Hugging Face's Local Voice-Agent Pipeline Gains 600+ Stars a Day

Hugging Face’s speech-to-speech project climbed GitHub trending with roughly 628 stars in a day, passing 9,400 total. It is a modular VAD → STT → LLM → TTS pipeline exposed behind an OpenAI Realtime-compatible WebSocket API, which means an existing realtime client can point at a locally hosted stack without code changes. Every stage is swappable: Silero VAD, Parakeet TDT or Whisper for transcription, any OpenAI-compatible endpoint or local model, and Qwen3-TTS, Kokoro or ChatTTS for synthesis. CUDA, CPU and Apple Silicon are all supported, under Apache 2.0. No latency figures are published.