Reproduce “ZipEnhancer: Dual-Path Down-Up Sampling-based Zipformer for Monaural 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
Haoxu Wang, Biao Tian
ZipEnhancer is a speech enhancement model that uses dual-path processing with efficient time and frequency domain down-up sampling within a modified Zipformer architecture. It achieves state-of-the-art PESQ scores of 3.69 on DNS2020 and 3.63 on Voicebank+DEMAND datasets, while maintaining a compact parameter count of 2.04M and 62.41G FLOPS.
Key insights
- Applying learnable down-up sampling in both time and frequency domains effectively reduces computational complexity while preserving multi-scale speech information.
- The dual-path processing framework is crucial for capturing both temporal and spectral characteristics, as removing either path significantly degrades performance.
- Leveraging a modified Zipformer architecture, including an efficient non-linear attention mechanism and reused attention weights, enables robust feature extraction with reduced overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.