TL;DR
A modular pipeline framework for building low-latency voice agents locally using open-source models.
Key features
Modular Pipeline: Runs VAD, STT, LLM, and TTS in separate threads, with every component being fully swappable.
OpenAI Realtime-compatible API: Provides a WebSocket API that works directly with OpenAI Realtime clients.
Full Local Execution: Supports fully local stacks using open-source models like Parakeet TDT (STT), Qwen3-TTS (TTS), and local LLM serving via llama.cpp.
Flexible LLM Backends: Uses OpenAI-compatible protocols to point to hosted providers, Hugging Face Inference Providers, or self-hosted vLLM/llama.cpp servers.
When to use it
When building voice assistants or conversational agents locally while maintaining data privacy.
When implementing a custom voice interface compatible with the OpenAI Realtime API.
When integrating voice conversation capabilities into existing robots (e.g., Reachy Mini) or IoT devices.