Dual-correlation graph attention network with supervision augmentation for polyphonic instrument recognition

Although neural networks have achieved remarkable progress in the field of musical instrument recognition, the collaborative interaction patterns within musical ensembles have not been fully explored. Musical instrument recognition differs from other time-series analysis tasks in that the expression of consistent musical themes typically relies on analogous instrumental combinations and arrangement techniques, with the latter often reflected in rhythmic characteristics. This necessitates effective temporal continuity modeling on sequential data, which is closely linked to diverse instrumental combinations. Such combinations further exhibit semantic coherence centered around musical themes. However, existing research rarely focuses on the joint modeling of these factors, making it difficult for conventional neural network-based time-series analysis methods to achieve satisfactory performance when directly applied to instrument recognition tasks. To address this issue, this paper proposes a Dual-Correlation Graph Attention Network with Supervision Augmentation (DC-GAT) for musical instrument recognition. The core innovation lies in a unified framework that adapts to the characteristics of instrument recognition by integrating temporal continuity modeling and semantic reasoning based on label correlation. Specifically: (1) Temporal continuity modeling is implemented via dynamic temporal correlation graphs, which are constructed using windowed Mahalanobis similarity combined with k-NN sparsification. (2) Semantic reasoning is supported by label correlation graphs built on CLIP-based embeddings, where instrument co-occurrence patterns are captured under the guidance of genre-aware auxiliary supervision. These two types of graphs are fused through a graph attention mechanism, enabling joint optimization of temporal perception and modeling of instrumental collaborative relationships. Experimental results on the OpenMIC and OpenMIC-IRMAS datasets demonstrate that the proposed method achieves state-of-the-art performance with statistically significant improvements over baseline approaches.