Efficient Punctuation Restoration via Weighted Lookahead Scoring Method for Streaming ASR Systems

In three linesNon-autoregressive scoring method for punctuation restoration in streaming ASR with bounded K-subword lookahead. Preserves input transcript, decides at word boundaries via hypothesis comparison. IWSLT 2017: macro F1 0.893 without fine-tuning (K=2) and 0.937 with fine-tuning, outperforming prompt-based and ELECTRA baselines.

Summary generated by Claude — human-verified