Brown Deep Learning 2024 Final Project
KAL-Duet is a deep learning model that can generate piano accompaniments given an input melodic sequence. It is a Transformer model that uses the attention mechanism to learn the relationship between melodic notes and the harmonic structure accompaniments.
The model is trained on a combination of the Lakh and Maestro datasets, which provide a diverse range of MIDI files. The Lakh dataset offers a large volume of data, while the Maestro dataset provides high-quality piano recordings. Preprocessing information and directions can be found in the README.md files in src/preprocessing and src/data.
Run src/main.py with arguments to train the model. More information can be found at the top of the main.py file.
Read the README.md in src/testing for more information on how to generate outputs. With our training, we were able to achieve an accuracy of around 60% on LAKH and Maestro datasets. We were able to generate valid MIDI outputs, but they were often chaotic or dissonant.
This project was an attempt at recreating the first paper below, and gathered inspiration from the rest of the linked papers:
http://arxiv.org/pdf/2002.03082
https://www.duo.uio.no/bitstream/handle/10852/95694/1/UiO_Master_Thesis_benjamas.pdf