Reproduce “WeNet 2.0: More Productive End-to-End Speech Recognition Toolkit” on OpenResearch

About this paper

Binbin Zhang, Di Wu, Zhendong Peng, Xingchen Song, Zhuoyuan Yao, Hang Lv, Lei Xie, Chao Yang and 2 more

WeNet 2.0 from Horizon Robotics and NPU's ASLP group presents a comprehensive update to its end-to-end speech recognition toolkit, focusing on production-oriented features. The toolkit includes a U2++ framework achieving up to a 10% relative error rate reduction, an n-gram LM integration, a contextual biasing framework, and a Unified IO system enabling efficient training on 10,000-hour corpora in 75 minutes per epoch.

Key insights

• Integrating a Right-to-Left (R2L) Attention Decoder into the U2 framework (U2++) enhances the shared encoder's representation and system generalization by leveraging future context.
• Efficiently compiling and integrating n-gram Language Models (LMs) with E2E models via WFSTs allows for robustly leveraging external text data in production ASR systems.
• A unified contextual biasing framework that constructs on-the-fly WFST graphs with 'failure arcs' enables rapid, user-specific adaptation without expensive model retraining, while preventing over-biasing.

Summary by alphaXiv.