Reproduce “Dynamic Language Group-Based MoE: Enhancing Code-Switching Speech Recognition with Hierarchical Routing” 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, Shenghui Lu, Yahui Shan, He Qu, Fengrun Zhang, Wenhao Guan, Qingyang Hong, Lin Li
The Dynamic Language Group-based Mixture of Experts (DLG-MoE) model enhances code-switching Automatic Speech Recognition through a hierarchical routing mechanism that explicitly identifies languages and implicitly models sub-linguistic variations. This approach achieves significant performance improvements over existing methods and offers flexible deployment options.
Key insights
- The combination of explicit language routing and implicit sub-linguistic routing within language groups effectively decouples and models the multi-faceted complexities of code-switching speech.
- A dynamic top-k training strategy allows the model to generalize across various inference-time computational budgets, maintaining performance while adapting to different efficiency requirements.
- The use of CTC for language identification, even without explicit frame-level LID labels, provides an accurate and causal mechanism suitable for streaming ASR and enables flexible model pruning for monolingual tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.