Reproduce “Breaking the Script Barrier: Enabling Automatic Alignment for PoS-based ASR Error Analysis in Non-Latin Scripts” 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
Prasenjit K Mudi, Dahlia Devapriya, Sheetal Kalyani
Researchers at the Indian Institute of Technology Madras developed a character-spacing-aware Needleman-Wunsch algorithm to accurately align ASR hypotheses and references across diverse writing systems, particularly non-Latin scripts, which existing tools struggle with. This improved alignment facilitates automated Part-of-Speech (PoS) based error analysis and, when integrated into a Transformer decoder, reduces Word Error Rate (WER) by 0.5% for Russian and 1.3% for Tamil.
Key insights
- Standard ASR alignment tools fundamentally fail for non-Latin scripts due to fixed character width assumptions, making robust character-spacing-aware alignment crucial for accurate linguistic error analysis.
- Part-of-Speech (PoS) based error analysis reveals distinct, language-specific error patterns and model weaknesses (e.g., high error rates for nouns/proper nouns in agglutinative languages), providing fine-grained insights beyond aggregate metrics.
- Integrating fine-grained PoS-wise error vulnerability information directly into ASR model architectures, such as weighting decoder attention, can systematically improve recognition accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.