Dementia is a progressive neurodegenerative disorder that affects speech, language, memory, and cognitive function. Early diagnosis is essential for timely clinical intervention; however, conventional diagnostic procedures are often expensive and require specialized medical infrastructure. This work presents a multimodal deep learning framework for speech-based dementia detection using the DementiaBank Pitt Corpus. The proposed approach integrates complementary acoustic, behavioral, and linguistic speech representations extracted from Cookie Theft, Verbal Fluency, and Sentence Repetition tasks. Acoustic features are encoded using a Conformer network, linguistic information is represented using BERT embeddings, and behavioral speech characteristics such as pause statistics, speech rate, filler words, and repetitions are incorporated to capture cognitive impairment. Cross-modal attention is employed to fuse the different modalities into a unified representation, while prototypical few-shot learning and contrastive learning improve robustness under limited training data. Experimental evaluation demonstrates an accuracy of 85.22%, an F1-score of 90.17%, and an AUC of 91.19%, outperforming several ablation configurations. The proposed multimodal framework provides an effective, scalable, and non-invasive approach for automated dementia screening and highlights the potential of speech-based artificial intelligence for early cognitive assessment.