Species-specific audio detection: A comparison of three template-based classification algorithms using random forests

Abstract

We developed a web-based cloud-hosted system that allow users to archive, listen, visualize, and annotate recordings. The system also provides tools to convert these annotations into datasets that can be used to train a computer to detect the presence or absence of a species. The algorithm used by the system was selected after comparing the accuracy and efficiency of three variants of a template-based classification. The algorithm We developed a web-based cloud-hosted system that allow users to archive, listen, visualize, and annotate recordings. The system also provides tools to convert these annotations into datasets that can be used to train a computer to detect the presence or absence of a species. The algorithm used by the system was selected after comparing the accuracy and efficiency of three variants of a template-based classification. The algorithm computes a similarity vector by comparing a template of a species call with time increments across the spectrogram. Statistical features are extracted from this vector and used as input for a Random Forest classifier that predicts presence or absence of the species in the recording. The fastest algorithm variant had the highest average accuracy and specificity; therefore, it was implemented in the ARBIMON web-based system.