Reproduce “Improving Noisy Student Training on Non-target Domain Data for Automatic Speech Recognition” 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
Yu Chen, Wen Ding, Junjie Lai
The LM Filter strategy improves Noisy Student Training (NST) for Automatic Speech Recognition (ASR) by effectively selecting valuable pseudo-labeled data from noisy, non-target domain unsupervised corpora. This approach leads to enhanced ASR performance and faster convergence compared to standard NST without filtering, especially in challenging real-world data conditions.
Key insights
- CER-Hypo serves as a reliable unsupervised proxy for pseudo-label quality, showing a strong positive correlation with the true Character Error Rate (CER-Label).
- Iteratively relaxing the data selection threshold allows the ASR model to progressively learn from more diverse and potentially noisier data as its accuracy improves.
- The quality of pseudo-labels generated by the teacher model is highly sensitive to the domain similarity between the supervised and unsupervised data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.