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.

• →Loss of punctuation during ASR processing causes a massive 20.7% relative BLEU score drop in translation quality.
• →The optimized pipeline with Punctuation Restoration Module achieved 36.38 BLEU score versus 31.48 baseline on custom dataset.
• →Wav2Vec2-XLS-R-300m model achieved state-of-the-art 2.72% Character Error Rate on OpenSLR-54 benchmark.
• →Human assessment confirmed superior Adequacy (3.673) and Fluency (3.804) scores with reliable inter-rater agreement.
• →The research establishes architectural insights applicable to other low-resource language translation systems.