Nari Labs, a startup founded by two Korean undergraduates with no formal AI background, has launched Dia, a new open-source AI model capable of generating synthetic podcast-style conversations. Inspired by Google’s NotebookLM, Dia offers users greater control over dialogue tone, speech disfluencies, and nonverbal cues like coughs and laughs.
Co-founder Toby Kim said Dia was trained using Google’s TPU Research Cloud and weighs in at 1.6 billion parameters. Available on Hugging Face and GitHub, Dia runs on consumer PCs and supports random or custom voice generation, including cloning. Kim said Nari’s goal is to build a socially-driven voice AI platform, while acknowledging the risks of misuse and pledging to publish a technical report soon.
