Instrument identification remains an interesting problem, and one that we feel should be given further consideration. Increasing the size of the windows that the network was given yielded slight increases in accuracy, but we were unable to explore the extent to which this continues. Overall, we recommend further research with the aim of identifying which factors have the largest effect on accuracy.
Software like this, once fully-functional, has several useful applications. The obvious is for music information retrieval purposes, where it could be quite useful for aiding researchers in annotating non-notated music.
Looking away from MIR techniques, the same sort of algorithm could be combined with speech-recognition software to automatically annotate panel discussions or podcasts, making them much more accessible for hearing-impaired users, as well as more easily searchable.