tonybeltramelli/deep-lyrics

Lyrics Generator aka Character-level Language Modeling with Multi-layer LSTM Recurrent Neural Network

Deep-Lyrics

Lyrics Generator aka Character-level Language Modeling with Multi-layer LSTM Recurrent Neural Network

The goal of this project is to generate completely new original lyrics inspired by the work of an arbitrary number of artists.

Description

This repository contains 4 main components:

• A web parser to gather lyrics online
• A preprocessing program to transform the lyrics into a computation-friendly format
• A program to train a LSTM model to fit the data
• A sampling program to generate new lyrics based on the learned data

The Deep Learning algorithm is implemented and tested with TensorFlow version 0.10.0rc0.
The parser is gathering lyrics from songmeanings.com which does not provide any API to request data. Therefore, it is needed to manually find the IDs of the artists you want to get inspired from and pass them to the script. The fun thing is that it does not matter if the artists you pick are related in style or not, the algorithm will learn from all of them; which can obviously lead to some cool results!

Usage

Try it yourself by running example.sh with your own data.

# parse web pages to retrieve lyrics, concatenate them and save them locally in a single file
./gather.py --output_file data.txt --artists "artist_ID_1, artist_ID_2, artist_ID_3, artist_ID_4, artist_ID_5"

# create a vocabulary file containing a binary representation for each character
./preprocess.py --input_file data.txt

# train the LSTM model to fit the lyrics data
./train.py --training_file data.txt --vocabulary_file data.vocab --model_name lstm_regression_model

# generate new lyrics and save them in a file
./sample.py --model_name lstm_regression_model --vocabulary_file data.vocab --output_file sample.txt --seed "Oh yeah"

Note

I highly recommend these two great articles to anyone willing to understand how Recurrent Neural Networks works and particularly LSTM:

• The Unreasonable Effectiveness of Recurrent Neural Networks
• Understanding LSTM Networks

Example

Example of lyrics generated with this code (slightly edited to fix typos). The resulting text is understandable but does not make any sense at all, which is quite funny!

Oh yeah, you made my clung,
When you wan't see there love what's gone on the back falling

These spired on the light for seeing

Whatever you say, whatever you do, you're the one the different that feel
You're gonna take it
And you wanna feel right

The baby gone far, kiss in the rain, but you, back home
They'll never get back when you've go

So you have been work

They come back home
But hidden I can't be there
The baby give is all the give

There is no on the way mind
We clink to go do the back and long, come

Have fun!