Reproduce “Dynamic Context-Aware Streaming Pretrained Language Model For Inverse Text Normalization” 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
Luong Ho, Khanh Le, Vinh Pham, Bao Nguyen, Tan Tran, Duc Chau
A streaming Inverse Text Normalization (ITN) system integrates a pre-trained language model with a dynamic context-aware mechanism to process spoken text in real-time. This system achieves an exceptionally low latency of 6.85 ms for 5-word chunks and an RTF of 0.138, enabling accurate text normalization for real-time applications, specifically demonstrated on Vietnamese.
Key insights
- Pre-trained language models provide rich linguistic representations that substantially improve ITN accuracy and robustness, even in streaming contexts and for low-resource languages.
- While limited context inherently reduces streaming ITN accuracy, incorporating dynamic right context during inference and training significantly recovers performance.
- Adapting to variable ASR output chunk sizes through dynamic context-aware mechanisms is critical for practical streaming ITN, preventing errors from fixed-chunk limitations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.