Untrained Blind Sound Source Separation
by Unsupervised Synchrony Loop Network
A newly instantiated, completely untrained Synchrony Loop Network listens to this (single audio channel) musical duet.
The network learns each instrument's sound profile in real time without supervision. Within 1 second of audio onset, it can simultaneously track and precisely extract each instrument.
The separated audio outputs are below.
Another newly instantiated, completely untrained Synchrony Loop Network listens to a saxophone in a crowded auditory scene.
Within 0.1 seconds of audio onset, the network identifies the saxophone as a distinct sound source within the scene and begins to follow and extract it.
Within 1 second it learns its sound profile precisely enough to separate the saxophone from all other sounds.