An Explainable and Lightweight Music Genre Classification Framework for Resource-Constrained Settings
Year: 2026
Type: conference-paper
Abstract: This paper proposes a lightweight music classification framework that combines handcrafted acoustic features with traditional machine learning models. Here, acoustic descriptors (RMS, spectral features, MFCC, chroma, and tempo) are initially derived using the audio data, and sample-level feature representations are built by statistical aggregation. Then, the three baseline models, SVM, Random Forest, and XGBoost are compared in a single feature space, and additional optimization is carried out on the most successful Random Forest model. To improve the performance of the model, this paper uses Optuna and Stratified K-Fold cross-validation to tune the hyperparameters of the Random Forest model and SHAP model to explain the decision-making process and the role of various acoustic features in improving classification. Besides the main experiment, this paper does extended experiments using the FMA small dataset to assess the transferability and the generalization ability of the method to a second dataset. The experimental findings illustrate that the tuned Random Forest beats all the baseline models on the GTZAN dataset, with the final model on FMA_small also producing better results than the baseline Random Forest. In general, the approach suggested in this paper is practically valuable regarding the lightweight character, interpretability, and reproducibility of music classification models.
Author: Kaixin Li
Institution: Sichuan University
Language: English
Topic: Music and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203708151