Researchers have introduced Andha-Dhun, the first dataset of human-authored audio descriptions (ADs) in Hindi, collected from eight full-length movies. This work addresses the gap in ADs for Indian languages, particularly for Blind and Low Vision (BLV) audiences, following mandates from India's Central Board of Film Certification. The study explores generating Hindi ADs through direct translation of English descriptions or by translating existing English ADs, evaluating them with perplexity and LLM-as-a-judge metrics. Findings indicate that direct translation can introduce artifacts and reduce diversity compared to original Hindi ADs, highlighting the need to adapt content for the target audience's accessibility rather than strictly adhering to the source. AI
IMPACT This work advances accessibility for visually impaired audiences by enabling AI-generated audio descriptions in a new language, potentially influencing future multimodal AI development for diverse linguistic contexts.
RANK_REASON The cluster contains a research paper introducing a new dataset and methodology for audio descriptions in Hindi. [lever_c_demoted from research: ic=1 ai=1.0]
- Andha-Dhun
- arXiv
- Central Board of Film Certification
- English
- Hindi
- Hugging Face
- India
- LLM-as-a-judge
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
