Hugging Face's speech-to-speech pipeline now enables Reachy Mini to run conversations fully locally without cloud dependency. Amir Mahla and Andres Marafioti detail a cascaded VAD, STT, LLM, and TTS architecture that serves a Realtime API-compatible WebSocket endpoint. Users can serve LLMs via llama.cpp or vLLM, with opinionated defaults including Parakeet-TDT for speech recognition and Qwen3TTS for synthesis. The setup allows component swapping for customization across latency and quality trade-offs.
No score is assigned. Sources and their independence are shown in the citation chain below.