Melody AI: RNN-based Melody Generator

Author: Yuzhe You

In this study, three RNN-based models are implemented with different RNN variants (vanilla RNN, LSTM & GRU). A set of MIDI files are collected and partitioned into single-instrument melodies to be used as a starting sequence for the models.

Environment

Python3
Keras Tensorflow

Getting started

Data Processing & Model Training

Place MIDI files (for training) inside the midi_songs folder
Run python train.py [rnn/gru/lstm]
Using a GRU-based or LSTM-based model is recommended for better results due to vanilla RNN suffering from the vanishing gradient problem

Input Sequence Preprocessing

Place the starting sequence (in the form of a MIDI file) in the input folder
Run python input.py
This will process all the sequences that will be used to generate melodies from the MIDI files located within the input folder

Melody Generation

Run python predict.py [rnn/gru/lstm]
This will generate a melody of sequence length 100 using trained model

Sample Outputs

A playlist of sample outputs can be heard here: Melody AI SoundCloud Playlist.

References

The implementation of this project is partially based on the implementation from How to Generate Music using a LSTM Neural Network in Keras.