UC San Diego, Stability AI, Arm · Text-to-Speech

ARC Text-to-Audio Generation

Introduces Adversarial Relativistic-Contrastive (ARC) post-training, a novel acceleration algorithm for diffusion/flow models that generates 12 seconds of 44.1kHz stereo audio in ~75ms on H100 GPUs and ~7 seconds on mobile edge devices. Combines relativistic adversarial formulation with contrastive discriminator objective to improve prompt adherence without distillation.