Answers are grounded in this paper and cite their sources.
Automatic speech recognition (ASR) often performs poorly on regional dialects. Standard models are trained on standardspeech, so they struggle with dialect sounds and replace regional words with standard ones. We introduce a Dialect-AwareSparse Mixture-of-Experts (MoE) Conformer for Marathi ASR. Our model routes speech frames autonomously withoutneeding dialect labels at test time. We replace feed-forward modules in encoder blocks 5–12 with a shared trunk and threeregional experts. Total model size stays identical to the dense baseline at 130.8M parameters, but each frame activatesonly 97.2M parameters (a 25.7% saving). On the IISc RESPIN test set (160 unseen speakers), our model achieves 19.50%Word Error Rate (WER), outperforming the dense baseline (36.13% WER) by 16.63 percentage points (46.0% relativeerror reduction, p < 0.0001). It also beats the 4.6× larger IndicConformer-600M (23.94% WER) and oracle-guided LoRAadapters (34.49% WER), while improving recall of regional dialect words from 62.9% to 91.2%. All code, checkpoints, andevaluation scripts are publicly available.
No takes yet. Share an insight, caveat, or question.
Kinjawadekar et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: