Birdsong recognition with DSP and neural networks

Analysis of speech often begins with study of the vocal tract that created it. Bird vocalizations and human speech are generated by similar processes. This suggests that LPC coefficients extracted from birdsong samples could retain enough information to permit identification of species. In this paper we train a back-propagation neural network to recognize bird songs. We generated test and training data sets using 133 songs from six common bird species. Initially, identification performance was good for some species, and poor for others. We attributed this to a lack of temporal context information in the data. By changing the type of spectral information presented to the network, we were able to improve performance. We conclude that a neural network combined with digital preprocessing can be used to identify a bird by its song.