COMPSTAT 2022: Start Registration
View Submission - COMPSTAT2022
Title: A nearest neighbours Gaussian Process model for time-frequency data: An application in bio-acoustic analysis Authors:  Hiu Ching Yip - Politecnico Di Torino (Italy) [presenting]
Gianluca Mastrantonio - Politecnico of Turin (Italy)
Enrico Bibbona - Politecnico di Torino (Italy)
Marco Gamba - University of Turin (Italy)
Daria Valente - University of Turin (Italy)
Abstract: In comparative bio-acoustic studies, one area of interest is to understand the acoustic structures of non-human primates in order to provide insights into the evolution of the communication mechanism of our closest relatives. The most common practices are feature engineering methods, which involve selecting a set of basis-features for quantitative comparison. The identification of meaningful features in the vocal repertoire relies on biologists to observe and interpret the behavioural contexts in which the animals emit the signals. These interpretations are costly to acquire, inaccurate due to human subjectivity and difficult to generalize for cross-species comparison. Furthermore, feature selection always treats the time-frequency bins of spectrograms as independent features or extracts common statistics from waveforms. This ignores the time-varying effect of observed vocalizations on the latent acoustic structure as well as the periodic nature of time-frequency data. The aim is to propose a Nearest Neighbour Gaussian Process model that accounts for the time varying components as well as the circular nature of time-frequency for latent spectral structure inference from bio-acoustic data. The dataset that will be available for model implementation are vocal signals of lemurs that were recorded in Madagascar.