PARSA-Bench evaluates what audio-language models actually understand when they listen to Persian: 8,000+ audio samples across 16 tasks, combining real recordings with TTS-generated audio.
Alongside standard tasks — ASR, speech translation, intent detection, named-entity recognition, question answering, emotion recognition — it introduces culturally grounded tasks that had never been benchmarked in any language: classical poetry meter (vazn), poetry style (sabk), and Dastgah classification of Iranian classical music. Ten of the sixteen tasks are the first of their kind.
With M. Amini, P. Bathayan, H. Faili, and A. Shakery.