StuPASE: Low-Hallucination Studio-Quality Speech Enhancement

Researchers have introduced StuPASE, a generative speech enhancement model that achieves studio-level quality while minimizing hallucinations. Built upon the PASE framework, StuPASE addresses two key limitations: it improves dereverberation by finetuning with dry targets instead of those with simulated early reflections, and it replaces the GAN-based module with a flow-matching module to handle strong additive noise. Experiments show StuPASE consistently produces high-quality speech with low hallucination, outperforming state-of-the-art methods. Audio demos are available online.