sedak326/ML-for-vocal-individuality

Using machine learning to determine whether individual white-naped mangabeys (monkeys) can be identified by their vocalizations. Done in collaboration with WAPCA in Ghana as part of a conservation effort for this endangered species. Published in African Primates, 2025. # Vocal Individuality in White-naped Mangabey Males **Published:** African Primates 19(1): 1-8, 2025 **Authors:** Seda Kavlak, Mireia M. Martin, Foster Poasangma, Mª Teresa Abelló, Andrea Dempsey, Núria Badiella-Giménez **Institutions:** West African Primate Conservation Action (WAPCA), Ghana; Universitat Oberta de Catalunya, Spain; Parc Zoològic de Barcelona, Spain Full paper included as `APVol191Kavlaketal.pdf`. --- This project marks my pivot from animal biology to machine learning. I was presented with a real conservation problem, needed to identify individual animals by their vocalizations, and solved it by extracting acoustic features and training a linear discriminant classifier. It was the first time I understood what ML actually was and how directly it could be applied to problems that matter. That's what made me want to work in this space. --- ## The Problem The white-naped mangabey (*Cercocebus lunulatus*) is an endangered West African primate. WAPCA breeds individuals in captivity and releases them into the wild to reinforce shrinking wild populations. Once an animal is in the forest, recapturing it to check on it is invasive and stressful. Passive Acoustic Monitoring (PAM) is the non-invasive alternative, placing audio recorders in the habitat to detect vocalizations. But species-level detection isn't enough. To know whether a specific released individual is surviving and integrating, you need to identify it by voice. That requires a classifier that can tell individuals apart from their calls alone, and first requires knowing whether the signal even exists. No one had tested this for *C. lunulatus*. --- ## Data and Feature Extraction We recorded the "wahoo" alarm calls of three adult male white-naped mangabeys at Accra Zoo. From each recording we extracted 10 acoustic features using Praat: duration of the "wa" and "hoo" syllables, dominant frequency, fundamental frequency, and the first four formants of the full call. These hand-crafted feat …