Evaluating the Importance of Demographic and Technical Factors in Creating Authentic-Sounding AI-Generated Human Voice Clones; Deepfakes and the Social Construction of Trust: A SCOT Analysis of Stakeholder Responses

This project pairs engineering tests of AI voice-cloning tools with a social-science look at how deepfakes chip away at trust. My team and I built a library of authentic and cloned voices, then asked listeners to judge each clip’s realism; the same clips were also scored by a machine evaluation model (NISQA). Most clones still sounded less genuine than the authentic voices, but clones trained on short samples from young male speakers and mixed with background noise fooled listeners as often as the originals. NISQA scores did not track these perceptions, highlighting a gap between technical ratings and what people hear. Using the Social Construction of Technology framework, I compared how platforms, banks, courts, and the public frame the deepfake problem. Banks, facing direct losses, quickly adopted multi-factor security, while media firms, regulators, and courts still disagree on standards. The analysis suggests four steps toward closure: cryptographic watermarks tied to platform checks, early courtroom reliability hearings, annual audits of detection accuracy, and broad media-literacy programs. Technical insights guide where safeguards are most needed, and social insights show how they might gain acceptance.