Cross‐continental zero‐shot anuran call classification with CLAP and denoising‐based synthetic embedding generation

Using machine learning models to classify bioacoustic signals of animal species is increasingly important for conservation monitoring because manual expert labelling is time‐consuming and tedious. Monitoring endangered species is particularly challenging because these species are rare, making it difficult to collect the large, representative datasets needed to train conventional machine learning models. Zero‐shot learning (ZSL) offers a promising solution by enabling models to recognise new species that are absent from the training data.

We developed a multi‐stage ZSL approach to detect and classify frog calls in natural soundscapes recorded by an automated acoustic monitoring device. Our multi‐label ZSL method uses a denoising model for embedding generation. It is designed for complex real‐world conditions, including substantial background noise, long periods without frog vocal activity, and overlapping multi‐species choruses. These settings are rarely represented in the curated benchmark datasets used in many previous ZSL studies.

To evaluate our approach, we curated and annotated a new dataset of Australian frog calls and conducted a three‐part ablation study. The ZSL model was trained on AnuraSet, a large corpus of Neotropical anuran calls, and therefore represents a zero‐shot transfer of anuran call classification across continents. Using our evaluation set, the model achieves an F1 score of 65.54% for frogs overall and 64.14% for the endangered green and golden bell frog ( Litoria aurea ), outperforming a distribution‐matched random baseline of 25.9% by more than double.

Our findings show that ZSL can support practical acoustic monitoring of rare or previously unrecorded species under real‐world conditions. We present a proof‐of‐concept system that operates in noisy, multi‐label soundscapes and demonstrates cross‐continental transfer of anuran call classification from the Neotropics to eastern Australia. These results highlight the feasibility of applying ZSL to biodiversity monitoring and identify key design considerations for future implementations.