In this talk, I will present a hierarchical model designed to analyze the spectrograms of animal vocalizations, with a focus on grunt calls from different lemur species. The primary goal is to uncover a latent spectral shape that characterizes each species and allows us to quantify dissimilarities between them. A key challenge lies in aligning calls of varying durations and temporal dynamics. To tackle this, we incorporate a synchronization function to manage non-stationary temporal features and adopt a circular representation of time to handle artifacts caused by the discretization of analog signals. Given the high dimensionality of spectrogram data, we use a Nearest Neighbor Gaussian Process for efficient computation and sample from the posterior distribution using MCMC. The model is applied to recordings from eight lemur species. For each species, we identify a representative vocal pattern and use a simple distance metric to compare them. Predictive performance is assessed via crossvalidation, and we also explore some special cases that highlight the model’s flexibility.
16 Maggio 2025, ore 12
Gianluca Mastrantonio
Politecnico di Torino, Dipartimento di Scienze Matematiche (DISMA)
In person: Room 24 (4th floor) building CU002 Scienze Statistiche
Webinar: https://uniroma1.zoom.us/j/83625004899?pwd=bXCtz0mp759PUh2lkqT0BUoVa0Uegg.1
ID riunione: 836 2500 4899
Passcode: 123456