WAND: Windowed Attention and Knowledge Distillation for Efficient Autoregressive Text-to-Speech Models
This research introduces WAND, a novel framework that optimizes autoregressive text-to-speech models by replacing full self-attention with a hybrid global-lo...