Bengali Air Writing Using Wearable Wireless Motion Sensors

Abstract

Air-writing with body-worn inertial sensors is a promising, camera-free pathway to text input on constrained or screenless devices. Yet Bengalione of the worlds most widely used scriptsremains sparsely represented in public corpora, limiting repro ducibility andcomparativeprogress. Weintroduceamotion-sensorBengaliair-writing dataset comprising 3,996 single-character trials spanning 60 core symbol classes (vow els, consonants, digits), captured with a wrist/hand-mounted 6-DoF IMU at 200 Hz under a controlled 2×3×2 protocol varying handedness (dominant/non-dominant), writingspeed(slow/normal/fast),andenvironment(normal/noisy). Acue-basedready go–write–hold procedure, deterministic file naming, and self-describing CSVs with user/session metadata support transparent filtering, canonical splits, and faithful re analysis. We benchmark recognition using classical machine-learning pipelines built on com pact time/frequency and cross-axis features, alongside sequence models that consume minimally processed IMU streams. The baselines verify the feasibility of IMU-based Bengalicharacterrecognitionwhilesurfacingpersistentfine-grainedconfusionsamong kinematically similar graphemes, cross-condition generalization gaps (hand/speed/en vironment), and the difficulty of writer-independent performance without personal ization. We further outline evaluation tracks for streaming, online decoding (joint boundary detection and classification) to bridge from trial-bounded clips to real-time interfaces. As of our study, this is the second publicly described IMU-captured Bengali air-writing corpus with adocumented acquisitionprotocol andreproduciblebaselines. Byreleas ing data, canonical splits, and challenge definitions, weprovidearigorousfoundation for future work on rotation-/tempo-robust representations, few-shot personalization, and low-latency decodingadvancing inclusive, privacy-preserving text input for non Latin scripts