VoxMem: Benchmarking Multimodal Memory in Large Audio Language Models
The benchmark says leading audio language models still fail most multimodal memory tests.
VoxMem tests 15 large audio language models on 3,196 instances drawn from 34,743 spoken sessions, covering speech content, speaker identity, paralinguistic cues, and environmental sound. No model scores above 40% at a 32K context budget. The paper says models remember spoken words better than audio-only evidence such as who spoke, how they spoke, or what else was audible, especially as histories grow and tasks require reasoning across sessions. HF Daily Papers' note
VoxMem tests 15 large audio language models on 3,196 instances drawn from 34,743 spoken sessions, covering speech content, speaker identity, paralinguistic cues, and environmental sound. No model scores above 40% at a 32K context budget. The paper says models remember spoken words better than audio-only evidence such as who spoke, how they spoke, or what else was audible, especially as histories grow and tasks require reasoning across sessions. HF Daily Papers' note
score 4