SEAM: Shortcut-Aware Real-Time Detection of Scripted vs. Spontaneous Speech for Interview Guardrails
Researchers have developed SEAM, a new framework designed to accurately distinguish between scripted and spontaneous speech in real-time, specifically for interview guardrails. This system addresses the issue of benchmark performance being inflated by corpus-specific shortcuts rather than genuine speech style detection. SEAM incorporates uniform preprocessing, seam-aware sampling, non-speech augmentation, and a compact DistilHuBERT backbone, achieving a 0.971 ROC-AUC on an external evaluation set with 8-second windows. The framework's effectiveness is attributed to its shortcut-prevention components, demonstrating that robust detection relies on both the model architecture and careful data design and evaluation. AI
IMPACT This framework could improve the accuracy of AI-powered interview analysis tools by better distinguishing between genuine and rehearsed responses.