Vui
A real-time voice assistant built around a small, open-weight text-to-speech model that keeps the whole spoken conversation, including the caller's own audio, in context so replies carry natural prosody across turns.
Vui is a real-time voice assistant: it transcribes speech from a microphone, runs a local language model to decide a reply, and streams a synthesized voice response back, all from a single Python server. Its text-to-speech component, Vui Nano, has 219 million active parameters, and generates each reply from inside the ongoing dialogue rather than as an isolated utterance: the conversation so far, including the actual audio of the caller's turn, stays in the model's cache across roughly six minutes of context, so replies carry breaths, hesitations and prosody that carry across turns.
Vui Nano was trained on two-speaker dialogue with an explicit speaker-change token, runs on a CPU through a dependency-free C build with no Python or PyTorch at runtime, and also ships as a standalone model usable on its own through a Gradio demo.
It is licensed under Apache 2.0; the authors offer a separate paid production API with newer models and a more advanced turn-taking system.