Musical instrument classification, as a fundamental task in music information retrieval (MIR), has broad applications in music analysis, education, and content management. However, existing research primarily focuses on short monophonic samples for classification, which fails to capture the timbral variation characteristics in real performance scenarios. Meanwhile, traditional deep learning models still have limitations in extracting complex timbral features. To address these challenges, this paper proposes ICKAN, a deep instrument classification model that incorporates the Kolmogorov-Arnold Network (KAN), and constructs a large-scale dataset containing 30,824 complete musical phrases. Experimental results demonstrate that ICKAN achieves a classification accuracy of 95.74% in a 20-class instrument classification task with 10-second audio segments, significantly outperforming current methods. This research introduces learnable nonlinear activation functions and comprehensive musical segments, offering new insights into improving the accuracy and practicality of instrument classification and contributing valuable references for the advancement of music information retrieval technology. The code and dataset are available at https://github.com/NMLAB8/ICKAN .
Keywords: Audio feature extraction; Deep learning; Kolmogorov-Arnold network; Music information retrieval; Musical instrument classification.
© 2025. The Author(s).