Linear discriminant analysis based hidden Markov model for detection of Mysticetes’ vocalisations

Mysticetes' produce distinctive vocalisations which are used for echolocation, communication, and other marine functions. These cryptic vocalisations are studied by marine scientist to determine the behavioural patterns and movement of this suborder of cetaceans within their ecosystem. In practice, these vocalisations are gathered using passive acoustic monitoring over days, weeks, months, and even years. Therefore, it is complex to study these sounds using traditional visual inspection techniques because the gathered datasets are huge. Machine learning (ML) tools such as Gaussian mixture models (GMMs), support vector machines (SVMs), and hidden Markov models (HMMs) have been adopted in recent times to proffer analytic solutions to automatically detect and study these cryptic vocalisations. Notwithstanding, the feature extraction techniques employed play a vital role in determining the performance of these ML tools. In most cases, the performance of the feature extraction technique is directly proportional to the performance of the ML tools. Thus, the method of linear discriminant analysis (LDA) is introduced in this article as a feature extraction technique that can be adapted with the HMMs (LDA-HMM) to seamlessly detect the vocalisations of Mysticetes. The performance of the proposed LDA-HMM detector is compared with other recent detectors for Mysticetes' vocalisations in the literature using two different species: Humpback whale songs and Bryde's whale pulses. Experimental results show that the developed LDA-HMM detector is a performance-efficient alternative in comparison to the recent detection techniques studied in this article. Besides, the LDA-HMM detector offers less computational time complexity; as such, it is more suitable for real-time applications.