Speech naturalness is one of the most critical challenges in text-to-speech (TTS) systems, especially for low-resource languages such as Uzbek. While recent advances in deep learning have significantly improved the intelligibility of synthesized speech, achieving natural prosody—including appropriate intonation, rhythm, stress, and timing—remains a complex problem. This study focuses on improving speech naturalness in Uzbek TTS systems through deep learning-based prosody modeling. The paper analyzes existing approaches to prosody modeling, discusses the linguistic characteristics of the Uzbek language that affect prosodic patterns, and proposes the integration of neural network-based methods to capture expressive and natural speech features. The findings highlight the potential of deep learning architectures to enhance the quality and naturalness of Uzbek speech synthesis and contribute to the development of more human-like TTS systems.