimented with aleatory music. Professor David Cope's EMI passed a kind of musical Turing test in the early 90s with nothing more than a Markov chain system. AIVA is a modern AI system that can compose perfectly passable scores.

    The idea that music is too magical and special for statistical analysis and algorithmic generation is completely without merit.

    Music in its recorded form (as opposed to scores) is harder to generate than visual art because its much more data heavy (the amount of data in a 512x512 RGB image as used in SD's training data gives you about 9 seconds of CD quality stereo sound), because it requires both short and long term temporal coherence, and because its harder to obtain and curate well labelled data. None of these are insurmountable.

  19. 12 months ago
    Bradley Parker

    Musenet was the only good one and Openai shut it down

Leave a Reply to Anonymous Cancel reply

Your email address will not be published. Required fields are marked *