Whisper fine-tuned with LoRA outperforms MFA in word boundary detection

Fine-tuning Whisper with LoRA on 47 hours of human-labeled boundary data achieves a word-boundary error of 11.6 ms, surpassing the Montreal Forced Aligner (MFA). The model predicts start and end times as timestamp tokens and performs better across English, Korean, and Dutch.