Reproduce “Adaptive Contextual Biasing for Transducer Based Streaming Speech Recognition” 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
Tianyi Xu, Zhanheng Yang, Kaixun Huang, Pengcheng Guo, Ao Zhang, Biao Li, Changru Chen, Chao Li and 1 more
Researchers at ASLP@NPU developed an adaptive contextual biasing method for streaming speech recognition that intelligently applies bias only when context words are detected. This approach reduces Word Error Rate (WER) by up to 96.7% in common speech scenarios, mitigating false alarms while maintaining robust performance in personalized contexts with negligible Real-Time Factor (RTF) increase.
Key insights
- Integrating an Entity Detector (ED) into the Transducer architecture allows dynamic control over contextual biasing, enabling the system to intelligently apply or disable bias based on predicted context word presence.
- Utilizing biased encoder and predictor embeddings as input to the ED improves its ability to differentiate between acoustically and semantically similar words, reducing false alarms.
- The auxiliary cross-entropy loss for the ED (L_bias) not only trains the detector but also provides beneficial supervision during the overall ASR model's training, leading to better optimized performance even when the adaptive switching isn't fully utilized.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.