저전력 임베디드 엣지 기기에서의 음성 전사: 음성인식 AI 모델 벤치마크

초록

Embedded speech systems increasingly run where cloud offload is a poor fit?factory robots, clinic kiosks with intermittent connectivity, classroom recorders with strict privacy rules, and mobile carts that must transcribe on the move. We benchmark on-device ASR with Whisper.cpp on NVIDIA Jetson TX1/TX2 boards. Using 10 s and 60 s audio clips, we measure real-time factor (RTF), power, and energy. Results show TX1 consistently transcribes faster than TX2 for all model sizes, while TX2 is more power-efficient, revealing a clear speed?energy trade-off. Our benchmarking pipeline provides a reproducible baseline for deploying Whisper ASR on power-constrained devices