In the process of working on models that decompose audio into sparse representations, I've found that it can be fun to explore the space of possible musical events that the learned decoders can produce. Most modern "neural" codecs encode fixed-size "frames" of audio. This work models sound as a sparse set of events and physical resonances, which could yield benefits for compression, intrepetability, classification, and beyond.Here, I map 32-dimensional event vectors onto a 2D plane via t-SNE. Each event can be played by clicking or tapping.