From Birdsong to Rumbles: Classifying Elephant Calls with Out-of-Species Embeddings

A new paper on arXiv (2605.00225) from researchers Christiaan M. Geldenhuys and Thomas R. Niesler demonstrates a practical approach to classifying elephant vocalizations using pretrained acoustic embeddings without any fine-tuning. The team evaluated a range of models from general audio, speech, and bioacoustic domains—all out-of-species (no elephant data used during pretraining). Perch 2.0, originally trained on bird sounds, performed best, achieving AUCs of 0.849 on African bush elephant (Loxodonta africana) calls and 0.936 on Asian elephant (Elephas maximus) calls. This is within 2.2% of a fully end-to-end supervised neural network, a remarkable result given the scarcity of annotated bioacoustic data. The fixed embedding networks serve as feature extractors; only lightweight downstream classifiers (linear models or small neural nets) are trained, making the system computationally efficient.

Perhaps most interestingly, the researchers conducted a layerwise analysis of transformer encoders like wav2vec2.0 and HuBERT. They found that the second layer encodes sufficient information for effective elephant call classification, while later layers add little improvement. By truncating the network at this layer, they retained only about 10% of the original parameters without significant performance loss. This compact representation is ideal for on-device processing in resource-constrained environments, such as field recorders or drones. The study highlights the potential of transfer learning from out-of-species audio models for bioacoustic monitoring, reducing the need for expensive labeled datasets. For conservationists, this could enable real-time elephant call classification in the wild, aiding in population tracking and human-elephant conflict mitigation.