Latent diffusion text-to-audio model — generate music, sound effects, and speech from natural-language prompts. Free open weights and inference code.