Scalable and Personalized Oral Assessments Using Voice AI
Summary
Two cohorts show operational feasibility, while agreement among model graders does not establish validity.
The July 26 v3 covers 73 students and engineering fixes for stacked questions and biased case selection; interruptions persisted.
It distinguishes marginal grading costs from subscription and voice costs, and explicitly lacks blinded expert validation and subgroup fairness analysis.
Two cohorts in one course with simultaneous changes; the 2025 blog belongs to the same research lineage.
Sources & reading record
- Panos Ipeirotis & Konstantinos Rizakos · arXiv / CACM ↗
- Read v3 design, costs, reliability, failures, student experience and validity limitations; system not run.
- Panos Ipeirotis · original classroom blog, 2025-12-29 ↗
- Read the original report and paper-expansion note; demonstration video not watched.