PiSoundSensing demonstrates CNN-based audio tagging on a low-cost standalone embedded device, with both physical and browser-based interaction.
Strong CNN audio classifiers are typically evaluated on workstation-class hardware. This project examines what happens when a pre-trained audio model is moved onto a resource-constrained embedded platform intended to operate as a standalone sensing device.
Users can interact locally through a physical button or remotely through a web interface. The browser interface controls the device and visualises detected sound events over time.
The demonstration documents the hardware and software stack and compares observations from deployment on the Raspberry Pi with software-only execution, making the gap between model evaluation and real embedded operation visible.
