CAST-Bench is a benchmark for multi-shot audio‑video generation that traces every generated sound back to its intended source. The key idea formulates T2AV correctness as an audiovisual contract: each requested sound must be emitted by the correct visual entity, synchronized with the action, and carry the right semantic meaning. The benchmark organizes 900 prompts into 9 sound‑source categories (speech, animals, instruments, ambient sounds, etc.), 100 prompts per category. Among them, 439 are single‑shot and 461 are multi‑shot narratives, all derived from a structured specification that links each audio event to its emitter, subject, and temporal slot. An 18‑metric suite evaluates Visual Performance, Content Consistency, and Audio‑Semantic Alignment. Testing 12 representative generators shows visual quality is near saturation, while audio semantic scores still vary widely. Multi‑shot prompts receive lower aggregate scores under subset‑specific rules, altering model rankings. Persistent challenges include multi‑person speech, consistent speaker identity across shots, and cross‑shot event retention. All prompts and evaluation code are publicly released.
Review: CAST-Bench highlights the current gaps in sound‑source attribution for multi‑shot audio‑video generation and offers a fine‑grained evaluation framework that can guide future model improvements.