Introduction

This project was first started during the Summer of 2025, as a side project with a friend of mine. We wanted an ambitious project that would drive us to learn more about generative AI in general. Most of the work during that time was some preliminary exploration of the available datasets.

I then pursued the project with other teammates at Georgia Tech for the CS 7643 (Deep Learning) class. It was then that we built the actual model and trained it on the available GPU cluster.

As a proof of concept, we decided to train the model on completion of audio sequences. We made various experiments to study the effect of some hyperparameter on the performances of the model. The final model we trained was about ~100M parameters. For IP reasons, we cannot provide the repo with the source code of the model. However, we gave some samples of generated audio sequences. Those are completion of famous classical pieces.

For more information, please contact me at tristanscuillereng@gmail.com.

Samples generated by the model

Mozart - Sonata in F Major K280

Rachmaninoff - Etude-Tableau Opus 39 No. 8 in D minor

Rachmaninoff - Prélude Opus 32 No. 10

Schumann - Toccata in C Major Opus 7