Reproduce “Contextualized End-to-End Speech Recognition with Contextual Phrase Prediction Network” 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
Kaixun Huang, Ao Zhang, Zhanheng Yang, Pengcheng Guo, Bingshen Mu, Tianyi Xu, Lei Xie
Researchers at Northwestern Polytechnical University introduce a Contextual Phrase Prediction Network and a two-stage filtering method to provide explicit supervision for contextual biasing in end-to-end Automatic Speech Recognition. This approach achieved a 40.5% relative reduction in Biased Word Error Rate and maintained performance effectively with large contextual lists.
Key insights
- Providing explicit supervision for contextual phrase recognition through a dedicated network and loss function substantially enhances the effectiveness of deep biasing in ASR.
- The proposed two-stage contextual phrase filtering strategy effectively addresses the challenge of large biasing lists, preventing Unbiased Word Error Rate (U-WER) degradation and maintaining efficient inference.
- Deep contextual biasing methods, especially with explicit supervision, can yield much greater improvements in recognizing rare words compared to traditional shallow fusion techniques.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.