Most mixture-of-experts approaches for accented speech recognition assign experts to specific accents, assuming accent identity is the main source of variation. The authors of this paper argue instead that accents systematically distort phoneme realization, so adaptation should be organized along phonetic categories. Their proposed framework, MoPE-LoRA, fixes six low-rank experts aligned with manner-of-articulation classes and routes frames through them using a hybrid mechanism that blends phoneme supervision with learned acoustic gating.
Because the expert set is shared across accents, the model needs no accent-specific modules and can in principle handle unseen accents. Experiments on the L2-ARCTIC benchmark show consistent gains over single LoRA and full fine-tuning. In a zero-shot setting where one accent is withheld during training, MoPE-LoRA achieves a 12.3% relative improvement over single LoRA.
The authors interpret these results as evidence that phonetic-category decomposition is a scalable and transferable alternative to accent-level adaptation. The paper was accepted at Interspeech 2026 and is available on arXiv.