ChiroEcho AI extends bat species identification beyond training data using geography

ChiroEcho is a new deep learning framework designed to expand automated bat vocalisation classification beyond the species it was explicitly trained on. Developed by Burooj Ghani, Dan Stowell, A. Leonie Baier and colleagues, the system predicts both species and genus from echolocation calls, then merges the coarse genus-level output with transparent geographic distribution maps at inference time. When only one species of a predicted genus occurs in a given region, the framework can resolve a species that was completely absent from the training taxonomy.

The team evaluated ChiroEcho using recordings spanning 35 European bat species, looking at closed-set classification accuracy, the instability of performance estimates for rare species, and a controlled held-out proof-of-principle experiment. Their rare-species analysis reveals how limited evaluation data can mask real species-level performance—a critical insight for conservation-driven monitoring. The held-out test showed that combining genus predictions with location data successfully recovers labels that the species classification head alone could not provide.

In practice, the geographic resolution extends operational coverage from 35 to 41 of the 48 native European bat species, raising coverage from 73% to 85%—the broadest automated classification coverage reported for European bats to date. The researchers frame this as a general proof of principle: coarse model predictions combined with external, interpretable constraints can resolve unseen fine-grained classes. That principle could apply beyond bioacoustics, offering a path for AI systems to handle novel categories in fields like ecology, medical imaging, or industrial quality control where training data can never capture every possible variant.