Google releases MTG-AD benchmark for inline audio descriptions

The 36-hour set has more than 8,000 descriptions, and end-to-end scoring checks whether narration fits existing video silences.

Researchers present MTG-AD, a professionally annotated dataset of 36 hours of videos, averaging about five minutes each, with more than 8,000 audio descriptions co-created by paired blind and sighted professionals. MTG-AD-Eval scores end-to-end inline descriptions by rendering scripts to audio and checking fit inside natural silences. Across frontier multimodal models, the authors report accurate visual-event description and fewer speech overlaps, with remaining gaps in narrative pacing and flow. Automated reference-free metrics correlated with ratings from sighted raters and seven blind or low-vision participants; users narrowly preferred human descriptions but found model output highly useful for comprehension. The dataset and evaluation code are released.