Automating Annotations for Leopard Vocalizations

Animal vocalizations provide a wealth of information on animals and their surrounding environment.This acoustic data can help us understand the behavioral, physical, and mental state of an animal, which can further help biologists better support the animal's health and well-being.Our project aims to create an automated process using which biologists can identify and annotate Leopard vocalizations from recordings made at animal enclosures.These annotations will be used to study correlations between vocal characteristics and estrus in leopards.Currently, each 24-hour recording takes upwards of 74 minutes to annotate, and each clip has to be manually extracted from the main recording.This is a manual task that requires an individual to carefully go through the recording and verify instances of leopard calls.Our focus is, therefore, not only to reduce this annotation time but also to better allocate resources to more pressing tasks.For this, we use 'auditok', an Audio Activity Detection tool to isolate sounds that are above a certain energy threshold and then store them as individual clips while also retaining timestamp information.Preliminary tests on denoised recordings reveal that most animal vocalizations last about 30 seconds and occur in groups of 15 -20 individual calls.Using 'auditok' and a Python pipeline, we can successfully isolate these individual calls and store them as clips.We also store start and end times for each vocalization in an output file for reference.Currently, we are able to identify all instances of grouped calls and ~80% of individual calls within each group within 30 -50 seconds per file.