Reproduce “SNR-Based Teachers-Student Technique for 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
Xiang Hao, Xiangdong Su, Zhiyu Wang, Qiang Zhang, Huali Xu, Guanglai Gao
A speech enhancement method leverages an SNR-based teachers-student technique to achieve robust performance across a wide range of signal-to-noise ratios, outperforming existing state-of-the-art baselines with a PESQ of 2.304 and STOI of 0.744 on a challenging test set. The approach effectively generalizes to unseen noise conditions and SNRs.
Key insights
- Specialized teacher models, trained on narrow SNR ranges, can become highly proficient within their designated domains.
- Knowledge distillation can effectively transfer this conditional expertise from multiple specialized teachers to a single student model, enabling robust performance across a wide SNR range.
- A time-domain U-Net is effective for direct waveform-to-waveform speech enhancement, which is crucial for preserving phase information, especially at low SNRs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.