SEAM: Real-time AI detects scripted speech in interviews

SEAM (Shortcut-Aware Detection of Scripted vs Spontaneous Speech) tackles a critical problem in automated interview integrity: many benchmark models inflate performance by learning shortcuts tied to recording artifacts or corpus identity rather than actual speaking style. The authors, Vsevolod Kovalev and Pranay Manocha, designed SEAM to strip away these shortcuts through three key preprocessing strategies—uniform preprocessing across datasets, seam-aware sampling that avoids splicing artifacts, and non-speech augmentation. The core model uses a compact DistilHuBERT backbone optimized for real-time inference on 8-second audio windows.

On an external interview-domain evaluation set, SEAM achieves 0.971 ± 0.004 ROC-AUC, demonstrating robust generalization. Crucially, removing the shortcut-prevention components improves internal metrics but sharply degrades external performance, confirming the presence of shortcut learning in prior work. Quantization reduces the model footprint to just 41.8MB with minimal accuracy loss, making deployment on edge devices feasible. Accepted at Interspeech 2026, the authors have released code and model checkpoints, enabling hiring platforms, interview practice tools, and academic researchers to implement fairer, more accurate spoken-language assessments.