Audio-Language Model Hallucination Detection

The field studies when audio-language models make claims that are not supported by the sound, including errors in transcription, event descriptions, and question answering. Evaluation uses human-checked audio-question pairs, speech recordings, and environmental-sound and music benchmarks, often testing temporal reasoning, absent events, and adversarial prompts. Detection methods range from language-model judges and consistency or uncertainty scores to probes of hidden states and attention; mitigation includes contrastive decoding, evidence verification, and preference training with perturbed audio. Models still struggle to ground claims in complex acoustic scenes, while stronger filtering can also remove correct answers, and methods may not generalize across models or tasks.