Skip to content

using machine learning to generate 4 chord song lyrics

Notifications You must be signed in to change notification settings

avparmar/instant-popstar

Repository files navigation

Context: This project was made for Artificial Intelligence Spring 2018 at Northeastern University!

My name: Aditya Parmar

My Project Proposal: Artificial Intelligence to find the next top pop song!

PROJECT DESCRIPTION: What is my project: I am hoping to use some ai algorithms to construct a top pop song by means of scanning in lyrics of other super popular top songs and using a prediction system to spit out a brand new one.

The Problem: I want to figure out how to write the perfect radio song. I don't know how to myself, but I have plenty of past examples.

Additional Detail: Many of the top songs of the past 50 years or so of radio music are songs that can be classified as "I-V-vi-IV" (If you've ever heard of a "four chord song" this is an example of a type of those). Without getting too technical into musical theory, this simple chord pattern works because the chords played in this way are in harmony. This is seen in practice when these types of songs repeatedly end up on the top charts and even end up topping up other such songs, while consumers seem to be fooled into thinking they're getting plenty of variety. My mission is then this: Now that I have a winning chord pattern, all I have to do is come up with the perfect lyrics, and it should be guaranteed to make plenty of money, or at the very least, impress a few people. I compose, but I don't write lyrics, so I figured I would use algorithms to do this part for me. I was inspired to do this project by an AI that participated in a film festival by writing its own screenplay. The result was hilariously entertaining when produced with real people so I'm very curious to see what happens in my application. Obviously, there are a few key differences between this project and the screenplay writer, so there may be some gaps I'll need to fill on my own, but I think this is definitely achievable.

Input: There are several lists of I-V-vi-IV songs. A fairly large one exists on wikipedia: https://en.wikipedia.org/wiki/List_of_songs_containing_the_I-V-vi-IV_progression This list contains a little under 200 songs, which I think should be enough, at least for a class project. The results could be amusing at the least. I plan on scraping the lyrics from youtube subtitles so that I can time index them as well. From there, I will construct an AI with python to make a prediction on words and their timings.

Motivation: Curiosity, Potential Wealth, Learning experience, Fun.

Output: One song's worth of lyrics.

ALGORITHMS: On examining the screenwriting AI's, consulting the internet for similar tasks, and also speaking with more advanced CS students on their opinions for recommended algorithms, a couple algorithms were mentioned in common. They are: LSTM: Long short term memory. While I don't know much about this yet to talk about it in detail, it can produce predictions for time based word lists. A common example on pages is an AI that translates text from one language to another. If I can extract lyrics and timestamp them, I can potentially achieve a similar result.

Sequence to Sequence: This is more advanced, but it's still a recurrent neural network, like lstm. The theory is that the
grouping of the words can be used for prediction and these carry over. 

I have decided to move forward with LSTM because it seems promising and well documented.

Resources for learning: There exist plenty of papers and tutorials on these algorithms.

  Wikipedia: https://en.wikipedia.org/wiki/Long_short-term_memory
  Tutorial on lstm: https://deeplearning4j.org/lstm.html
  A dissectable, practical application: http://colah.github.io/posts/2015-08-Understanding-LSTMs/
  tensorflow details: https://www.tensorflow.org/tutorials/recurrent#lstm

RESULTS: I expect to have, in some form, a song, created from the average of all other top similar tune songs. I will then attempt to produce a real song out of this, perform it on ukulele, and record the performance, and note how much traction it gets. A surprising amount of radio songs come from obscure spots, so it's not unlikely that my song could later appear on the radio and I acquire immense wealth.

About

using machine learning to generate 4 chord song lyrics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages