A Hybrid Signal-Processing and Neural Network Pipeline for Low-Latency Acoustic Event Detection at the Edge

Abstract: This paper presents an edge-computed AI framework that detects a specific acoustic event — a human handclap — and actuates a wireless IoT device in response, without touching a cloud server at any point. Audio is captured through a professional condenser microphone, streamed into a Python processing pipeline via a non-blocking callback queue, and subjected to two lightweight deterministic gates before any neural computation occurs. The first gate discards frame whose peak amplitude falls below a fixed noise floor, keeping the classifier idle during silence. The second gate enforces a post-detection bypass window, preventing room echoes from generating spurious follow-on triggers. Frames that clear both gates pass through a two-stage neural pipeline: Google's YAMNet model extracts 1024- dimensional acoustic embeddings, and a purpose-trained Keras Sequential classifier maps those embeddings to a binary confidence score. When confidence exceeds 0.28, the system sends a UDP packet over the local Wi-Fi network to an ESP32 microcontroller, whose on-board LED toggles as hardware confirmation of successful end-to-end delivery. The entire loop — from live audio capture, through AI inference, to physical actuation — runs within a sub-second latency budget on commodity CPU hardware, with no internet dependency. The result is a privacy-preserving, network- independent architecture that generalises to any latency-critical acoustic IoT trigger application.

Keywords: Edge Computing; Acoustic Event Detection; YAMNet; Transfer Learning; Keras; ESP32; Real-Time Signal Processing; Deterministic Frame-Skipping; IoT Actuation.