Passive acoustic monitoring provides continuous, high-resolution recordings, but real ocean soundscapes are messy: noise is non-stationary and biologically important click events can be sparse. These conditions often challenge purely discriminative detectors. This poster explores Information Field Theory (IFT) as a Bayesian alternative for acoustic denoising. IFT treats signals as latent fields and reconstructs their posterior mean together with calibrated uncertainty.
Methodologically, the spectrogram is modeled as data generated from a latent acoustic field plus noise. A time-periodic, frequency-random prior encodes expected click structure, and variational optimisation of the Gibbs free energy yields the reconstruction. Minimising this free energy corresponds to variational Bayesian inference within IFT.
Implemented in NIFTy.re, the numerical IFT library for Gaussian-process priors and scalable variational inference, the approach suppresses background clutter in sperm-whale recordings and recovers both regular and slow click types without hand-crafted filters.
Outlook: next steps are (1) scaling to automatic click detection in large, sparse archives, and (2) reconstructing fine click morphology (envelope, phase, sub-pulse spacing) to study internal click structure.
Presented at HAICON25, Karlsruhe (03.06.2025) — Schmieder J., Albrecht S., Mousavi H., Fais A.