First pop songs composed (semi)-entirely by a computer

From Fact Magazine comes the first pop songs composed entirely by computer using artificial intelligence (AI) programs. Now of course the machine had to be programmed; otherwise it would just emit random combinations of sounds and noises. And, as Turing predicted, it couldn’t write comprehensible lyrics, either; those came from a human. As the site notes:

The song, which is called ‘Daddy’s Car’, was composed by an AI system called Flow Machines.

The Flow Machines software got its music knowledge from a huge database of sheet music with songs in varying styles and wrote that track after being given a style prompt from a human composer. The melody and harmony was composed by AI and then a human musician, French composer Benoît Carré, produced, mixed and wrote lyrics for the track.

‘Daddy’s Car’, which you can hear below, is expected to be on an album of songs entirely composed by AI due out in 2017.

This one is said to be composed “in the style of the Beatles”. Well, yes it is, but it isn’t anywhere near as good as anything the Beatles ever did—except, perhaps, Octopus’s Garden. Still, it’s okay, and I have to say it’s at least as good as a lot of the autotuned, repetitive, soulless crap produced by today’s pop stars. It’s clear that, like classical music and opera, rock has run its course. It’s equally clear that it’s a long way before computers will even come close to replacing human composers.

Here’s another AI-composed song, “Mr. Shadow,” described like this:

“Mister Shadow” is composed in the style of American songwriters such as Irving Berlin, Duke Ellington, George Gershwin and Cole Porter. French composer Benoît Carré arranged and produced the songs, and wrote the lyrics.

Now this one just sucks! George Gershwin and Duke Ellington my tuchas! It sounds more like a mashup of Rudy Vallee and Davie Bowie. But feel free to disagree.

 

61 Comments

  1. GBJames
    Posted September 26, 2016 at 1:36 pm | Permalink

    Well, OK, the music is composed by AI. But what/who is performing?

  2. sensorrhea
    Posted September 26, 2016 at 1:37 pm | Permalink

    Far prefer “Come on Dad, Gimme The Car” by the Violent Femmes.

    • frednotfaith2
      Posted September 26, 2016 at 8:32 pm | Permalink

      I’ll go with that as well as the Beatles’ “Drive My Car”

  3. Posted September 26, 2016 at 1:42 pm | Permalink

    I like the first one better than a lot of what is called music these days.

    • peepuk
      Posted September 27, 2016 at 2:45 am | Permalink

      Agree, it’s not really my taste (maybe because I don’t like the Beatles) but it’s above average.

      • Ken Kukec
        Posted September 27, 2016 at 10:19 am | Permalink

        I recognize that more than a few Gen Xers have said they hate the Beatles simply as a reaction to having Boomer music shoved down their throats. But I’ve never heard a music aficionado say they dislike the Beatles’ music qua music. It would be akin to hearing an oenophile announce that they loathe Chateau Lafite.

  4. Randall Schenck
    Posted September 26, 2016 at 1:49 pm | Permalink

    Seems the computer has a ways to go musically. Not for me thanks…

  5. Posted September 26, 2016 at 1:49 pm | Permalink

    Coming from someone who dislikes techno and most synthetic sounds, excepting a few from the 80s (I can’t help it; grew up listening to U2), these are crap!

    They sound like machines.

  6. Posted September 26, 2016 at 1:53 pm | Permalink

    “it’s at least as good as a lot of the autotuned, repetitive, soulless crap produced by today’s pop stars”

    I felt that is was “autotuned, repetitive, soulless crap” so there may well be a market, just not me.

    • Posted September 26, 2016 at 2:06 pm | Permalink

      “I felt that is was “autotuned, repetitive, soulless crap” so there may well be a market, just not me.”

      At least a computer has an excuse for autotuned soullessness, and unlike much of that this isn’t at the top of the pop charts. I guess we’ll have to wait for computers to be sexy with big butts for that to happen.

  7. Posted September 26, 2016 at 2:12 pm | Permalink

    Well, to claim that this was, “was composed by an AI system” is, it seems to me, a little excessive.

    The system can only compose within the rules supplied by the programmers. So, who is really composing here? How can you show there’;s even a parallel between what the algorithm does and what a human composer does?

    This isn’t like chess, where there’s a pre-defined goal that the system can seek on.

    • compuholio
      Posted September 27, 2016 at 4:27 am | Permalink

      The system can only compose within the rules supplied by the programmers

      Actually, that is not quite true (depending on what exactly you mean by “rules”). Although machine learning algorithms are programmed and the computer follows this programming precisely, it is a very different type of programming.

      You only define a mathematical function (typically with many free parameters) and the machine is programmed to optimize the parameters with respect to an error function.

      Just to come up with an example in music: Assume you have lots of music available for the machine to learn on. The machine could learn a probability distribution for a note, given the last 10 notes.

      It then can “compose” music by sampling from this probability distribution. The result will be something that is based on the observed patterns but is not pre-determined in any way.

      The programmer has not explicitly coded the patterns by which notes are to be generated (at least not in any meaningful sense)

      • Posted September 27, 2016 at 11:24 am | Permalink

        Also, one can ask the fundamental response to many criticisms of AI: *how do you think humans do it*?

    • compuholio
      Posted September 27, 2016 at 4:32 am | Permalink

      Addendum: If you like to read more about how this works I can only recommend the following blog post by Andrej Karpathy who has done some amazing work on recurrent neural networks.

      In this case he teaches them to generate texts:
      https://karpathy.github.io/2015/05/21/rnn-effectiveness/

  8. Posted September 26, 2016 at 2:16 pm | Permalink

    Seems to me the AIs have not gleaned the hierarchical relationship between scale-degrees in tonal music from the songs they were fed. It’s this hierarchy of harmonic function (which may or may not be observed by a composer, depending on context) that makes one harmony following another sound “right”. These harmonies sound more or less random to me.

    • Posted September 26, 2016 at 2:20 pm | Permalink

      (Although the first is better than the second.)

    • Posted September 26, 2016 at 2:20 pm | Permalink

      Agreed. And very wooden (in the stilted sense).

  9. Posted September 26, 2016 at 2:19 pm | Permalink

    I find the first one uninteresting. (As you noted, nothing like the real Beatles — not even close to Octopus’s Garden IMO.)

    The second is just awful.

    I don’t think a machine (at least as they stand today — I’m not predicting the future of AI, from which I think consciousness is not precluded) can produce the subtleties (timing, phrasing, dynamics, etc.) that makes music really interesting to humans.

    If you listen to some favorite music played in a computer (e.g. Guitar Pro; similar SW) where the score is interpreted literally, it sounds dead. (Unless the score is insanely prescriptive.)

  10. Craw
    Posted September 26, 2016 at 2:24 pm | Permalink

    Classical run its course? I suggest this means you don’t know much contemporary classical music, which is vibrant and varied, having cast off the grip of serialism.

    • Posted September 26, 2016 at 2:39 pm | Permalink

      “having cast off the grip of serialism”

      Thank Hank for that! 🙂

  11. bric
    Posted September 26, 2016 at 2:27 pm | Permalink

    Alan Turing programmed the Manchester University computer to play music, and a recording made by the BBC in 1951 has turned up today

    https://www.theguardian.com/science/2016/sep/26/first-recording-computer-generated-music-created-alan-turing-restored-enigma-code?CMP=twt_gu

    • Flemur
      Posted September 26, 2016 at 6:54 pm | Permalink

      Playing, but not generating a new song.

    • Posted September 27, 2016 at 11:26 am | Permalink

      I’ve always found it interesting that producing art (including music) has always been a thing with computers from early on. I wonder why that is? I know it seems to echo my experience – as a child when I learned programming, my reaction was that I had discovered the most wonderful artistic medium ever.

  12. Posted September 26, 2016 at 2:30 pm | Permalink

    Reblogged this on The Logical Place.

  13. Kevin
    Posted September 26, 2016 at 2:32 pm | Permalink

    Second one was better. I’ve heard better AI music. This stuff is trying to sound to popular. The best AI will be innovative for me. The best AI for most everyone else will sound like a Disney channel song….and that’s not hard.

  14. David Harper
    Posted September 26, 2016 at 2:58 pm | Permalink

    “Now of course the machine had to be programmed; otherwise it would just emit random combinations of sounds and noises.”

    Much like the works of Karlheinz Stockhausen, then.

    • bric
      Posted September 26, 2016 at 3:59 pm | Permalink

      This is a strange thing to say about Stockhausen, an adherent of ‘total serialism’, which extends rational control into all musical elements such as duration, dynamics, and register as well as pitch.

      • Posted September 26, 2016 at 6:19 pm | Permalink

        Yes, serialism is not random, but it can certainly seem that way when compared with tonal music. The pitch-choosing heuristics in tonal music are derived from intrinsically musical considerations and how humans typically perceive pitched sound(s). This leads to an organic, perceptible musical logic. Serialism, by contrast, imposes a pitch-choosing algorithm based on musically extrinsic logic. The result doesn’t cohere the way tonal music does. This can come across as sounding random.

        • Barry Lyons
          Posted September 26, 2016 at 6:32 pm | Permalink

          And if anything has run its course in the classical realm, it’s serialism. Many composers today embrace conventional melody and tonality however heavily chromatic (“dissonant”) some if might be. True, there’s still some avant-garde stuff being written, but because of this widespread (re)embrace of tonality some writers refer to this as the “New Tonality”.

          As for Stockhausen, I never cared for him. I much prefer the orchestral works by Ligeti (“Lontano” and “Atmospheres” are two favorite works). There’s a sensuous quality to his works, whereas Stockhausen (from the little I know of him) sounds arid to me.

        • darrelle
          Posted September 26, 2016 at 7:23 pm | Permalink

          I love it when you talk like that.

          • Posted September 26, 2016 at 7:43 pm | Permalink

            lol

            Sexy does come naturally to musicians. Especially organists.

        • bric
          Posted September 27, 2016 at 3:56 am | Permalink

          I broadly agree, and yet there have been serial/twelve-tone pieces that both cohere musically and can move the receptive listener – Berg’s Violin Concerto and ‘Lulu’, Schoenberg’s ‘A Survivor from Warsaw’ and Piano Concerto, Stravinsky’s ‘Agon’. Perhaps all these examples have a whiff of tonality about them that gives us an entry-point . . .
          I even heard performances of Stockhausen’s Kontakte and Gesang der Junglinge at the Goethe Institue in Paris with sound direction by the composer which were quite affecting.

          • Posted September 27, 2016 at 12:11 pm | Permalink

            Yes, the row Berg uses in his violin concerto is constructed so as to suggest tonal relationships.

          • Posted September 27, 2016 at 12:25 pm | Permalink

            I should add that my original comment wasn’t intended to pass judgment on either system.

            I also agree with your point about the suggestion of tonality conferring a familiar kind of coherence.

  15. BobTerrace
    Posted September 26, 2016 at 3:01 pm | Permalink

    Sounded to me like it was processed through a meat grinder. Maybe the genre could be called meat loafyor burger-raw.

  16. darrelle
    Posted September 26, 2016 at 3:21 pm | Permalink

    Rock run its course? Not so! There is plenty of good, new rock music. Though I must admit, pop is certainly much bigger. It’s all about making money. I think the “problems” with the current popular music industry and the current movie industry (both producing mostly highly polished turds, or mediocrity at best) are a result of the same influences.

    Classical music isn’t dead yet either!

    • Posted September 26, 2016 at 6:21 pm | Permalink

      I’m getting better!

    • Bric
      Posted September 27, 2016 at 12:41 am | Permalink

      During the course of his epic three night stint on BBC Four over last weekend Keith Richards opined that playing rock without the blues you are ‘just a pop star’. On technology, he owns neither a mobile phone or a computer so I think we can guess his position on AI music

  17. keith cook +/-
    Posted September 26, 2016 at 3:46 pm | Permalink

    From LiveScience:
    “These are trends – but the perception among many listeners is that this homogenisation of music has taken a big leap forward in recent years. And there are a couple of important technological developments that have made this happen.”

    http://www.livescience.com/56140-invention-made-pop-singers-sound-the-same.html

    And into the BLANDNESS we go, well some… if Starwars is anything to go by at the very lest, the musicians will get weirder and probably a bit more interesting.

  18. Ken Phelps
    Posted September 26, 2016 at 4:01 pm | Permalink

    The second one sounds a lot like regular music being played backwards by one of those bat-shit crazy preachers a few years back. It would be interesting to play it backwards and see if it sounds any better, and perhaps learn who’s trying to subliminally influence us.

    Well, enough of this. I just had a sudden urge to go sacrifice a goat…

  19. Dick Veldkamp
    Posted September 26, 2016 at 4:11 pm | Permalink

    Back to the pre-computer age: recreation of modern song with 1930s instruments:

    [PCC(E): If this embeds automatically, feel free to remove.]

  20. Posted September 26, 2016 at 4:39 pm | Permalink

    These are pretty horrible songs. Rock has not lost its moxie, I think, rather, radio has as has the moneyed interests behind popular music. There are hundreds of Indie groups making good rock ‘n’ roll but you have to look for it rather than just scan thru the am/fm radio channels.

    I have a Spotify list that I think proves this, with about 400+ songs from 2010 forward that exemplify music every bit as good as much from the 1950s-1970s and later. I have a small sample (130?) on YouTube.

    If you think the scene is moribund, please listen.

    • Posted September 26, 2016 at 5:58 pm | Permalink

      Sorry, link from my name was truncated – now fixed? Or direct link is

  21. Henry Fitzgerald
    Posted September 26, 2016 at 6:02 pm | Permalink

    Not able to listen to that here, so I can’t comment on quality.

    But I’ve come round to liking the idea of AI-composed music.

    I don’t think that, for instance, the classical era really was fully exhausted. I think it’s just that Mozart, Haydn and Beethoven all died and there was nobody left with the right skill set to compose the same music they did. Now, marinated in the sounds of the 21st Century, it’s simply impossible for any human, however talented, to write like that.

    But maybe it’s not impossible for a computer.

  22. Barry Lyons
    Posted September 26, 2016 at 6:04 pm | Permalink

    Beethoven and friends have NOT run their course! In fact, just the other day I was listening to my favorite Beethoven symphony, the Sixth (“Pastoral”).

    But if you mean to say that music by Haydn, Mozart, Beethoven, Schubert, Chopin, Brahms et al. (including some “moderns” such as my favorite, Prokofiev) has run its course in terms of the day-to-day lives of most people, I’m afraid to say that it’s probably correct. Too bad. Beethoven is king of the hill in my book. Nobody can touch him.

    • Martin Knowles
      Posted September 27, 2016 at 9:09 am | Permalink

      Anyone who thinks classical music has run its course, probably never got much into it in the first place and has obviously never taken the time to hear the magic of Schubert’s 600 timeless
      lieder.

      • Barry Lyons
        Posted September 27, 2016 at 9:36 am | Permalink

        Indeed. And the big Schubert moment for me in recent years was the discovery of his four-hand piano music. I knew the “Fantasia” but lots more of this music is incredible. I love this box: https://www.amazon.com/Franz-Schubert-Piano-Music-Hands/dp/B00JGE6EM4/ref=sr_1_9?s=music&ie=UTF8&qid=1474986853&sr=1-9&keywords=schubert+four+hands

      • bric
        Posted September 27, 2016 at 9:59 am | Permalink

        I listen to a lot of very varied music, but Schubert is there every day; I have 24 versions of Winterreise (not counting the many on Spotify) so go figure

        • bric
          Posted September 27, 2016 at 10:06 am | Permalink

          I had to check – 188 Winterreisen on Spotify, some will be duplicates but still enough there to justify the subscription 🙂

        • Barry Lyons
          Posted September 27, 2016 at 10:37 am | Permalink

          Wow. I’m more of a piano solo/chamber music/orchestral music guy, and so at the moment I only have one recording of the work (Souzay). A friend of mine says he has about a dozen. But I do want to get around to Schubert’s songs (especially the cycles), and I suspect a few more “Winterreise”s will enter my collection. (I also use Spotify, but I haven’t subscribed yet: I just sit at my desktop Mac and hit the mute key when adds come on).

          • Martin Knowles
            Posted September 27, 2016 at 10:21 pm | Permalink

            With Schubert’s lieder you can’t go wrong with the song cycles. I have about a dozen versions of Winterreise and Die Schone Mullerin. Can’t go wrong with Fischer Diskau as a jumping off point.

            • Barry Lyons
              Posted September 28, 2016 at 7:33 am | Permalink

              I was just remarking to a friend a few weeks ago that I don’t much care for Fischer-Dieskau’s voice. Something glottal there that bugs me. I know that’s heresy, but there it is. I much prefer a classic recording by Souzay, who not only has a beautiful voice but seems more emotionally engaged with the music (I’m referring to Souzay’s peak years captured in a Philips 4-CD box that is now out of print).

              • Martin Knowles
                Posted September 28, 2016 at 12:10 pm | Permalink

                Thanks for the tip. I’m unfamiliar with Souzay and look forward to hearing his work.

              • bric
                Posted September 28, 2016 at 4:43 pm | Permalink

                My favourite contemporary interpreter is Christoph Pregardien (historically it would be Hans Hotter, just listen to Der Leiermann it’s heart rending). Pregardien has made four recordings so far, including two with chamber orchestrations, but the recording that really caught my attention is with Andreas Staier playing a fortepiano.
                I find Souzay’s German rather over-careful, but he is marvelous in Ravel and Faure

  23. Flemur
    Posted September 26, 2016 at 6:40 pm | Permalink

    I have a bunch of music generation software (not written by me), some of which generates based on fractals (pick parameters) or other no-input-source algorithms. The output doesn’t sound like normal songs, but it’s sometimes pleasant and often amusing to listen to.

    But …
    “The Flow Machines software got its music knowledge from a huge database of sheet music with songs in varying styles and wrote that track after being given a style prompt from a human composer. The melody and harmony was composed by AI and then a human musician, French composer Benoît Carré, produced, mixed and wrote lyrics for the track.”
    …could be the “Nth Order Eddington Monkey” algorithm. I used the text above and ran it thru an NOEM (which I wrote) with orders 2,3,4 & 5. The order is the number of characters it looks at to determine the next output character. Replace characters with note values (tone, length, etc) and you may have something like the “Flow Machine” algorithm.

    Here’s the first sentence from each:
    2:
    The promposer trach compt the goter Benching gote goter track a hat Caren and wromposed, Flody wroduced an, produced being gote lyrician and an a huge lyrics musics its composer.

    3:
    The melody and the Flow Machines songs in varying style produced, mixed and harmony was composer. The that track.

    4:
    The melody and then a styles and wrote lyrics for the track.

    5:
    The Flow Machines software got its music with songs in varying styles and wrote that track.

  24. Jenny Haniver
    Posted September 26, 2016 at 6:44 pm | Permalink

    Not sure just how this relates but David Cope, a prof. of music at UC Santa Cruz, developed an algorithmic or should I write “algorhythmic” ‘music generator’ some years ago. http://artsites.ucsc.edu/faculty/cope/; http://www.radiolab.org/story/91515-musical-dna/. His program, called EMI “deconstructs the works of great composers, finding patterns within the voice leading of their compositions, and then creates brand new compositions based on the patterns she [EMI] finds. But it’s not just copy and paste. She brings something new to the pieces.”

    It really is pretty amazing; but once I had an epic musical hallucination, which was active for around ten days, every waking minute, in which I created the same kind of music — in the style of all the great classical composers and genres, from early music to contemporary — and one Zulu chorus. The most perfect fidelity, surround-sound, but all in my head. I’ve been trying to figure out what was going on ever since, but alas, Oliver Sacks is dead. Was it epilepsy or Memorex?

  25. Saul Sorrell-Till
    Posted September 27, 2016 at 4:25 am | Permalink

    I don’t think rock has run its course any more in this decade than in the previous three. It’s not moving forwards perhaps, but that’s not the same as saying there aren’t any good bands around.

    The first song is very strange…almost plausibly human but with the weird wooziness and semi-random feel of all this kind of stuff. I don’t see how an AI is ever going to narrow down on a good melody without the help of a human to filter the vast majority of crap that they produce. After all, that’s all I’ve ever done when writing: trying things out, semi-randomly, and listening out for anything that sounds good.

    Besides, I like to think there are three good ways of defining art(they’re the best I’ve managed to come up with anyway):

    – It amounts to something greater than the sum of its parts

    – It’s useless in any practical sense

    – it’s by definition impossible to reduce to a strict blueprint that can be used to automate the process.

    So I don’t think there’s a secret ‘how to create art’ instruction manual that can ever be produced, no matter how knowlegeable we become about neuroscience or physics. This may be just because art is genuinely irreducible even in principle, but even if that turns out not to be the case any ‘success’ at automating the process would render the art in question worthless, like discovering an endless supply of gold or diamonds.

  26. TJR
    Posted September 27, 2016 at 4:31 am | Permalink

    It should be fairly easy to get a computer to write hip-hop.

    A slow, plodding beat and someone chanting inane rhyming couplets about how great he is.

  27. Dominic
    Posted September 27, 2016 at 6:02 am | Permalink

    You may be interested to hear this very early computer music
    http://newatlas.com/oldest-computer-music-recording-restored/45617/

  28. Martin Knowles
    Posted September 27, 2016 at 9:12 am | Permalink

    I really liked the 2nd song Mr. Shadow. It’s quirky and unexpected. Forget about the claims of Americana, it’s an interesting piece in its own right.

  29. Ken Kukec
    Posted September 27, 2016 at 9:45 am | Permalink

    Seems the first AI has been listening to Brian Wilson’s post-Pet Sounds oeuvre, some White Album outtakes, and odds & ends from Ray Davies’ catalogue with the Kinks. (Which, come to think of it, could also describe the Gallagher brothers from Oasis.) Hell, even the title comes from the opening lyric of “Fun, Fun, Fun.”


%d bloggers like this: