VoiceGuard: An Evaluation-Honest, Multi-Signal Decision Layer for Real-Time Voice-Clone Detection in High-Stakes Telephone Calls

VoiceGuard: An Evaluation-Honest, Multi-Signal Decision Layer for Real-Time Voice-Clone Detection in High-Stakes Telephone Calls This paper investigates the generalization limitations of speech anti-spoofing systems for detecting voice-clone fraud in telephone calls. It examines evaluation artifacts in the ASVspoof 2019 Logical Access benchmark and re-evaluates the published AASIST model using an anti-shortcut evaluation protocol. The paper proposes VoiceGuard, a multi-signal decision layer that combines acoustic spoof detection, speaker consistency, prosodic features, and call context to generate ALLOW, VERIFY, or ESCALATE recommendations. It also introduces a feature-only audit log to support transparency and human review. Additionally, a five-language pilot study evaluates the transferability of an English-trained detector across Indian languages, highlighting the challenges of multilingual voice-clone detection. The work emphasizes evaluation integrity, cross-dataset generalization, explainable decision-making, and the need for further validation before deployment in high-stakes environments. Authors: Ghanshyam Kumavat, Sunil Saini, and Vicky PukaleAffiliation: Department of Computer Engineering, Vidyalankar Institute of Technology, Mumbai, India.