Energy-Efficient Hybrid AI Tutor for Learning Varieties of English

Energy-aware design is becoming a practical constraint for speech-first language tutors, especially when learners expect fluent interaction across multiple varieties of English (AmE, BrE, Indian English, AusE, and CanE). We compare edge-only, cloud-only, and hybrid edge-cloud tutoring under a fully specified simulation that includes an on-device ASR profile, a compressed local tutor model, cloud GPU allocation with PUE, transport energy, and repeated trials. The revised evaluation reports not only energy and latency but also normalized learning gain ΔS and learning efficiency LE=ΔS/Etotal. Across 150 simulated 5-minute sessions, the hybrid design with caching uses 0.37+/-0.04 Wh, offloads 15.6+/- 4.8% of turns, preserves dialect-style accuracy within about one point of the cloud baseline, and improves LE over cloud-only by 46% when LE is reported as gain per Wh. These results suggest that uncertainty-gated hybrid inference can provide a practical energy- accuracy trade-off for dialect-aware tutoring, while the present findings should be interpreted as simulation-based rather than as evidence from physical deployment or learner trials.