Research map / Large language models / Language, safety and interpretability
Convex Low-resource Accent-Robust Language Detection in Speech Recognition
Miria Feng, William Tan, Mert Pilanci
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 |
|---|---|---|---|
| add CLD convex language-detection head on Whisper-Small instead of default detector | Whisper-Small · Multiclass ASR, 5 languages | WER: 139.37 -> 31.74 | Improved |
| add CLD head instead of default detector | MMS-1B · Multiclass language ID | detection accuracy: 0.6701 -> 0.9702 | Improved |
| use convex CLD head instead of non-convex NN head (ablation) | Whisper-Small · Multiclass language ID | detection accuracy: 0.7737 -> 0.9715 | Improved |
| use CLD instead of NN head on Whisper-Large-V3 (ablation) | Whisper-Large-V3 · Multiclass language ID | detection accuracy: 0.9605 -> 0.9806 | Improved |
| use CLD instead of linear or kernel SVM on MMS-1B (ablation) | MMS-1B · Multiclass language ID | detection accuracy: 0.5701 (kernel SVM) -> 0.9702 | Improved |
| use KNN as language detector (ablation) | Whisper-Small/Large/MMS · Multiclass language ID | detection accuracy: 0.7154 (default) -> 0.6123 (KNN) on Whisper-Small | Hurt |
| use CLD instead of fine-tuned NN or WSP-SFT as training data shrinks (ablation) | Whisper-Small · Binary English/Mandarin accents, 100-10000 samples | accuracy and WER: CLD 96.94% to 99.14% accuracy across sizes | Improved |
| use CLD instead of NN or WSP-SFT for training cost (ablation) | Whisper-Small · Multiclass dataset | training time (s): 840.30 -> 64.45 (NN), 1,096.74 (WSP-SFT) | Improved |
| use CLD instead of NN on dialect-imbalanced English/Chinese (ablation) | Whisper-Small · English/Chinese dialects, 500 samples | accuracy (total): 0.5580 -> 0.9695 | Improved |
| use CLD with Whisper-Large-V3 versus NN on WER (ablation) | Whisper-Large-V3 · Multiclass ASR | WER: 29.25 -> 28.60 | No clear effect |
Closest papers by meaning
- “Allot?” is “A Lot!” Towards Developing More Generalized Speech Recognition System for Accessible Communication (AAAI 2024 · similarity 0.85)
- SwitchLingua: The First Large-Scale Multilingual and Multi-Ethnic Code-Switching Dataset (NeurIPS 2025 · similarity 0.85)
- DialectLLM: A Dialect-Aware Dialog[ue] Generation Framework Beyond Standard American English (NeurIPS 2026 · similarity 0.85)
- Objective Soups: Multilingual Multi-Task Modeling for Speech Processing (NeurIPS 2025 · similarity 0.85)
- MultiTalk: Scaling Full-Duplex Speech Models to Long, Multi-Party, Bilingual Conversation (NeurIPS 2026 · similarity 0.84)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.84)
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.
