A Self-Supervised Approach for Hainan Gibbon Call Separation in Passive Acoustic Monitoring

Passive acoustic monitoring (PAM) provides an effective approach for ecological research and endangered species conservation. However, automated analysis of PAM recordings from the critically endangered Hainan gibbon (Nomascus hainanus) remains challenging due to environmental noise and the lack of paired clean reference recordings. Supervised source separation methods often require extensive annotated datasets, limiting their applicability to real-world ecological recordings. To address these challenges, this study proposes MC-RemixIT, a self-supervised source separation approach designed for Hainan gibbon calls under complex acoustic conditions. Building on RemixIT, MC-RemixIT introduces a lightweight call enhancement module with attention mechanisms, signal cross-correlation constraints to improve target-noise separation, and multiple training constraints to enhance separation robustness. The proposed approach operates directly on raw waveforms and does not require fully clean reference signals during training. Experimental results demonstrate that MC-RemixIT consistently improves separation performance compared with RemixIT and bioacoustic denoising frameworks. Moreover, downstream call presence detection experiments show that target-call estimates generated by MC-RemixIT improve detection performance across multiple deep learning architectures, indicating enhanced usability of the separated signals for automated acoustic analysis. MC-RemixIT provides a practical self-supervised solution for improving wildlife acoustic recordings in noisy environments and supports large-scale bioacoustic monitoring of endangered species.