Reproduce “BLR-MoE: Boosted Language-Routing Mixture of Experts for Domain-Robust Multilingual E2E ASR” 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

Guodong Ma, Wenxuan Wang, Lifeng Zhou, Yuting Yang, Yuke Li, Binbin Du

The BLR-MoE architecture enhances Multilingual End-to-End Automatic Speech Recognition by addressing language confusion within both the self-attention mechanism and the language identification router. This framework from Yidun AI Lab at Netease achieved a 16.09% relative reduction in average WER/CER and a 9.41% absolute improvement in LID accuracy compared to the LR-MoE baseline, particularly in out-of-domain scenarios.

Key insights

  • Language confusion in Multilingual ASR can be independently and effectively mitigated within both the self-attention mechanism and the language routing component.
  • Applying the Mixture of Expert (MoE) principle to the Value (V) and Output (O) linear layers of the self-attention module successfully creates language-specific processing pathways, reducing cross-language interference.
  • Enhancing the Language Identification (LID) router's modeling capacity and enabling its decoupled fine-tuning significantly improves its robustness to domain shifts and thereby the overall ASR accuracy.
  • Expert pruning allows a single pre-trained multilingual model to be dynamically optimized for specific deployment scenarios by eliminating irrelevant language experts, reducing runtime confusion without retraining.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.