Tag
The paper introduces StyleInstructCaps, the first standardized benchmark for controllable speaking-style captioning (SSC) in audio-language models, evaluating groundedness, hallucination, instruction-following, and speaker consistency across nine state-of-the-art models.