We saw a post by Chris Johnson on training a char-rnn on Eminem lyrics and thought that this seemed very interesting so we decided to try it ourselves.

What it does

The model was trained on Drake's song lyrics and learned how to rap like Drake.

How I built it

We decided to use a combination of selenium and beautifulsoup4 to scrape all the song lyrics. We scraped approximately 600 songs and used these for the dataset. Then we used Torch7 a Deep Learning library for Lua, and an implementation of a char-rnn created by Andrej Karpathy. After the model was trained we used INOVA a text to speech library and outputted the text to speech for mixing with a beat.

Challenges we ran into

The first challenge was scraping websites using beautifulsoup4 for python. After hacking away at scraping the webpages, we trained the model. Training the model can take very long especially on a computer without a fast GPU, so we were limited in how big the model could be. Furthermore, adjusting hyper parameters for the model also poses a challenge and in the constrained time limit we did what we could. Tuning hyperparameters, scraping websites, and mixing audio.

Accomplishments that we're proud of

We started rather late and came up with the idea around 12 and still managed to finish.

What we learned

We gained a better understanding of how char-rnn's work and the strengths, and weaknesses of the char-rnn.

What's next for DeepDrizzy

We would like to use more data, perhaps dozens of rappers discroaphys to create a model for many rappers. The increase in data will lead to a better

Built With

Share this project: