quillman

Why we included this project

QuiLLMan ships a complete voice chat app rather than a bare model wrapper, which makes it a genuinely useful reference for anyone building conversational speech interfaces. On the backend it runs Kyutai Lab's Moshi speech-to-speech model alongside the Mimi streaming encoder, and bidirectional websocket streaming with Opus compression keeps audio round-trips short enough that replies can approach real conversational cadence instead of the pause-then-answer pattern typical of voice assistants. The code splits neatly into a React frontend served through FastAPI and a separate Moshi websocket inference module, both deployed as serverless Modal apps that scale to zero when idle. A developer or small team can lift it as a starting point for a low-latency voice bot, study how the backend manages streaming audio state, or run the dev server to test the websocket pipeline before adapting it. Treat it as an illustration and playground rather than a turnkey product: plan to build on top of it instead of deploying it as-is.