Prism: Neural Modeling of Multiband Mixtures of Nonlinear Analog Effects

This paper introduces Prism, a neural audio modeling framework for real-time emulation of multiband mixtures of nonlinear analog effects, including overdrive, distortion, and fuzz pedals. While conventional approaches involve models specifically trained to emulate the response of single devices on the whole frequency spectrum, Prism exploits band-aware conditioning, allowing independent end-to-end processing of multiple frequency bands directly on the input waveform. The model leverages latent representations to modulate a processing network, supporting simultaneous real-time emulation of multiple effect types and complex interband interactions. Prism is evaluated through benchmarks on audio quality and computational efficiency, yielding performance comparable with state-of-the-art virtual analog models despite the broader scope of the task. Prism allows extended control on hybrid and creative audio manipulation beyond conventional virtual analog modeling.