It has recently been shown that certain types of modern Neural Nets for sound classification have parallels to the human auditory system. If we train such a network to recognize musical genres, what does it actually hear? How does music sound to an artificial ear? We can try to get an idea by letting the network ‚dream up‘ new audio clips that it thinks fit well into a given genre. The network however has learned to define genres via very specific patterns that we humans might not recognize. So the generated segments may not even sound like music to us. But they allow us a eavesdrop on the often tightly sealed neural black box.

For all of the following musical genres the network generated multiple examples: pop and chart / pop rock / folk rock / punk, punk rock / indie, indie pop, indie rock / folk / new wave / industrial / progressive rock, psychedelic rock / gothic rock / post-rock, shoegaze, shoegazing / blues rock, hard rock / synthpop / techno, trance, house / dance and electronica, electronica / dubstep, idm / chillout, downtempo / ambient / new age / hip hop rnb and dance hall, hip hop, hip-hop, hiphop, rap / metal, heavy metal, trash emtal, black metal, death metal, alternative metal / jazz, bebop / blues / classical / country, country rock / chinese, cantonese, mandarin / desi, punjabi, bhangra / arabic / latin, mexican, mexico / francophone / world / ska / reggae, dub / soul, rnb / funk / disco / gospel / easy listening  / 80s / big band, swing


Audio file of the above mentioned music genres as ‘dreamed’ by the neural networks.


Audio spectrograms