DSFlow: Dual Supervision and Step-Aware Architecture for One-Step Flow Matching Speech Synthesis

Researchers have introduced DSFlow, a new distillation framework for flow-matching text-to-speech models. It enables high-quality speech generation in just one step, drastically cutting computational costs compared to traditional iterative methods. The method uses a dual supervision strategy and step-aware tokens to improve stability and parameter efficiency. Experiments show it outperforms standard distillation, achieving strong synthesis quality while reducing both model size and inference cost significantly.