Parlor

Parlor is an early research-preview project for fully on-device, real-time multimodal voice interaction, built to approach the experience of OpenAI's GPT-Live while running entirely on a local machine such as an Apple Silicon MacBook. After an attempt to fine-tune a model directly into a full-duplex voice model did not succeed, it uses a classic cascade of separate speech, language and speech-synthesis stages instead.

Turn-taking is hands-free: browser-side voice activity detection with a short silence cutoff decides when audio has stopped, and a separate turn-detection model judges whether the user has actually finished a thought rather than paused mid-sentence, so it can wait through pauses instead of interrupting. Replies stream as they are generated: a transcript of what was heard appears immediately, and the response is spoken sentence by sentence while later sentences are still being produced. An optional background-research mode can hand a query to a cloud-based frontier model with web search while the local conversation continues, but stays off, keeping everything on-device, unless a corresponding API key is configured.

It is written in Python, released under the Apache-2.0 license, and documented as an early experiment with expected rough edges.