Measuring What Was Said, Not Only How It Sounded: A Hybrid Deterministic and Generative Architecture for Automated Assessment of Children's Impromptu Speech

Automated speech coaching tools have converged on a narrow set of metrics: speaking rate, disfluency counts, and pause frequency. These are measured because they are cheap to measure, not because they are what makes a spoken argument good. The result is a category of products that can tell a speaker they said "um" eleven times but cannot tell them their argument was unsupported. This paper describes the architecture of a system that assesses both halves, for speakers aged 6 to 14. It advances four design claims: that delivery measurement should be deterministic rather than generative, since a language model receives text and cannot hear audio; that content quality is assessable and is the higher-value signal; that acoustic baselines must be within-subject when the speaker population is developmentally heterogeneous; and that withholding numeric scores from the end user is an assessment design decision rather than a presentation preference.