Audio QA

Audio question answering studies how models perceive speech and other sounds, connect acoustic cues to a question, and produce reliable answers. Approaches include audio-language models trained on question-answer data, structured perception and reasoning, retrieval over timestamped events, and tool-based evidence fusion. Benchmarks span speech, music, medical audio, long recordings, and multilingual or noisy settings, while evaluation increasingly tests temporal grounding, uncertainty, and whether answers truly depend on audio. A central challenge is that strong scores can coexist with sensitivity to wording or answer order, shortcuts from text priors, and failures when multiple reasoning steps must be combined.