1 of 13

Dynamical System Modeling and Stability Investigation�DSMSI-2025

May 08-10, 2025, Kyiv, Ukraine

AUTOMATIC IDENTIFICATION OF MUSICAL INSTRUMENTS

Volodymyr Kuz, Andrii Shatyrko

Taras Shevchenko National University of Kyiv,  Ukraine

University of L’Aquila, Italy

​

�

2 of 13

Introduction

The goal of this study was to investigate whether traditional spectral analysis methods combined with machine learning can achieve reliable classification of musical instruments.��We expected that even with simple models, careful feature extraction would lead to high accuracy in recognizing solo instruments — which is important for music classification, retrieval, and automatic annotation.

Problem Statement:�Given a musical piece, it is required to develop a program that can classify the musical instruments present in it.

Dynamical System Modeling and Stability Investigation, DSMSI-2025

3 of 13

Dataset and Preprocessing

  •  

4 of 13

5 of 13

6 of 13

Models and Training

After extracting all the spectral features necessary for training the model, we face the final—and perhaps the most interesting and challenging—task: applying a machine learning algorithm to train the model using a dataset of musical pieces.

7 of 13

Approximate Nearest Neighbor (ANN)

�The approximate nearest neighbor algorithm returns data points whose distance from the query point does not exceed more than a factor of C compared to the distance to the exact nearest neighbors.�The strength of this approach lies in the fact that, in many practical cases, an approximate nearest neighbor is almost as effective as an exact one.�In particular, when the distance metric accurately reflects the user's notion of similarity or quality, small differences in distance become negligible.

8 of 13

9 of 13

  •  

10 of 13

11 of 13

Results and Comparison

​

  • ANN accuracy: 97.5%
  • SVM accuracy: 97%
  • Model choice not critical – both performed well
  • Feature quality was key to high accuracy

12 of 13

Discussion and Future Extension

Based on this foundation, we can now move toward more complex applications.

  • Apply pipeline to the recognition of military vehicles and equipment from battlefield audio.
  • Of course, this introduces many new challenges:�- High levels of noise�- Low-quality or corrupted recordings�- Very limited training datasets
  • Needs: source separation, noise filtering, data augmentation.
  • Potential for defense and military applications.

13 of 13

Thank you for your attention