HuggingFace's Modular Voice Agent Pipeline Runs Locally, OpenAI Realtime-Compatible

solyarisoftware · x · 2026-08-09

HuggingFace introduces a speech-to-speech voice agent pipeline that runs entirely on open-source models and is fully modular. Each component is swappable, and stages run in separate threads connected by queues. It exposes an OpenAI Realtime-compatible WebSocket API, allowing existing clients to switch by changing one URL. The pipeline includes VAD (Silero VAD v5), speech-to-text with live partial transcripts, LLM response generation, and text-to-speech synthesis. Multiple interchangeable backends are available via CLI flags, and a fully local setup can use llama.cpp with models like Gemma 4.

Original post →

More from coding & agent

coding & agent channel →