From Pretrain to Primate: Decoding Chimp Nocturnes for Wellbeing

Sanctuaries play a crucial role in ensuring primate welfare, yet limited economic and human resources constrain animal behavior from being systematically monitored. While daytime behavior can be observed directly, nocturnal behavior (equally critical for wellbeing, social dynamics, and management) remains largely inaccessible. Chimpanzees vocalize at night in high arousal contexts, shaping circadian rhythms, emotional states, and subsequent cooperation in management routines. Current approaches to accessing this information, such as manual audio and video inspection, are unsustainable for long-term sanctuary operations. We propose an automated framework to analyze nocturnal vocal behavior using self-supervised speech representation learning. By retraining HuBERT, to which we added a sequential classification layer, we aim to detect and classify vocal interactions from audio recordings. This frugal and scalable approach provides ethologically meaningful insights, enabling informed decisions from keepers without requiring intensive human labor or costly infrastructure. As a case study, we focus on Fundació MONA, a European primate sanctuary where manual night analyses have already informed management decisions. Automation will extend these efforts, offering timely insights that directly enhance chimpanzee wellbeing, and will allow to share the methodology ai to other sanctuaries worldwide.