FastRTC: The Real-Time Communication Library for Python

Source: HuggingFace Date: 2025-02-25 URL: https://huggingface.co/blog/fastrtc

Summary

Library release: FastRTC — HF’s Python library abstracting WebRTC for real-time audio/video AI applications. Key abstractions: ReplyOnPause for automatic voice turn-detection, built-in Gradio WebRTC UI for local testing, fastphone() for receiving calls on a generated phone number, one-line FastAPI deployment. Built-in integrations: Moonshine Base (STT), Kokoro (TTS), and LLM providers. Demo: full voice chat pipeline in ~20 lines (STT → Llama-3.2-3B on SambaNova → TTS streaming). No benchmark numbers.

Implications

HF as open-source ML hub. FastRTC abstracts WebRTC’s complexity into a Python-native API that non-specialist developers can use — the bottleneck for voice AI applications has been WebRTC’s C++ origins and protocol complexity, not the AI models themselves. Positioning Moonshine/Kokoro as built-in defaults makes FastRTC a complete voice stack from day one.

Open-weights ecosystem health. Real-time audio AI (Moshi, Qwen2-Audio, Ultravox) has emerged rapidly but deployment tooling lagged. FastRTC addresses the deployment gap directly — the fastphone() function for inbound phone calls is the most novel feature, enabling open-weights voice models to serve as phone agents without Twilio or similar telephony abstractions.