How well can VLMs rate audio descriptions: A multi-dimensional quantitative assessment framework

TL;DR

The paper tackles the challenge of scalable, high-quality audio description (AD) for full-length videos by introducing a multi-dimensional assessment framework that combines content and formatting dimensions, grounded in DCMP guidelines and expert input. It then operationalizes this framework in a comprehensive workflow using Item Response Theory (IRT) to compare human raters and state-of-the-art vision–language models (VLMs) against expert ground truth, across 10 videos and 40 AD variants. Key findings show VLMs can closely approximate expert judgments on most dimensions, though their justifications are less actionable, underscoring the value of hybrid human–AI evaluation for scalable AD quality control. The work advances accessible media evaluation by formalizing formatting considerations (delivery method and timing), providing a replicable methodology, and highlighting the complementary strengths of humans and AI for delivering reliable, scalable AD quality assessments.

Abstract

Digital video is central to communication, education, and entertainment, but without audio description (AD), blind and low-vision audiences are excluded. While crowdsourced platforms and vision-language-models (VLMs) expand AD production, quality is rarely checked systematically. Existing evaluations rely on NLP metrics and short-clip guidelines, leaving questions about what constitutes quality for full-length content and how to assess it at scale. To address these questions, we first developed a multi-dimensional assessment framework for uninterrupted, full-length video, grounded in professional guidelines and refined by accessibility specialists. Second, we integrated this framework into a comprehensive methodological workflow, utilizing Item Response Theory, to assess the proficiency of VLM and human raters against expert-established ground truth. Findings suggest that while VLMs can approximate ground-truth ratings with high alignment, their reasoning was found to be less reliable and actionable than that of human respondents. These insights show the potential of hybrid evaluation systems that leverage VLMs alongside human oversight, offering a path towards scalable AD quality control.

Figures (19)

• Figure 1: Conceptual diagram of the major components in our evaluation workflow. From 10 videos, we collected 40 audio descriptions (1 human-authored and 3 VLM-generated per video). Three experts established ground-truth ratings using our seven-dimension assessment framework. Additional respondents, four humans and eight VLMs, applied the same framework, and their ratings were compared to the ground truth. Item Response Theory (IRT) was then used to model evaluator proficiency and item difficulty.
• Figure 2: Ten YouTube videos used in the evaluation, grouped by category: Entertainment (top), How-to & Style (middle), and Education (bottom). Subcaptions show the YouTube titles and video length.
• Figure 3: VLM-generated description pipeline. Videos are segmented into scenes, a VLM generates scene-level descriptions, a second VLM pass optimizes them, and the outputs are resynthesized and synchronized with the video.
• Figure 4: Viewing interface for synchronized playback of audio description with video.
• Figure 5: A snippet of the prompt used to instruct the VLMs at applying the 7 dimensional assessment framework to evaluate AD.