Reproduce “Back to Mel: Learning to Bridge Fidelity and Perception in Generative Speech Enhancement” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Andong Li, Kun Zhou, Xiaodong Li, Chengshi Zheng
BaME combines Mel subbands and degraded-recording phase in a speech-enhancement system. It uses a structured generative backbone and one-step fine-tuning with fidelity and perceptual constraints. On DNS-Challenge 2020, its backbone scores PESQ 2.487, versus 2.448 for a data-prediction DiT, with 9.67M versus 331M parameters. These results support this comparison, not universal superiority or Mel’s optimality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.