A simulation based approach for enhancing automatic detection of cicada songs in challenging chorus conditions

Advancements in machine learning-based audio recognition have significantly contributed to ecoacoustics. However, past studies have primarily focused on bird vocalizations, leaving insect sounds under-studied. One challenge in this area is recognizing insect calls in chorus conditions where multiple species vocalize simultaneously at high population densities. This complexity requires labeled training data specific to each chorus pattern, making sound recognition challenging and resource-intensive. To address this problem, we propose a data augmentation approach using physical acoustic simulation. Our method, tested on five species of Japanese cicadas, utilizes simulation software to generate numerous pseudo-chorus recordings by combining focal recordings of individual cicada vocalizations with background sounds. Thousands of simulations with various species combinations and individual counts were used to train a convolutional neural network (CNN) multi-label classifier. The proposed method achieved a 22% improvement in F1-score compared to models trained on focal recordings alone. By optimizing simulation parameters through hyperparameter tuning, we achieved a mean F1-score of 0.83 on real-world soundscape recordings with minimal training data (approximately 10 focal recordings per species). This approach is readily applicable to other chorusing taxa such as orthopterans and amphibians, offering a scalable solution for biodiversity monitoring in complex acoustic environments. • Physical acoustic simulation generates realistic training data for complex cicada choruses. • Simulation-trained models outperform models trained on focal recordings by 22% (in F1-score). • Only around 10 focal recordings per species needed to achieve 0.83 mean F1-score. • Approach reduces annotation burden and is adaptable to various chorusing taxa.