CBM-net: Automatic detection and separation of cicada chorus from bee wingbeat audios

Passive acoustic monitoring is an important tool for long-term ecological studies of honey bees. However, audio recordings collected in natural environments are often severely contaminated by high-intensity cicada chorus noise, which substantially degrades the detection of wingbeat sounds and compromises subsequent analyses of ecological indicators. Previous audio separation methods are primarily optimized for speech or bird vocalizations and therefore struggle to cope with the low-frequency harmonic structure of bee sounds (150–500 Hz) as well as the complex acoustic conditions of real-world field environments. To address these challenges, this study proposes a deep learning audio separation framework for bee wingbeat and cicada chorus signals, termed as Complex Bee-Band Mask Network (CBM-Net in short). CBM-Net integrates complex mask separation mechanism with bee-specific frequency band enhancement. This design enables joint optimization of magnitude and phase information in the time-frequency domain. We constructed both a simulated dataset and a real-world mixed audio dataset using bee recordings and cicada sounds collected in natural environments. In the simulated dataset, pure bee wingbeat signals and cicada sounds were mixed using linear overlaying, spatial reverberation, non-linear and time-varying signal-to-noise ratio techniques to emulate realistic acoustic interference. Experimental results on the simulated dataset show that CBM-Net outperformed both traditional approaches and advanced deep learning models in terms of signal fidelity and preservation of bioacoustic characteristics, achieving approximately 9.5% improvement in SDR and 10.6% improvement in SAR. In subjective listening tests on real mixed audios, CBM-Net also received higher scores for sound quality and noise reduction than all other comparison methods. To the best of our knowledge, this is the first systematic study on separating cicada noise from bee audio recordings. The proposed method has strong potential as an effective noise suppression tool for acoustic monitoring of bees and may contribute to future ecological research.