Acoustic Feature Synergy and Self-Supervised Learning for Robust Tabla Stroke Classification

Accurate and robust automatic classification of tabla strokes is essential for music information retrieval and performance analysis, yet remains challenging due to complex timbral structures and subtle acoustic variability across stroke categories. To address this challenge, we propose a robust tabla stroke classification framework that integrates multidomain handcrafted features, spanning spectral, temporal, cepstral, and perceptual descriptors together with self-supervised learning (SSL) representation derived from a newly developed, manually annotated tabla dataset. This dataset is accompanied by an augmented counterpart that simulates realistic acoustic variability, enabling systematic evaluation under domain shift conditions. ANOVA F-test based feature selection is applied to retain the most discriminative attributes, and a range of machine learning classifiers are employed. Experimental results show that multidomain feature synergy significantly improves classification performance, with the Hybrid-8 configuration achieving up to 97.56% accuracy under in-domain evaluation, while SSL representation exhibits superior cross-domain robustness, attaining 94.07% accuracy when trained on original data and tested on augmented data. While handcrafted multidomain features yield near-ceiling accuracy in controlled settings, SSL representation provides stronger resilience to acoustic variability. These findings reveal a trade-off between peak discriminative performance and cross-domain generalization, highlighting the complementary strengths of handcrafted features and SSL representation for developing robust and generalizable tabla stroke classification system.