Hidden Markov models for detection of Mysticetes vocalisations based on principal component analysis

Abstract

The economic relevance of Mysticetes has prompted marine ecologists and biologists to investigate this suborder of cetaceans. Mysticetes produce distinct vocal repertoires, which are recorded to analyse the behaviour of the species within its ecology. Passive acoustic monitoring (PAM) is a standard technique for tracking Mysticete movement and vocalisation. PAM collects enormous datasets over a long period, making it practically impossible to analyse with typical visual examination methods. Machine learning (ML) techniques such as hidden Markov models (HMMs) have made automatic recognition and analysis of extensive sound recordings possible. Nevertheless, the performance of ML tools is determined by the adopted feature extraction technique. Hence, this article introduces the method of principal component analysis (PCA) as a performance-efficient alternative feature extraction technique for detecting Mysticete vocalisations using HMM. Performance of the developed PCA-HMM detector is compared with state-of-the-art detectors using two different Mysticete vocalisations (Humpback whale songs and Bryde’s whale short pulses). In both species, results show that the PCA-HMM detector has the best performance and is more suitable for use in real-time application since it exhibits less computational time complexity.