Parlor lets users converse with an AI that processes speech and video entirely on their own hardware. It stitches together Google’s Gemma 4 E2B for multimodal understanding and Hexgrad’s Kokoro for text‑to‑speech, routing audio and JPEG frames through a FastAPI WebSocket bridge. The demo runs on a Mac M3 Pro or any Linux box with a supported GPU, requiring about 3 GB RAM.

Creator fikrikarim built Parlor to keep a free English‑learning voice assistant off cloud servers, eliminating recurring costs. Earlier prototypes needed an RTX 5090 for real‑time voice, but the new model delivers sub‑three‑second end‑to‑end latency on Apple silicon: 1.8‑2.2 s for speech‑vision inference, 0.3 s for response generation, and another 0.3‑0.7 s for TTS. Users can interrupt mid‑sentence thanks to browser‑side VAD and smooth interaction overall.

Parlor’s open‑source repo pulls models automatically from HuggingFace, totaling roughly 2.6 GB for Gemma plus TTS assets, and ships with a lightweight HTML front end that handles voice activity detection and camera feed. By demonstrating real‑time multimodal AI without any cloud dependency, the project offers a blueprint for affordable, privacy‑first language tools on consumer hardware that can run locally on laptops.