Inspiration
In DSC 10 we saw a data visualization which used song lyrics from artists to determine how "creative" an artist was. The use of text data is also something we saw in DSC 80 so we were inspired to take things a step further and predict lyrics altogether!
What it does
When the user inputs some initial lyrics into our website, the Sequential model takes those lyrics and probabilistically imputes the next n lyrics, where the user specifies what n is.
How we built it
Using Keras and TensorFlow, we built a Sequential model trained on tens of thousands of words from an artist's lyrics.
Challenges we ran into
Scraping the data was a significant challenge since we wanted to avoid the classic "garbage in, garbage out". Furthermore, our initial model did not produce very coherent lurics
Accomplishments that we're proud of
The final model, while not perfectly coherent, produced some decent lyrics which could certainly be sung! It was truly rewarding to see the model be able to come up with that.
What we learned
We learned the importance of clean, iid data. We learned that NLP models take massive amounts of data to properly train.
What's next for Song Lyric Generator
With more time, we could further improve this project by feeding it more data and using a more advanced model.
Log in or sign up for Devpost to join the conversation.