Reproduce “Enhancing Code-Switching Speech Recognition with LID-Based Collaborative Mixture of Experts Model” 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
Hukai Huang, Jiayan Lin, Kaidi Wang, Yishuang Li, Wenhao Guan, Lin Li, Qingyang Hong
A Collaborative Mixture of Experts (MoE) model is introduced to enhance code-switching ASR by explicitly integrating Language Identification (LID) for expert routing and multi-level collaboration. This approach from Xiamen University achieves a 10.17% relative performance improvement over a Conformer Baseline, while demonstrating reduced computational complexity for Mandarin-English code-switching scenarios.
Key insights
- Explicit Language Identification (LID) supervision for routing significantly enhances expert specialization and mitigates language confusion in MoE models.
- Combining LID-weighted inter-group collaboration with unsupervised intra-group collaboration effectively integrates diverse linguistic knowledge and refines attribute modeling.
- The model's design supports flexible allocation of experts to specific language groups, enabling adaptation to varied task requirements while preserving computational efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.