Energy Efficient Speech Command Recognition for Private Smart Home IoT Applications

Speech command recognition (SCR) services for smart homes rely on enterprise cloud servers and a fast internet connection. This negatively impacts service availability in offline use, and it raises privacy concerns in online use. The utilization of high computational resources on Internet-of-Things (IoT) nodes and IoT gateways, allows for traditionally cloud-based services to be implemented closer to the end users. In this paper, we propose a SCR system that consists of an energy-efficient IoT node that communicates over a private BLE home network with an IoT gateway. Feature extraction on sound signals is performed on the IoT node, and speech command classification is performed on the IoT gateway using a novel deep neural network (DNN) model. We evaluate the DNN model accuracy of the proposed system using k-fold cross validation, and tune it based on the effect of different feature extraction parameters on the prediction accuracy, the IoT node energy consumption and system latency. The proposed system performs a task (sound acquisition, feature extraction, transmission and classification) with accuracy higher than 87.8%, latency ~1.136 sec, and consumes ~0.87 J of energy on the IoT node per task. The estimated system battery life is more than 422 days when performing a task per minute using a 3.7 V, 2000 mAh battery. The long-term battery life and the small footprint of the proposed system, make it ideal for cable-free discrete smart home installations and for wearable devices, while its offline availability reduces privacy concerns.