Reproduce “Voice activity detection in the wild: A data-driven approach using teacher-student training” 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
Heinrich Dinkel, Shuai Wang, Xuenan Xu, Mengyue Wu, Kai Yu
A data-driven teacher-student framework improves Voice Activity Detection by leveraging large, weakly-labeled, unconstrained audio datasets, achieving superior noise robustness and generalization in real-world conditions while maintaining performance on cleaner speech, thus reducing reliance on costly frame-level annotations.
Key insights
- Teacher models trained on diverse, weakly-labeled real-world audio datasets exhibit inherent robustness to various noise conditions, even with clip-level supervision.
- Frame-level soft labels generated by a weakly-supervised teacher effectively guide a student model, leading to performance improvements over clip-level supervision across different noise levels.
- Training student models on large, unlabeled, out-of-domain speech datasets using teacher-generated labels facilitates cross-domain adaptation and enhances real-world VAD performance, with dataset characteristics sometimes being more influential than sheer volume.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.