Lyrics, Pt. 2: Automated Songwriting
For quick access to SongwriterBot, click here!
In Part 1 of my lyrics project, I scraped thousands of song lyrics from Genius and developed a model which classifies songs into genres based solely on their lyrical content. Now, in the second and final [okay, maybe not final] installment of the project, I’ll provide a brief explanation of how I developed SongwriterBot, which writes its own randomly generated songs.
This project utilizes Jeremy Singer-Vine’s amazing markovify package to create separate Markov models from the scraped songs of each genre. If you read Part 1, you’ll know that I scraped 1,500 songs per genre; here, I scraped approximately ten thousand country, metal, pop, rap, rock, and soul songs. For each genre, every song was split line-by-line and combined into a Markov model, which can randomly generate sentences from the text.
The following picture does a good job of demonstrating how the Markov models generate sentences. In this example, the Markov model includes just two sentences: “Mary had a little lamb” and “Mary had a giant crab.”
At the beginning of the sentence, the model only knows one possible word: Mary. Directly after Mary, the model has only seen one word: had. After had comes only a. After a, though, the model has seen two possibilities, each an equal number of times: little and giant. It randomly picks one of the two options, and continues as normal. It is this exact process – albeit, with far more possibilities – that occurs hundreds of times during the creation of one SongwriterBot song.
While back-to-back lines aren’t always thematically consistent, the content of the lines themselves tends to imitate each genre very well. Here is an example stanza for each genre:
Country: but i don’t understand / it would take my hand / and i won’t let me get my wheels / the screaming wheels and blackjack deals
Metal: look at the seventh veil / too blind to see the white whale / we’re on our crooked scale / divine - night in jail
Pop: you hate the word and we can come on and dance / don’t want to see the private jets to france / he’s giving me the vodka skip the criss / we are what they say that you would never miss
Rap: i smoke on some g s**t / international bring back piper’s pit / i scratch off on they dream / so when i’m with that guillotine killer team
Soul: don’t take your love from the start / love will never be apart / whenever you want to stick / that we have come here quick
Even better, we can also create “genres” that don’t actually exist. I added the entirety of the Harry Potter and A Song of Ice and Fire (more commonly known as Game of Thrones) book series as separate Markov models. markovify also allows users to combine multiple models – here are a couple amusing stanzas of a Harry Potter & rap song:
Harry Potter / Rap: i shouldn’t say s**t and they fall / the room at the yule ball / i let you have your mother’s eyes / but we’ve got to be cooked in pies / you give them to the ministry instead / grab the chrome to ya head / any form of cedric stood there / i walk in the cliff-top garden and into the air
If you’re interested in generating your own songs with SongwriterBot, click here. Because Heroku limits free web applications to 500MB, I removed rock, Harry Potter, and A Song of Ice and Fire models from the available options, as well as the ability to combine genres. Note that the app still occasionally runs into memory bugs and timeouts, which I am working to fix.
Once you generate a song, you can save it – which I recommend, seeing as the program will never create the same song twice – or try out your new lyrics against an instrumental track for any genre. Hope you enjoy!