The rapid developments within generative AI (GenAI) present new opportunities for personalized tutoring, particularly for students with learning difficulties. However, accessibility challenges persist, as many GenAI models rely on text-based interactions. This study, grounded in social presence theory, examines the effectiveness of interactive AI voice-based learning tutoring in a grammar-learning task. Thirty-seven eighth-grade students, with and without learning difficulties, engaged with three interaction modes: human-tutor text chat, text-based GenAI chat, and voice-based GenAI chat. The results indicate that interaction mode significantly influenced task persistence, cognitive load, cognitive load variability, trust, and perceived confidentiality, with human-tutor chat rated highest. However, interaction mode did not predict task success or engagement. Unexpectedly, voice-based AI was the least effective, as students struggled due to the technology responsiveness. Our findings highlight the need for adaptive AI learning tools and perception of social support to enhance trust and learning outcomes of diverse students. Future research should refine AI adaptations for specific learning challenges.