My Computer Sings a Love Poem

Love in the Public Domain

{{ Give me a new love carousel }}

The love poems in this carousel are all found in the public domain--they are not protected by copyright and are free to use without permission. I put together the collection because I wanted to set the love poems to music (computer-generated music, that is) and I didn't want to violate copyrights.

My original plan was to have my computer write and sing songs using lyrics from each love poem, but I soon discovered that generating one minute of song required the better part of a day. I decided instead to generate a song for just one poem--and just how ever many lines fit into one minute of music.

I chose Helen Hay Whitney's With Music, as it seemed like a good fit, and used Kevin MacLeod's "Ukelele Song" to kick things off--the first 10 seconds of the below piece are written by a human. The computer takes it from there...

With Music

The lyrics are unclear, but much better than in the other two versions generated. I include those tracks below, as well as more details for anyone interested in learning more about how the song was made. In the meantime, my husband Dave--who had not read the poem--offered to transcribe what he heard:

With Music

by Helen Hay Whitney

Dear, did we meet in some dim yesterday?
I half remember how the birds were mute
Among green leaves and tulip-tinted fruit,
And on the grass, beside a stream, we lay
In early twilight; faintly, far away,
Came lovely sounds adrift from silver lute,
With answered echoes of an airy flute,
While Twilight waited tiptoe, fain to stay.

Her violet eyes were sweet with mystery.
You looked in mine, the music rose and fell
Like little, lisping laughter of the sea;
Our souls were barks, wind-wafted from the shore—
Gold cup, a rose, a ruby, who can tell?
Soft—music ceases—I recall no more.

With Music

(as sung, transcribed by Dave):

There's a real need for some germs so gone
I half remember how the birds were singing
among the leaves
and to the tethered fruit and on
the grass beside the stream
remaining in the twilight
there too far away under the true
then sail my real sound
drip some silver reed leavee
real downstairs echo on and then

Here's What Happened:

The Victory of Eros by Angelica Kauffmann
Near Last Camp on Ross Fork, Snake River, Lander Wagon Road, Oregon by Henry Hitchings

For this project, I decided to try OpenAI's JukeBox, "a neural net that generates music, including rudimentary singing, as raw audio in a variety of genres and artist styles." I spent some time browsing the samples the AI made in a variety of genres and in the style of various artists, and settled on Abba/Pop as the sound I was hoping for.

To use the JukeBox model, I followed the instructions Marconan shared on Medium and used the coresponding notebook on Google Colab.

The tutorial recommended starting with a sample, as a sort of inspiration, and so I went over to freepd, where I found Kevin MacLeod's "Ukelele Song." I used ffmpeg to convert the .mp3 to a .wav format, and then primed the model with the first 12 seconds of 'Ukelele Song.' The model figured out the rest!

Here are the other two tracks:

With Music (version b)

With Music (version c-- somewhat punk?)

I hope you enjoy the public domain carousel! And the songs, of course. More carousels and projects here.

Further Reading

Dhariwal, P., Jun, H., Payne, C., Kim, J. W., Radford, A., & Sutskever, I. (2020). Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341.