A scalable transfer learning workflow for extracting biological and behavioural insights from forest elephant vocalizations

Animal vocalizations encode rich biological information—such as age, sex, behavioural context and emotional state—making bioacoustic analysis a promising non‐invasive method for assessing welfare and population demography. However, traditional bioacoustic approaches, which rely on manually defined acoustic features, are time‐consuming, require specialized expertise and may introduce subjective bias. These constraints reduce the feasibility of analysing increasingly large datasets generated by passive acoustic monitoring (PAM). Transfer learning with Convolutional Neural Networks (CNNs) offers a scalable alternative by enabling automatic acoustic feature extraction without predefined criteria. Here, we applied four pre‐trained CNNs—two general purpose models (VGGish and YAMNet) and two avian bioacoustic models (Perch and BirdNET)—to African forest elephant (Loxodonta cyclotis) recordings. We used a dimensionality reduction algorithm (UMAP) to represent the extracted acoustic features in two dimensions and evaluated these representations across three key tasks: (1) call‐type classification (rumble, roar and trumpet), (2) rumble sub‐type identification and (3) behavioural and demographic analysis. A Random Forest classifier trained on these features achieved near‐perfect accuracy for rumbles, with Perch attaining the highest average accuracy (0.85) across all call types. Clustering the reduced features identified biologically meaningful rumble sub‐types—such as adult female calls linked to logistics—and provided clearer groupings than manual classification. Statistical analyses showed that factors including age and behavioural context significantly influenced call variation (P < 0.001), with additional comparisons revealing clear differences among contexts (e.g. nursing, competition, separation), sexes and multiple age classes. Perch and BirdNET consistently outperformed general purpose models when dealing with complex or ambiguous calls. These findings demonstrate that transfer learning enables scalable, reproducible bioacoustic workflows capable of detecting biologically meaningful acoustic variation. Integrating this approach into PAM pipelines can enhance the non‐invasive assessment of population dynamics, behaviour and welfare in acoustically active species.