Voice-activated home automation system for IoT edge devices using TinyML

Abstract Home automation systems are popular because they enhance the quality of life and the way users interact with the environment. Deploying complex machine learning models on Internet of Things (IoT) devices with limited resources is still difficult. This study proposes a home automation system based on a TinyML (Tiny Machine Learning) model to recognize specific spoken keywords. The developed model runs effectively on IoT devices which usually have limited resources. Using TinyML, the limitations of memory size, processing power and latency associated with IoT devices are addressed. The objective of this research is to train a keyword-spotting model for devices with low computation and memory. The trained TinyML model can recognize specific voice commands associated with home automation tasks, such as controlling lights, thermostats, and other appliances. To test our approach, we ran experiments in real-world settings and on edge IoT devices with limited resources. The results show that our keyword spotting model is both highly accurate and efficient and uses minimum computational resources. This research helps in the advancement of TinyML applications in home automation and broadens the potential for voice interaction in constrained environments. The keyword spotting model in the proposed system is built using Deep Convolutional Neural Network (DCNN). Different data pre-processing techniques are also applied to refine the dataset. The trained model is then converted to be deployed on the low resource devices without compromising the model’s efficiency. The model attains an 96.67% test accuracy. The model is quantized for devices with limited resources. It operates with an 11 ms latency, using 19.8 K of RAM and 55.0 K of flash for recognizing and classifying users’ voice commands in real-time. This demonstrates how TinyML can create efficient and user-friendly smart home solutions. The main contribution of the work presented in this paper is that the designed model can be deployed on a wide range of IoT devices. Since the model is trained on voice instructions which limits the model’s robustness. In future work, this limitation can be eliminated by integrating multilingual instructions.