The Bob Dylain
Project
A creative AI exploration
Can humans and AI work together? Hoping to explore this question, I trained an AI model on the songwriting catalog of Bob Dylan and asked the model to generate new lyrics in Dylan's style. The model generated hundreds of songs, and I sorted through thousands of documents containing new lyrics. This video is an example of the lyrics that were generated—dense with opaque symbolic imagery, visceral descriptions of moments both routine and extraordinary, and a connection to the American west.
To generate new music from these lyrics, I partnered with musician and writer Tony Attwood, maintainer of the Untold Dylan website and Facebook Group. Tony created original music and melodies for the lyrics, and masterfully matched the tone of the lyrics to the music.
Whether or not the output is "good" or "just like Dylan" wasn't really the point. There are some similarities to Dylan's work, but his songwriting spans so many decades, genres and themes that it's impossible to say if the output is "Dylan-esque." What I enjoyed most about the process was the moment I first heard Tony's work—his ability to take what was essentially computer code and turn it into 3 minutes of human emotion shocked and delighted me.
In order to get to that moment, I immersed myself in the AI creative process as much as I could. There are many options for AI text generation, but I wanted to get as close to the code as possible. I found an open-sourced Google Colaboratory GPT-2 project I could train with basic code manipulation. I then tracked down a complete data set of Dylan lyrics—over 700 songs—and cleaned it up for use in my model. I then asked the model to generate lyrics of various length and "temperature" (the variable that determines how "like" or "unlike" the generated text will be relative to the original data set. then I sorted through thousands of documents to find a set of lyrics that felt appropriate to give to a musician. In some cases I lightly edited the output to make the format more song-like. Finally, I connected with Tony Attwood to generate new music. This process was essential in understanding the cycle of human/AI creativity. From idea to data, to output, it's clear to me that a human/AI creative partnership can be a rich one. As more creative people begin to incorporate AI into their work, it's crucial to maintain a human connection in everything we do.
I've written in more depth about the project at TheHuesThatMakeUpTheNight.com