Efficient Music Instrument Classification Using Mamba Models
Keywords
Loading...
Authors
Issue Date
2025-09-14
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Abstract
Music instrument recognition is a task in which the instrument(s) that play in a piece of music
need to be identified. It is not a fully solved problem, and is therefore an active area of
research. Recently, Mamba models have been explored as a potential alternative to models
such as the transformer. In contrast to the quadratic scaling in transformer models, Mamba
is a model that scales linearly with the sequence length. Mamba models have not yet been
deployed for the task of music instrument recognition. In this research project, Mamba
models that receive Mel-Spectrograms, tempograms and MFCCs as input representations
are trained. The models are trained and evaluated on the instrument recognition in musical
audio signals (IRMAS) and NSynth datasets. The effect of data augmentation on the IRMAS
dataset using a generative model called WaveGAN is also explored. This can be especially
useful for the IRMAS dataset, as some of the instrument classes have a low number of training
samples. The results indicate that the Mamba model that uses Mel-Spectrograms as
input performs best, obtaining a micro and macro F1-score of 0.61 and 0.51 respectively on
the IRMAS testing dataset, and an accuracy of 0.78 on the NSynth dataset. The data augmentation
approach and the use of ensemble models consisting of weighted combinations
of the base models did not improve performance. The models are competitive with some of
the existing models on these datasets, but performance is not quite matched with state-ofthe-
art models. The results could encourage further research into deploying Mamba models
for music instrument recognition and other tasks in the audio domain, as well as exploring
different data augmentation methods for the IRMAS dataset.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen