Lightweight Restricted Dialogue Visual Speech Recognition for Embedded Systems
Year: 2026
Type: conference-paper
Abstract: The rapid integration of Artificial Intelligence (AI) into industrial and commercial sectors has accelerated the development of sophisticated human-machine interfaces. Visual Speech Recognition (VSR), or machine lip reading, offers a robust alternative to traditional acoustic systems, particularly in environments with extreme acoustic noise or privacy constraints. However, the current trajectory of VSR research is dominated by deep learning architectures that require massive datasets and high-performance computing resources and are often referred to as “black box” models. For Thai, where largescale audiovisual corpora are unavailable, and for deployment on resource-constrained embedded devices, these heavy architectures pose a significant hurdle. This paper presents a novel methodology, Lip Geometry Analysis (LGA), which prioritizes a lightweight, deterministic framework suitable for deployment on low-power microcontrollers while addressing the specific linguistic challenges of low-resource languages such as Thai. Because the method operates on mouth geometry rather than language-specific phonology, it generalizes to other languages; we demonstrate this on a mixed Thai and Chinese food menu task with no cross-language confusion.
Institution: Assumption University
Language: English
Topic: Speech and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204467775
