Speech as a Tool: An Engineering Framework for High‑RoE Communication

Speech is humanity’s strongest coordination technology, yet we rarely treat it as anengineered system. This thesis reframes speech as a toolchain—Intention →Compilation → Actuation → Feedback—and proposes Return on Expression (RoE), therealized state change per unit cost, as the organizing objective. We present (i) aformal state-space/control model of utterances as interventions on beliefs, goals, andcommitments; (ii) an empirically calibrated affordance map that translates contextvariables (uncertainty, risk, expertise asymmetry, channel constraints) into designchoices (framing, redundancy, prosody); (iii) a failure taxonomy with detectablesignals and mitigation protocols; and (iv) a Speech IDE that instruments conversationwith an LLM-in-the-loop intention compiler, ambiguity linters, a commitment typechecker, and a grounding meter. We define standardized micro‑protocols (e.g., Query→ Calibrate → Commit; Goal → Constraint → Offer → Check) and preregisteredexperiments across lab and field settings to estimate treatment effects on RoE,commitment conversion, grounding depth, and rework. Bayesian hierarchical modelssupport estimation and transportability across domains. Ethics features(consent‑to‑influence, bias audits, kill‑switches) reduce misuse. If validated inpreregistered A/B tests (≥20% CCR lift and ≥25% ambiguity‑error reduction), thisframework provides a measurable, reproducible way to teach speech as acraft—raising coordination reliability in education, healthcare, and operations whileestablishing RoE as a first‑class design objective.