Research map / Learning theory and algorithms / Prediction and decision losses
CTC-DRO: Robust Optimization for Reducing Language Disparities in Speech Recognition
Martijn Bartelds, Ananjan Nandi, Moussa Koulako Bala Doumbouya, Dan Jurafsky, Tatsunori Hashimoto, Karen Livescu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CTC-DRO instead of baseline fine-tuning | MMS/XLS-R multilingual ASR · ML-SUPERB 2.0 sets 1-5 | worst-language CER: 90.0 -> 57.5 (MMS, set 5); 114.8 -> 71.5 (XLS-R, set 5) | Improved |
| use CTC-DRO instead of baseline fine-tuning | MMS/XLS-R · ML-SUPERB 2.0 sets 1-5 | average CER: 25.2 -> 22.5 (XLS-R set 1); 29.9 -> 23.8 (XLS-R set 5) | Improved |
| use original Group DRO instead of baseline | MMS/XLS-R · ML-SUPERB 2.0 sets 1-5 | average CER: 23.4 -> 30.5 (MMS set 1) | Hurt |
| use original Group DRO instead of baseline | MMS/XLS-R · ML-SUPERB 2.0 | worst-language CER: 60.8 -> 86.6 (MMS set 1); 90.0 -> 62.2 (MMS set 5) | Mixed |
| add length-matched group losses to CTC-DRO (ablation) | CTC-DRO XLS-R · Set 5 | worst-language CER: 115.2 -> 71.5 | Improved |
| add smoothed maximization objective to CTC-DRO (ablation) | CTC-DRO XLS-R · Set 5 | worst-language CER: 194.2 -> 71.5 | Improved |
| add length-matched group losses to CTC-DRO (ablation) | CTC-DRO MMS · Set 5 | average CER: 29.5 -> 24.3 | Improved |
| add smoothed maximization objective to CTC-DRO (ablation) | CTC-DRO MMS · Set 5 | average CER: 97.9 -> 24.3 | Improved |
| add additional unbalanced training data, CTC-DRO vs baseline | XLS-R · Set 2 (additional data) | worst-language CER: 97.2 -> 51.4 | Improved |
| scale to 18 languages | CTC-DRO · 18-language set (balanced) | worst-language CER: 8.9% (MMS), 9.2% (XLS-R) relative reduction | Improved |
Closest papers by meaning
- CR-CTC: Consistency regularization on CTC for improved speech recognition (ICLR 2025 · similarity 0.85)
- Speech Robust Bench: A Robustness Benchmark For Speech Recognition (ICLR 2025 · similarity 0.85)
- Objective Soups: Multilingual Multi-Task Modeling for Speech Processing (NeurIPS 2025 · similarity 0.85)
- Align With Purpose: Optimize Desired Properties in CTC Models with a General Plug-and-Play Framework (ICLR 2024 · similarity 0.85)
- LAMA-UT: Language Agnostic Multilingual ASR Through Orthography Unification and Language-Specific Transliteration (AAAI 2025 · similarity 0.85)
- Enabling Differentially Private Federated Learning for Speech Recognition: Benchmarks, Adaptive Optimizers, and Gradient Clipping (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
