Multispecies Discrimination of Seals (Pinnipeds) using Hidden Markov Models (HMMs)

Abstract

Hidden Markov Models (HMMs) were developed and implemented for the discrimination of 5 available Seals (Pinnipeds), namely the Bearded Seal (Erignathus barbatus), Harp Seal (Pagophilus groenlandicus), Leopard Seal (Hydrurga leptonyx), Ross Seal (Ommatophoca rossii), and Weddell Seal (Leptonychotes weddellii). The main objectives of the experiments were to study the impact of the frame size and step size and number of states for feature extraction and acoustic models on classification accuracy. Based on the experiments using Mel-Frequency Cepstral Coefficients (MFCCs) extracted from the vocalizations (15 ms frame size and 4 ms step size), HMMs containing 20 states with single underlying Gaussian Mixture Model (GMM) produced discrimination of 95.77%. From the results, the framework could be applied to analysis for other marine mammals for both classification and detection of vocalizations and species.