EchoLIFE: Zero-Shot In-Home ADL Recognition with LLM-Guided Active Acoustic Sensing

Abstract

Population aging is increasing the demand for in-home monitoring, yet practical recognition of activities of daily living (ADLs) remains hard to scale due to privacy concerns, installation overhead, and the limited cross-home generalization of data-hungry supervised models. We present an LLM-enabled, room-level ADL recognition system built on active acoustic sensing using one commodity speaker-microphone device per room. To enable zero-shot deployment across homes, we avoid end-to-end training and represent continuous echo streams as room-labeled segments with structured semantic summaries that remain comparable across different rooms and layouts. This representation separates low-level sensing evidence from high-level reasoning and provides a stable interface for semantic inference. An LLM then compiles the activity inventory, feature schema, home layout, and occupant routines into an executable RuleBook. During online monitoring, deterministic rule execution produces top-3 candidates with evidence, and a margin-triggered LLM verification step is invoked only for ambiguous segments to control cost. Evaluated on continuous in-home streams from 10 users over multiple days across 11 ADL classes, our system achieves cross-home zero-shot recognition on continuous time axes with 73.3% Top-1 and 88.5% Top-3 end-to-end accuracy, while reducing 75.7% LLM cost.