Reproduce “UCorrect: An Unsupervised Framework for Automatic Speech Recognition Error Correction” 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
Jiaxin Guo, Minghan Wang, Xiaosong Qiao, Daimeng Wei, Hengchao Shang, Zongyao Li, Zhengzhe Yu, Yinglu Li and 4 more
Huawei Translation Services Center developed UCorrect, an unsupervised framework for Automatic Speech Recognition (ASR) error correction that leverages pre-trained language models to reduce reliance on paired training data. The system achieves a 14.29% Word Error Rate Reduction with lower false alarm rates and demonstrates universality across various ASR models and decoding strategies.
Key insights
- Unsupervised learning for ASR error correction is viable and effective, significantly reducing data dependency compared to traditional supervised approaches.
- A modular Detector-Generator-Selector architecture provides improved interpretability and controllability, leading to lower False Alarm Rates.
- Incorporating phonetic similarity in candidate generation is crucial for effectively addressing typical ASR errors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.