OpenAI's Bidi 1 full-duplex voice model now testable in ChatGPT via Chrome sideload

OpenAI's experimental full-duplex voice model, internally labeled "Bidi 1," is now accessible in ChatGPT through a community-built Chrome extension. The model can listen to a user's speech while simultaneously generating its own audio output—a technical capability that enables the assistant to interrupt mid-sentence, much like a human conversation partner. Installation requires enabling Developer mode in chrome://extensions/, loading the unpacked extension folder, then navigating to chatgpt.com and clicking "use voice" in a new chat. The extension works on Chrome and Chromium-based browsers including Arc. Users can test the interruption capability by explicitly asking the assistant to cut them off while speaking.

Full-duplex audio is a long-standing challenge in conversational AI; most voice assistants operate in half-duplex mode, waiting for silence before responding. The feature remains experimental and unofficial—OpenAI has not publicly announced Bidi 1, and access depends entirely on sideloading community tooling rather than a native rollout. The voice settings menu includes a "model" dropdown listing Bidi 1 alongside an "intellect" slider, though the slider's functional impact is unclear. The extension's circulation suggests OpenAI has solved enough of the technical stack—low-latency streaming inference and echo cancellation—to begin limited testing.