<p>This study presents a novel framework for predominant instrument recognition in polyphonic music using Focal Modulation Networks (FMN). Trained on single-instrument data, FMN accurately identifies multiple instruments in unseen test data without sliding window analysis. It replaces self-attention with a focal modulation module, where focal context aggregation prioritizes key spectral regions, and adaptive query modulation refines feature representations. Extensive experiments demonstrate that FMN achieves a macro F1-score of 0.58 and a micro F1-score of 0.66, representing a 10 and 16% improvement over the state-of-the-art Han model, respectively. FMN consistently outperforms various transformer architectures, including Vision Transformers (ViT), Swin Transformers (SwinT), and Compact Convolutional Transformers (CCT), with enhanced accuracy and interpretability. Our findings highlight FMN’s potential in capturing complex musical patterns and relationships, establishing a new benchmark for predominant instrument recognition in polyphonic music, without the need for aggregation strategies or sliding window approaches</p>