While working on some deep learning related audio projects I came upon an interesting
auditory illusion that shows how our brain associates memorized phrases with auditory stimuli. Below is a sound generated by randomly stimulating a neural net that was pretrained on approximately 1000 recordings of brass sounds (both single notes and entire musical phrases). The resulting “drone” is totally uniform and has no temporal development at all (in the sense that there would be individual notes present in that mixture). If you first listen to the audio file it just sounds chaotic and maybe a bit disharmonic, nothing to see here, move along. But wait! Try running the file in the background for several minutes in a loop and you will notice that it appears as if there are suddenly notes, even entire melodies emerging from the mix. If you listen long enough this sound will trigger memories of music that you have listened to in the past, and it will appear as if you can actually make out melodies (even though they are at the brink of just being noticeable in the noise). They can be subtle like a car horn in the distance, church bells, or an ambulance, or (like in my case) an entire phrase played by a jazz trumpet. Try it out, it’s quite fun!