Researchers have introduced ScreenASR-Bench, a new benchmark designed to evaluate speech recognition models, particularly those incorporating visual context. The benchmark features three tiers: Explicit Textual Evidence (ETE), Explicit Visual Evidence (EVE), and Implicit Semantic Evidence (ISE), each presenting different levels of visual information. Experiments with various models, including Large Audio-Language Models (LALMs) and Omni models, demonstrated that visual context significantly enhances speech recognition recall compared to audio-only approaches. The study found that models performed best with direct visual evidence (ETE) and less effectively with indirect semantic cues (ISE), suggesting a need for further development in leveraging nuanced visual information. AI
IMPACT This benchmark could drive advancements in speech recognition by highlighting the benefits and challenges of integrating visual context.
RANK_REASON The cluster describes a new academic benchmark for evaluating AI models, published on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Explicit Textual Evidence
- Explicit Visual Evidence
- Hugging Face
- Implicit Semantic Evidence
- Large Audio-Language Models
- Omni models
- ScreenASR-Bench
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
