TEN Framework is an open-source framework for building conversational voice AI agents that talk, listen and respond in real time. Developed around the TEN Agent reference implementation, it lets you assemble low-latency voice pipelines from interchangeable parts and deploy agents that hold natural spoken conversations. The project has earned about 11,200 stars on GitHub and was still actively developed in 2026.
Instead of locking you into one model vendor, TEN treats speech recognition, language models, text-to-speech and turn detection as pluggable extensions that communicate over a real-time graph. Extensions can be written in Python, Go, TypeScript or C++, and the runtime handles streaming audio and video with WebRTC-grade latency. Reference agents demonstrate integrations with services such as the OpenAI Realtime API, Azure Speech, Deepgram, ElevenLabs and many open-source models, and the same graph can drive avatars, transcription bots and multimodal assistants that combine voice, video and text. The project ships with examples, a hosted playground and thorough documentation, so you can hear a working agent within minutes of cloning the repository. Because every pipeline stage is swappable, teams typically prototype with cloud speech services and later substitute self-hosted models to cut cost and latency as usage grows.
TEN Framework is free and open source. You self-host the runtime and only pay for any third-party speech or model APIs you plug into your agent graph.
TEN Framework is for developers building voice-first products — call agents, talking avatars, meeting assistants and realtime translators — who want an open framework instead of a closed voice API platform.