Dia is an innovative open-weights text-to-speech (TTS) model designed to generate realistic conversations directly from transcripts. It aims to enhance the dialogue generation process by creating entire conversations in a single pass, ensuring a more natural and efficient output than traditional TTS systems that stitch together speaker turns. Developed with a focus on user experience, Dia supports audio prompts allowing conditioning on specific voices or emotions, which adds a layer of versatility to its application. The creators, Toby and Jay, embarked on this project due to their dissatisfaction with the repetitiveness of existing technologies, such as the podcast feature of NotebookLM. After about three months of intensive learning and development, they are excited to release a lightweight technical report to share their findings and encourage open-source contributions. The feedback from the community is generally positive, with users praising the model's performance and suggesting potential for more language support in the future. Main concerns include stability in various contexts, particularly in specialized terminology, and hardware requirements for running the model effectively.