Reproduce “Voice Agents under Acoustic Stress: From Signal Degradation to Interaction and Action” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Amir Ivry, Kai-Wei Chang, Lin Zhang, Sharon Gannot, Carlos Busso
TRACE reframes voice-agent robustness as a task-level safety and interaction problem: the key question is not only whether acoustic stress harms recognition, but whether it causes incorrect actions, failed recovery, or extra user effort. The proposed paired workflow compares clean and acoustically stressed recordings while tracking goal attainment, prohibited actions, clarification success, timing, and interaction cost.
Key insights
- Acoustic robustness should be measured by the change between matched clean and stressed conditions, not by performance on noisy inputs alone.
- Clarification is only a successful recovery when the agent obtains the missing information and uses it correctly without having already performed a prohibited action.
- Final task completion and interaction quality are distinct: premature actions, interruptions, repeated information requests, and added effort can matter even when the eventual goal is reached.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.