This paper introduces a domain-adversarial Gaussian Mixture Variational Autoencoder (DA-MARTA) model developed to differentiate between Parkinsonian and healthy speech by generating knowledge-guided embeddings based on manner classes. The model constructs an interpretable latent space that effectively clusters phonemic groups, making it well-suited for both multidialect and cross-dialect contexts. By leveraging domain-adversarial training, DA-MARTA is able to overcome domain shift challenges inherent in multidialect speech analysis, thereby enhancing generalization across diverse datasets. Experiments were conducted on three distinct speech corpora: Albayzin [1], NeuroVoz [2], and GITA [3]. In the multidialect scenario, where all datasets were combined, DA-MARTA achieved AUC scores of 85 on running speech recordings from NeuroVoz and 81 on those from GITA, demonstrating its superior capability in exploiting shared phonemic features across dialects. In a cross-dialect evaluation, AUC values of 79 on NeuroVoz and 73 on GITA were obtained, highlighting the robustness of the domain-adversarial approach in generalizing to new dialectal variations. Moreover, the model’s practical applicability was further validated under scenarios where manual transcriptions were not available. When operating in a cross-dialect setting without manual transcriptions, DA-MARTA achieved AUC scores of 74 for running speech in NeuroVoz and 70 in GITA. This minimal performance loss underscores the feasibility of deploying the system in real-world contexts, where high-quality transcriptions may be impractical.