An advanced semi-supervised fusion architecture leveraging speech embeddings has been developed to detect Parkinson's disease with high sensitivity. This AI system analyzes nuanced vocal features including pitch and rhythm, capturing early-stage dysarthria symptoms often imperceptible clinically. By utilizing a semi-supervised learning approach, the model efficiently leverages both limited labeled and abundant unlabeled speech data, enhancing diagnostic accuracy and generalizability across diverse populations. This technology promises a non-invasive, accessible tool for earlier Parkinson’s diagnosis, fostering improved management and treatment strategies.