AINeutralarXiv – CS AI · Mar 34/104
🧠
Mitigating Structural Noise in Low-Resource S2TT: An Optimized Cascaded Nepali-English Pipeline with Punctuation Restoration
Researchers developed an optimized speech-to-text translation pipeline for Nepali-to-English that addresses punctuation loss issues in low-resource language processing. By implementing a Punctuation Restoration Module, they achieved a 4.90 BLEU point improvement over baseline systems, demonstrating significant quality gains for cascaded translation architectures.