Probably the most important aspect to the challenge species identification lies in the building of a robust data set to uniquely define the vocalizations of the individual species. This data representation of the signal must contain enough information to distinctly describe the species vocalization while discarding redundant and irrelevant information to avoid unnecessary processing and resource usage. It was discovered during the course of the investigation that both decreasing the block FFT size and using any sort of overlap between samples did not improve the results of classification by the neural network. This is fortunate since it decreases the size of the data set considerably since only one vector is produced on the entire length of the signal.

The analysis to date suggests that neural networks show great promise in the classification of bird species by their vocalizations. While the networks performed very well for small species subsets, the reliability of the network decreased as the number of species to be classified increased. Increasing the number of neurons in the hidden layers increased the reliability in the network for slightly larger subsets and greatly reduced the number of training epochs required to achieve these results. Adding hidden layers resulted in an increase beyond that which additional neurons could achieve but increased the number of training epochs required. Both of these improvements in the network come at the expense of an increasing consumption of computing resources.

Although it appears that species identification is possible with neural networks, more work must be done with larger, more robust data sets to determine the feasibility of implementing a species recognition system.