Reproduce “SpeechRefiner: Towards Perceptual Quality Refinement for Front-End Algorithms” 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
Sirui Li, Shuai Wang, Zhijun Liu, Zhongjie Jiang, Yannan Wang, Haizhou Li
SpeechRefiner introduces a universal, standalone post-processing module leveraging Conditional Flow Matching to enhance the perceptual quality of speech outputs from various front-end algorithms. The system demonstrates superior or competitive perceptual quality (measured by SIGMOS) compared to task-specific refiners and exhibits strong generalization to unseen front-end algorithms, improving overall quality scores by up to 42% on unseen systems.
Key insights
- Conditional Flow Matching is effective for learning complex mappings from degraded to perceptually clean speech, enabling high-quality refinement.
- A single, universally trained post-processing model can effectively improve perceptual quality across diverse speech processing tasks (enhancement, dereverberation, separation, target speaker extraction).
- Prioritizing perceptual metrics like SIGMOS in model development can bridge the gap between objective performance and human listening experience, leading to more natural and clearer speech outputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.