Developing Tinyml For Real-Time Intervention In Intimate Partner Violence Using Edge Computing On Esp32 Microcontroller

Sammanfattning

As Internet of Things (IoT) technologies have become increasingly common in domestic environments, the nature of Intimate Partner Violence (IPV) has also evolved. This study explores how IoT architecture can be leveraged to detect and intervene in incidents involving IPV. Guided by the Design Science Research Methodology, we developed a TinyML prototype capable of Speech Emotion Recognition (SER) on a simulated ESP32 microcontroller, using the EdgeImpulse platform, with the aim of detecting IPV-based emotional states directly on the edge in a clean audio environment. By utilizing the RAVDESS dataset, MFCCs were extracted from the audio and emotional states were regrouped into a binary classification of Safe and Distress. The model’s parameters were then optimized for the chosen microcontroller. The resulting prototype successfully executed inference in real-time without cloud reliance. Evaluation of the model demonstrated a weighted average F1-score of 78% during validation and a 73% during testing. Furthermore, the system achieved an inference time of 281 ms and utilized 55.2 kB of RAM. These findings conclude that deploying a TinyML model for IPV detection and intervention on an edge device is feasible, in a controlled audio environment.