Assistente Virtual Embarcada

Abstract

The Embedded Virtual Assistant (EVA) was researched, designed, implemented, and validated with the aim of offering a low-cost home automation solution that prioritizes local data processing and direct control of domestic devices, minimizing cloud dependency and user privacy risks. The system’s architecture follows an edge computing paradigm, dividing processing between an acquisition module (based on the ESP-32-WROOM-32 microcontroller) and a processing module installed on a local computer network. The acquisition module is responsible for audio capture via microphone and sound feedback reproduction, activating only when a button is pressed and communicating via Wi-Fi. Collected data is sent to the edge server via TCP/IP. The processing module, implemented in Python, performs computationally intensive tasks such as audio-to-text transcription using an artificial neural network framework (Vosk) and natural language processing for intention identification and device management. After processing, results are returned to the microcontroller in JSON format for action execution and user feedback. The methodology included a bibliographic study of Artificial Neural Networks for audio, comparative analysis of similar projects, and validation through manual prototype testing. The system is expected to interpret simple voice commands, such as lighting automation or answering questions, with high accuracy. This work demonstrates EVA’s technical and economic feasibility, combining energy efficiency with advanced processing capabilities. The modular architecture and focus on local processing optimize resources, meet demands for privacy and autonomy, and facilitate expansion and customization for various application scenarios, with security protocols ensuring reliable operation.