Repurposing Bird‐Based Acoustic Models for Amphibian Surveys: Strengths and Limitations in Real‐World Deployments

ABSTRACT Traditional biodiversity monitoring is often hindered by hazardous conditions and restricted human access. To bridge critical data gaps in these areas, we developed AnuraSearch (implemented as a frozen BirdNET v2.4 feature extractor coupled with a newly trained amphibian‐specific classification head). The 11 included species represent the majority of acoustically active Ukrainian anurans with available open‐access recordings in Xeno‐canto and Observation.org . The dataset comprised 15,213 3‐s audio clips partitioned into training (80%; 12,170 clips) and validation (20%; 3043 clips) subsets. We evaluated the model via a two‐stage process. On an external test set of 4389 manually annotated 3‐s clips from independent high‐quality, predominantly single‐species recordings from open‐access archives—Xeno‐canto and Observation.org—(Stage 1), AnuraSearch achieved macro‐averaged Precision = 0.994 and Recall = 0.62 (weighted Recall = 0.73). On a second external test set of 8686 expert‐verified field clips from 22 autonomous recording units deployed across diverse amphibian habitats within the Strict Nature Reserve Roztochya (Stage 2), macro‐Precision remained high (0.69) but macro‐Recall fell to 0.23, revealing species‐specific domain shift. Performance particularly declined for species with low‐amplitude calls or underwater vocalisations, highlighting the physical limitations of airborne monitoring for cryptic taxa. We applied the model to over 18,000 h of scheduled audio recordings (2 h around sunrise + 2 h around sunset + 5‐min clips every 30 min over the remaining 24‐h cycle), successfully reconstructing phenological and diel activity patterns without constant human presence. For acoustically cryptic and domain‐shifted species, AnuraSearch functions as a high‐precision clip‐prioritisation filter rather than a tool for reliable absence inference; reported confidence scores are not yet calibrated measures of detection reliability. Our results demonstrate that repurposing avian bioacoustic models via transfer learning is a viable, resource‐efficient strategy for maintaining ecological oversight.