Automated detection of wolf howls using audio spectrogram transformers

Abstract

The grey wolf (Canis lupus) is a pivotal species for ecological studies. As a key participant in ecosystem processes, it also serves as a model for investigating social structure formation and ecological adaptation. However, the species' complex social behavior, spatial dynamics, and expansive habitats make monitoring and population assessments across large areas particularly challenging. In recent years, audio traps have been used to collect extensive datasets of wolf vocalizations, particularly howls. Yet, manually detecting wolf howls in lengthy recordings remains a labor-intensive and inefficient task. We propose an approach leveraging modern machine-learning techniques to address this challenge. Following a comprehensive analysis of sound processing methods, we developed two state-of-the-art deep learning models based on the Audio Spectrogram Transformer architecture. The first model classifies audio for the presence of animal sounds with a precision of 98.3% and a recall of 99.3%. The second model distinguishes wolf howls from other animal sounds with a precision of 89.6% and a recall of 93.4%. These models significantly enhance the efficiency and accuracy of wolf vocalization detection, supporting ecological monitoring and research efforts.