Neural Implicit Representation for Spatial Audio Coding

This work revolves around the implementation of a neural implicit representation for spatial audio, specifically focusing on synthesizing impulse responses in a continuous space using a sparse dataset, and compressin the network to obtain a lightweight representation of the space. The proposed method, Neural Acoustic Fields (NAFs), models acoustic propagation as a linear time-invariant system and learns to map emitter and listener locations to a neural impulse response function. By optimizing and compressing the representation, NAFs enable the rendering of spatial acoustics at arbitrary locations with also being performant.