Download SCAPES: Semantically Conditioned Autoregressive Prior for Environmental Sounds This paper presents SCAPES, a semantically conditioned autoregressive prior for environmental sound generation. The system models discrete audio representations using an autoregressive architecture conditioned on semantic information, enabling the generation of environmental sounds that follow user-specified concepts. By learning a prior over audio tokens, SCAPES combines high-level semantic control with detailed temporal modeling. Experimental evaluation investigates the quality, diversity, and semantic consistency of generated sounds, demonstrating the potential of autoregressive priors for controllable environmental sound synthesis.