Detection and classification of captive coppery titi monkey calls

Ecoacoustic monitoring has many applications in conservation and welfare but generates large amounts of data that are extremely time-intensive to manually process. This has led to an increased interest in the use of machine learning methods to increase efficiency and reduce workload. Common issues within this area include noisy, unbalanced and limited datasets, making it challenging to make effective machine learning models. This study aimed to determine the vocal repertoire of the coppery titi monkey, Plecturocebus cupreus, and develop a machine learning model that can detect, segment and classify calls within streaming audio using a small and unbalanced dataset with overlapping calls from other species. Acoustic data were collected across three zoo populations of P. cupreus using passive acoustic monitors. From this, 3302 calls were manually labelled to use as training data. Ten call types were established manually, corresponding to three groups: short calls, long calls and harsh calls. A Long Short-Term Memory neural network was created that successfully detected calls (accuracy = 0.95) and classified call types (accuracy = 0.97). Potential applications for the model include welfare monitoring in captivity and population monitoring of P. cupreus and related endangered species in the wild.