TinyML-Based Vietnamese Keyword Recognition for Smart Home Voice Control

Abstract

This study develops a TinyML-based voice control system specifically for the Vietnamese language, addressing the demand for localized smart home solutions. Unlike existing research focused on English, our work provides an accessible option for Vietnamese speakers. We collected a dataset comprising one wake-up keyword and eight common commands, integrating DSP algorithms and machine learning models on the ESP32-S3 MCU. The model, with 129,098 parameters, achieved an average accuracy of 94\% and a real-time execution time of 368 ms for data windows of 500-1000 ms. This highlights the potential of TinyML in voice-controlled devices, offering a low-cost, Internet-independent solution that enhances data security and user privacy.