The highest single-test accuracy was produced by running the input through a Fast Fourier Transform prior to feeding it into the neural network; however, separate testing of that one revealed it to be not nearly as accurate, and we suspect there was an error in the input that allowed it to overfit to the sound of a guitar playing.

Feeding the networks raw samples, without the intermediate step of an FFT, yielded higher average accuracy in detection, routinely outperforming the 'random guess' level of accuracy that the FFT tests produced.