SR-FD: representation-space Fréchet auxiliary loss for tokenizer-free diffusion-autoregressive speech generation. EMNLP 2026 submission target. - View it on GitHub