Paper Title
A Lightweight Real-Time Lip-Reading System for Visual Speech Transcription And Translation

Abstract
Visual speech recognition, commonly known as lip reading, has gained increasing attention with recent advances in deep learning. Nevertheless, many existing approaches remain computationally demanding and are mostly evaluated in offline settings, which limits their applicability in real-world scenarios involving pose variation, illumination changes, and embedded hardware constraints. This study presents a prototype system for real-time English lip-reading transcription with an optional machine-translation interface. The proposed pipeline consists of mouth region extraction using MediaPipe Face Landmarker, fixed-length temporal windowing, a lightweight spatiotemporal recognition model based on an R(2+1)D architecture, and inference-level stabilization mechanisms including confidence gating, temporal buffering, and speech-boundary detection. The system is designed with quantization-friendly settings to support future deployment on low-cost edge devices such as the Raspberry Pi 5. Experiments were conducted using BBC-derived lip-reading datasets containing variations in speakers, lighting conditions, and camera viewpoints. The proposed model achieved 92.7% recognition accuracy under limited-vocabulary offline evaluation. For real-time operation, the system prioritizes stable and temporally aligned transcription rather than frequent low-confidence predictions. The results indicate that the proposed approach provides a practical bridge between dataset-trained lip-reading models and real-time transcription/translation applications, while highlighting key engineering considerations for deployment outside controlled laboratory environments. Keywords - Deep learning, edge AI, lip reading, multimodal translation, Raspberry Pi, real-time inference, TensorFlow Lite, visual speech recognition.