Abstract
Air-written character recognition technology has placed itself as a promising alternative to traditional text input methods in wearable devices and assistive systems, as they allow surface-free writing through motion tracking. However, recognizing complex scripts remains challenging due to high inter-character similarity, variations in writing styles and sensor noise. Prior studies in this area mainly focused on the English, Latin, and Arabic languages, and most of them used deep learning approaches. Furthermore, limited studies have been done on Bangla literature, with very few works employing shallow learning models and not enough available data on sensor-based datasets. To address this gap, this study explores IMU-based recognition of the first six Bangla vowel characters using four shallow machine learning models. A dataset comprising 3,006 samples, having 501 samples per character, was collected using a finger-mounted IMU sensor capturing six motion channels: three gyroscope and three accelerometer axes, preprocessed through labeling and resizing to fixed length for consistency, extracted 8 time-domain statistical features: minimum, maximum, mean, median, standard deviation, RMS value, skewness, and kurtosis, and employed SVM, Random Forest, KNN, and GBM on them. The SVM model achieves the highest accuracy of 81.68% among Random Forest (80.53%), KNN (77.23%) and GBM (74.92%). These findings demonstrate the feasibility of shallow learning methods for Bangla air-written vowel recognition and they highlight their potential for developing lightweight, real-time and language-inclusive input systems in wearable and assistive technologies.
