A hidden Markov model with selective time domain feature extraction to detect inshore Bryde's whale short pulse calls

摘要

An Increase in the study of cetaceans' sounds has motivated the development of different automated sound detection and classification techniques. Passive acoustic monitoring (PAM) is extensively used to study these cetaceans' sounds over a period to understand their daily activities within their ecosystem. Using PAM, the gathered sound datasets are usually large and impractical to manually analyse and detect. Thus, hidden Markov models (HMM) is one of the popular tools used to automatically detect and classify these cetaceans' sounds. Nonetheless, HMM rely heavily on the employed feature extraction method such as Mel-scale frequency cepstral coefficients (MFCC) and linear predictive coding (LPC). In most cases, the more reliable the extracted feature vector from the known sound label, the higher the sensitivity of the HMM. Although these aforementioned feature extraction methods are widely used, their design is based on filters and requires windowing, fast Fourier transforms (FFT), and logarithm operations. Consequently, this increases the computational time complexity of the HMM. Here, we describe a selective time domain feature extraction method that can be easily adapted with the HMM. This proposed feature extraction method uses a combination of some simple but robust parameters such as the mean, relative amplitude and relative power/energy (MAP), which are selected based on empirical observations of the call to be detected. The performance of this proposed MAP-HMM was verified using the acoustic dataset of continuous recordings of an inshore Bryde's whale (Balaenoptera) short pulse calls collected in a single site in False bay, South-West of South Africa. Aside from exhibiting a low computational complexity, the proposed MAP-HMM offers superior sensitivity and false discovery rate performances in comparison to the LPCHMM and MFCC-HMM.