AC/DC Song Made with a Markov Chain | Hacker News Reader

AC/DC Song Made with a Markov Chain

youtube.com

49 pointsimpostervt15 comments

Honestly, this doesn't read like Markov chain output. It feels more like a human writing what they think a Markov chain would do. Or maybe autogenerating a bit of it and then having the human riff on whatever nonsense the generator produced to fill in most of the song.

The creator of something similar (but with Nickelback) posted on Reddit last week and went into a bit more detail, but essentially, he created a bunch of lyrics with a markov chain, then arranged them into the structure of a song.

https://www.reddit.com/r/Music/comments/gc68t3/nickelbot_nob...

I don't buy that explanation, there's clearly a ton more human involvement here than just rearranging verses.

The output is far too grammatical and maintains too much thematic consistency across filler word distance to be output by a low-order Markov chain, but is too unique compared to the (very small) corpus to be output by a high-order one.

The music is more or less the structure to For Those About to Rock by AC/DC (1). Pretty amusing though, and AC/DC would be a good test drive for an AI generator because the song structure and lyrics don't have a ton of breadth.

(1) Compare to: https://youtu.be/fKhTk0IynHM?t=78

Yeah, the whole notion of this being something generated with a markov chain feels too much like hype.

The guy admitted on reddit that he manually cherry-picked and assembled the lyrics, and then he sang and recorded the song himself.

For all intents and purposes, he might as well have just fed a bunch of AC/DC lyrics through a word randomizer and then assembled the new lyrics by hand. No AI or markov chains required.

On the other hand, here is ACTUAL real music generated and played entirely by AI: https://openai.com/blog/musenet/

It seems like the author only used a Markov chain to produce the lyrics, but the rest is sung by an AC/DC impersonator. It's still quite impressive, but I'd love to see the music itself generated by GANs as well.

https://www.youtube.com/watch?v=iJgNpm8cTE8

Hot damn, technology is catching up with humans' manual creations. Your video is reminiscent of Adriano Celentano's song "Prisencolinensinainciusol"[1]–whereby he created a song to represent what English sounds like to non-native speakers[2].

[1] https://www.youtube.com/watch?v=-VsmF9m_Nt8

[2] https://en.wikipedia.org/wiki/Prisencolinensinainciusol

If the lyrics were all generated by a Markov chain from beginning to end, there's no way that it would have a repeating chorus.

Are all the 3-word tuples actually used in other AC/DC songs? (Output this grammatical could not be generated by a Markov chain that only looks at 2-word tuples.) If not, that also seriously calls their claim into question.

I think the maker would have done well to be more upfront: "AC/DC Song Co-authored by a Computer" or something like that.

Love it!! One day we can just “order” a “new” Hitchcock movie or Crichton novel?!

More than that. If you don't like the first page you can just start a fresh one.

> Also let me say that trying to sing like Brian Johnson is extremely hard.

I think he did a very good job! You can easily tell the difference side-by-side, but still, it's close enough that I thought it was deepfaked at first.

would love to see a blog post about it!

cooool!!!

1000x times more listenable than an actual AC/DC song.

Reply on news.ycombinator.com