PHmatch: A prototype-driven semi-supervised framework for insect sound recognition

• A semi-supervised framework for insect sound recognition. • Dynamic Prototype Regularization to enforce a discriminative feature space. • Prototype-Guided Gating for reliable pseudo-label screening. • Pulse-Harmonic Synergistic Attention module to capture critical acoustic structures. The high cost of expert annotation is a challenge in automated insect sound recognition, making semi-supervised learning (SSL) a promising approach. However, mainstream SSL frameworks are not optimized for the fine-grained nature and pulse-harmonic structures of insect spectrograms, which can lead to error propagation. To address these limitations, we propose PHmatch, an SSL framework. It integrates three components: Dynamic Prototype Regularization (DPR) to enhance feature separability, Prototype-Guided Gating (PGG) to improve pseudo-label quality, and a Pulse-Harmonic Synergistic Attention (PHSA) module to capture critical acoustic structures. Experiments on the public InsectSet66 and our InsectAcousticData24 datasets demonstrate that PHmatch achieves more discriminative feature representations and reduces error propagation for fine-grained insect sound classification.