Probabilistic bioacoustic signal extraction within spectrograms.

Abstract

A methodology for probabilistic bioacoustic signal extraction within spectrograms of natural sound recordings is proposed. Probabilistic models of signal attributes are described for generating a host of likelihoods used to estimate the probability that individual pixels within a spectrogram represent part of a bioacoustic signal. The pixel probabilities result in a transformation of the spectrogram into a probability map of bioacoustic signal presence. It is shown that these probability maps create a dramatic increase in the bioacoustic signal-to-noise ratio within the spectrogram. These probability maps along with threshold filtering provide a means for image segmentation of the spectrogram, creating blocks of pixels that represent bioacoustic signals, facilitating feature and signal extraction. This methodology is applied to natural sound recordings of three quality types in a wide range of signal-to-noise ratios. In each instance, the probability mapping greatly increases the signal-to-noise ratio and, when applied as a threshold filter, is far more selective with pixel inclusion than threshold filtering applied based on sound level. Suggested applications include automated call recognition of birds, frogs, and insects from field recordings within a wide range of ambient noise.