> Word-level timestamps. Each word has precise start/end times and confidence scores.

Neat. This is missing from many models.