<![CDATA[Visual acuity (VA) testing is a fundamental procedure for evaluating visual function. Conventional VA examinations require an examiner to interpret patients' responses, which may limit automation and scalability in repeated or remote testing scenarios. This study develops an embedded speech-based optotype recognition system to support digital visual acuity testing workflow. Mel-Frequency Cepstral Coefficients (MFCCs) were used to represent spoken responses, and a Convolutional Neural Network (CNN) was trained to classify nine Snellen optotype letters. Voice recordings were collected from 20 participants, with data augmentation including noise injection, pitch shifting, time stretching, and random gain adjustment to increase acoustic variability. Under the evaluated experimental protocol, the proposed MFCC-CNN model achieved 97.42% accuracy, 94.89% precision, 94.33% recall, and an F1-score of 94.61%, outperforming the evaluated kNN, Random Forest, and SVM baselines. The model was deployed on an NVIDIA Jetson Orin Nano, occupying approximately 8.3 MB and achieving an inference latency of approximately 30–40 ms per prediction. These results demonstrate the technical feasibility of embedded spoken-optotype recognition for supporting visual acuity testing workflows. However, the present study evaluates speech-recognition performance rather than clinical agreement between automatically generated VA scores and clinician-administered measurements. Formal clinical validation and stricter leakage-safe evaluation are therefore required before clinical applicability can be established.]]>
Copyrights © 2026
