Sparse coding for large scale bioacoustic similarity function improved by multiscale scattering

The bioacoustic event indexing requires to be scaled in space (oceans and large forests, multiple sensors), and at inter and intra specie level. The usual time-frequency featuring is inefficient for long term precise correlation analysis. We propose that a sparse transform of the time frequency domain or so, can generate an efficient representation that allows light similarity computations. In this paper we illustrate it with the tracking of minke whales (Balaenoptera acutorostrata) with a sparse coding of their 'boing' vocalizations. This sparse coding confers several advantages : it makes the structure in natural signals explicit and it represents complex data in a way that is easier to read and compute at subsequent levels of processing. We discuss on the required properties of this sparsed representation, first based on usual Mel Filter Cepstral Coeff. We then demonstrate that the recent scalogram for audio representation will produce more contrastive cosine similarity measure for any species, yieldin...