Frequency tracking using Monte Carlo methods: Application to bat echolocation signals

This paper presents a sequential Monte Carlo method for tracking an unknown varying number of time-varying frequencies. A reversible-jump sampler is used to implement model-order determination. It is shown that for a linear-in-the-amplitudes observation model in white Gaussian noise, the amplitudes and noise variance can be analytically marginalised out of the posterior distribution resulting in a reduced dimension state estimation problem. A sum of linear chirps model is chosen as a local observation model and this basis is used to determine the instantaneous frequency of the signal. We present frequency tracking results from synthetic as well as field-recorded bat echolocation calls and compare results with Fourier based frequency tracking.