Evaluating Convolutional and U-Net Architectures for Tonnetz-Based Music Generation and Prediction
Author:
- Akshesh ShahASAkshesh Shah
- 27 June 2025
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-8
Abstract
This paper evaluates the performance of a simple convolutional neural network (CNN) and a modified U-Net model for predicting and generating musical chords using Tonnetz graph representations. The simple CNN provides a baseline for convolutional models, while the U-Net, adapted for smaller grids, includes skip connections to improve performance. Results show that the U-Net excels in individual chord prediction but struggles with key changes and long-term generation. The simple CNN performs better in a continuous generation but suffers from repetitive outputs. The study suggests future work could involve advanced methods like diffusion models to enhance musical generation.Keywords
This publication has 5 references indexed in Scilit:
- GiantMIDI-Piano: A Large-Scale MIDI Dataset for Classical Piano MusicTransactions of the International Society for Music Information Retrieval, 2022
- Exploring Musical Structure Using Tonnetz Lattice Geometry and LSTMs2020
- Hexagonal Convolutional Neural Networks for Hexagonal GridsIEEE Access, 2019
- CNN architectures for large-scale audio classification2017
- Black Atlantic Rhythm: Its Computational and Transcultural FoundationsMusic Perception: An Interdisciplinary Journal, 2002