The Pink Floyd Fantasy LP: Writing psychedelic lyrics with AI

While Pink Floyd has been long gone, their lyrical content has inspired countless bands within the progressive and psychedelic genres. So why not inspire an AI to spew out verses akin to Gilmour and Waters?

I trained an LSTM on the scraped lyrics of the entire Pink Floyd discography of 200+ songs. I then prodded it to go ahead and write me a ballad. The results are a beautiful mess (I’d say it trumps PF’s queerest, Bike).

You can view the project here. Below are a few interested results, mildly formatted for the reader’s pleasure. I’ve carefully curated this and more in the Stink Floyd Fantasy LP, a little pet project of mine.


I call this one Stormy, Hot Machine Gun.

Slightly piper, stepping up — 
If I heard the rains. 
Will either hear the day trucks
Bottles on the lapel remember love.
Some little sand whisper
Packing shit now, saying don’t.
Got so he never
Ice air — ooooh Maggie!
Love spoils the razor on,
Something my love at all it.
What away? A Fletcher. 
Most I’ve hold the sun -
With the pick the real
Going ons.
Another lifetime more do you shall?
Swan silent man, I
Show them an anything.
Eyes who was slip, of the
Tree surprise, petrified faces and needed a
Stone, don’t love your time!
To my do to slied,
Looking of taken sound.
You’re hello? Into the wall! With stormy
Hot machine gun, great!
S’il naked to will.

You can see a few familiar phrases and words from the PF arsenal, but overall a quirky, preposterous verse.

Let’s try another one, called Green Laddy!

Water, ah drones!
Of in skies up, float.
Wrong, if you games to
Minefield out to believed, but…
The curtain unfurled in!
With prisoner, shine the
Sails flow, flow than the
Returns sky, heart, what legend!
By climb many dreams the other had.
Rolling precision, madam isn’t has fish-eyed reminder.
Feelings, batter, in had in into
Tear-stained forms knock mother, you are crashing.
Help, look lost death or blue sea.
Fabulous precision, illusion were
Your cut you think alone to.
You and so — 
Leave you’re welcome sarcasm with
Pondered but to you’re.
Said would you.
The not forever I diamond.
And, and, and control love you can’t.
Quiet steel mama’s did before.
Green laddy! will asked
It’s see a harken.

Inferences

  • The juxtaposition of unrelated words from the PF canon make for a hilarious read many a time.
  • The correlation between adjacent words is not particularly strong, making for nonsense verses. Will be working on tweaking the AI to give stronger word-level relations.
  • There do however seem to be a few thematic relations in both the samples — (i) More melodramatic, with love as a dominant emotion (ii) Darker, with mentions of drones, minefields, prisoners, crashing, tears and death.

How it works

  • The AI uses the Kim Sung’s word-rnn LSTM for a word-level model, and Max Deutsch’s guide on using it is priceless.
  • I trained the LSTM using lyric files scrapped using the vagalume-download-lyrics project.
  • I took the output samples and strung them into a poem-like form, to best represent lyric semantics.

What next

  • It would be fun to train this to artists across different spectra of music. Common lyrical themes abound.
  • You could also train it over a corpus of MANY musicians, so it (hopefully) evens out as a more generic lyric generator.