Overall outcome analysis and conclusions

An overall analysis of our program's output suggests a quite disappointing result: notwithstanding that some small gains can be achieved by different classifiers and a careful hyper-parameter tuning, any major performance improvement arises from the feature extraction algorithm, with the Mel Frequency Cepstral Coefficients approach getting consistently much more promising results than the Fast Fourier Transform.
As a consequence of our finding, we think that a future extension of our work would concentrate mostly on leveraging a smarter feature extraction process, such as the one relying on psychoacoustic features or the Auditory filterbank temporal envelopes: the former is based on estimates of the perceived roughness, loudness and sharpness, whereas the latter is based on a model representation of temporal envelope processing by the human auditory system.