PetTalk: A Latency-Aware Edge--Cloud System for Real-Time Affective Interaction

Companion-animal vocalizations are brief and event-driven, making real-time interaction a joint problem of continuous sensing, selective cloud invocation, and response latency. We present PetTalk, a deployed edge–cloud system that closes the loop from companion-animal vocalization to spoken affective interaction. An always-on edge gate filters non-target audio before cloud-based affective-vocalization classification, response generation, and streaming speech synthesis. A pre-synthesized greeting provides early audible feedback while the complete LLM–TTS response continues. Across 40 stratified test clips, greeting playback onset is directly measured end to end at the speaker at 1.737 s p50 and 3.003 s p95. Complete-response onset is instead estimated from stage-wise percentiles at 2.86 s and 5.37 s, with LLM generation dominating the latency budget. On a deterministic 90-minute household replay, the full edge gate reduces would-be cloud activations to 4.0% for cats and 3.1% for dogs, with 100% and 48% target recall, respectively. Supporting evaluation on an expert-annotated 1,346-clip set yields 95.7/94.0% classification accuracy for cats/dogs, while a convenience-sampled formative study (N=20) reports 9.0/10 latency acceptability. Together, these results support an end-to-end latency-aware interaction loop under the evaluated setting while exposing the current dog-gating limitation.