Bird species identification via transfer learning from music genres

Humans possess the ability to apply previously acquired knowledge to deal with novel problems quite efficiently. Transfer Learning is inspired by exactly that ability and has been proposed to handle cases where the available data come from diverse feature spaces and/or distributions. This paper proposes to transfer knowledge existing in music genre classification to identify bird species, motivated by the existing acoustic similarities. We propose a Transfer Learning framework exploiting the probability density distributions of ten different music genres for acquiring a degree of affinity between the bird species and each music genre. To this end, we exploit a feature space transformation based on Echo State Networks. The results reveal a consistent average improvement of 11.2% in the identification accuracy of ten European bird species.