speech-to-speech

Why we included this project

If you want a voice assistant that runs without phoning home, this project hands you a working conversation server instead of a pile of speech models to integrate yourself. It chains voice activity detection, speech-to-text, an LLM, and text-to-speech through queues, and every stage is swappable from the command line, so you can start with hosted inference and later move the language model onto a local vLLM or llama.cpp box while leaving the rest alone. Since it exposes an OpenAI Realtime-compatible WebSocket API, an app that already talks to that endpoint can be pointed at this server almost unchanged. The fact that it powers real consumer robots suggests the pipeline holds up under genuine turn-taking and latency pressure, not just scripted demos.