<![CDATA[Abstrak. Pengenalan bahasa isyarat berbasis sensor wearable menjadi pendekatan yang semakin relevan karena ketahanannya terhadap variasi pencahayaan dan oklusi jari dibandingkan pendekatan berbasis kamera. Penelitian ini bertujuan untuk mengklasifikasikan 26 abjad Sistem Isyarat Bahasa Indonesia (SIBI) menggunakan model Multilayer Perceptron (MLP) yang diimplementasikan secara on-device melalui TensorFlow Lite (TFLite) pada platform Android. Data masukan berdimensi 33 fitur diperoleh dari sarung tangan pintar yang dilengkapi lima flex sensor dan satu unit IMU MPU-6050, dikirimkan melalui Bluetooth Low Energy (BLE) setiap 300ms. Pra-pemrosesan data menggunakan normalisasi Z-score (StandardScaler), dan prediksi distabilkan menggunakan mekanisme Majority Voting 10-frame. Evaluasi menggunakan metode K-Fold Cross Validation (K=5) menghasilkan akurasi rata-rata sebesar 97.31% pada skenario subject-specific. Pengujian integrasi sistem menunjukkan rata-rata latensi end-to-end sebesar 285ms dan packet loss BLE sebesar 0.26%, keduanya jauh di bawah ambang batas target. Hasil ini membuktikan bahwa arsitektur MLP berbasis data sensor mampu menjadi Proof of Concept yang andal untuk sistem penerjemah isyarat ringan, luring, dan berlatensi rendah di perangkat Android. Abstract. Sensor-based wearable sign language recognition has emerged as an increasingly relevant approach due to its robustness against lighting variations and finger occlusion compared to camera-based methods. This study aims to classify 26 alphabets of the Indonesian Sign Language System (SIBI) using a Multilayer Perceptron (MLP) model implemented on-device via TensorFlow Lite (TFLite) on the Android platform. The 33-dimensional input features are acquired from a smart glove equipped with five flex sensors and one IMU MPU-6050 unit, transmitted via Bluetooth Low Energy (BLE) every 300ms. Data preprocessing employs Z-score normalization (StandardScaler), and predictions are stabilized using a 10-frame Majority Voting mechanism. Evaluation using K-Fold Cross Validation (K=5) yields an average accuracy of 97.31% in a subject-specific scenario. System integration testing demonstrates an average end-to-end latency of 285ms and BLE packet loss of 0.26%, both well below target thresholds. These results demonstrate that an MLP architecture based on sensor data can serve as a reliable Proof of Concept for a lightweight, offline, low-latency sign language translation system on Android devices.]]>
