Fitri Utaminingrum, Aulia Riza Mufita, Aldiansyah Satrio Kabisat, I. Komang Somawirata
This study introduces a voice-activated smart wheelchair system engineered to support individuals with physical disabilities, especially in noisy environments. The proposed system employs GFCC for noise-resistant feature extraction and utilizes a ResNet50 deep learning architecture for command classification, implemented on an NVIDIA Jetson TX2 embedded platform. The model is designed to accurately identify Indonesian vocal commands related to wheelchair movement directions. Experimental evaluation encompasses epoch-wise performance analysis, confusion matrix evaluation, computational time measurement, and comprehensive testing within real-world environments under both calm and noisy conditions. The best model was found at epoch 72, when the validation accuracy was 94.6%, the validation loss was 0.221, and the macro-averaged precision, recall, and F1-score values were 0.955, 0.957, and 0.956, respectively. The average GFCC extraction and inference durations are 0.089 and 0.578 s, respectively, culminating in a total system latency of 0.667 s, thereby meeting real-time control specifications. Integrated testing shows that the proposed system works 88% of the time in quiet settings and 73.33% of the time in noisy ones. These findings demonstrate that the proposed GFCC–ResNet50 framework exhibits robust noise resistance and dependable real-time performance, rendering it appropriate for practical assistive mobility applications. © by the authors
Department of Informatics Engineering, Brawijaya University, Indonesia; Department of Electrical Engineering, National Institute of Technology, Indonesia
