Endophasia: Utilizing Acoustic-Based Imaging for Issuing Contact-Free Silent Speech Commands
summary
A phone app emits an up‑converted GSM sequence through its speaker and uses the microphone to recover channel impulse responses while a user mouths words silently, from which differential CIRs become 2D motion profiles that feed a ResNet18 CNN. Evaluation on 14 participants with 20 commands and 30 repeats each yields 92.5% accuracy after cold/warm/online training, tolerable angle and distance variation, and a cold‑start leave‑one‑user‑out baseline near 40%.
AI-written summary · about