Reproduce “CTC-DRO: Robust Optimization for Reducing Language Disparities in 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
Martijn Bartelds, Ananjan Nandi, Moussa Koulako Bala Doumbouya, Dan Jurafsky, Tatsunori Hashimoto, Karen Livescu
CTC-DRO adapts group distributionally robust optimization for multilingual speech recognition, where CTC losses can be misleadingly high for reasons such as longer audio. By matching batch durations across language groups and smoothing group-weight updates, it lowers worst-language character error rates across most tested settings, with a maximum reported reduction of 47.1% versus the baseline and only a modest training-time increase.
Key insights
- Robust optimization depends not only on how group weights are updated, but also on whether the group losses being compared are meaningfully comparable.
- Duration matching and smoothed updates both contribute to CTC-DRO’s results; ablations show that removing either component can hurt performance.
- Lower character error does not guarantee improvements on every metric or in every setting: language-identification and word-error results are mixed, and some gains vary across random seeds.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.