LiveTalking
A real-time streaming digital human engine: feed it text or audio and the avatar lip-syncs, streamed to a browser or a live platform.
Does it fit your team?
- Consider it if
- You have an RTX 3060 or better and someone who can set up Python and CUDA, and you want an avatar that lip-syncs to text or audio and streams to a browser or a live platform.
- Look elsewhere if
- You need spoken conversation with wake-word interruption out of the box: that is in the paid commercial edition, priced on request. Also skip it if your videos on Bilibili, WeChat Channels or Douyin cannot carry the LiveTalking watermark and logo.
- Can you use it commercially?
- Yes, with a watermark on videos
Apache-2.0, so commercial use and modification are allowed. The README adds a statement: videos made with it and published on platforms such as Bilibili, WeChat Channels or Douyin must carry the LiveTalking watermark and logo.
- Is it still maintained?
- Active
Last push on 2026-10-09. The repository is not archived.
- How hard is it to deploy?
- Moderate
Tested by the authors on Ubuntu 22.04, Python 3.12 and CUDA 12.8. You install dependencies, download models from a cloud drive, and start it with one command. The official FAQ lists twelve install errors, mostly dependency version conflicts. A Windows bundle and ready-made cloud images also exist.
- What does it cost?
- Needs a GPU
The software is free; the GPU is the cost. The authors recommend an RTX 3060 or better for the wav2lip model and an RTX 3080 Ti or better for musetalk. How many avatars can speak at once depends on the GPU; idle sessions depend on the CPU.
- What else you need
- A GPU and text-to-speech; conversation also needs an LLM and speech recognition
A GPU is required. For speech you can use the free EdgeTTS, host GPT-SoVITS or CosyVoice yourself, or call Tencent Cloud (needs an API key). For it to hold a conversation, add an LLM: Qwen and OpenAI-compatible endpoints are supported (needs an API key). For spoken questions you also need a FunASR recognition service, or the browser’s built-in recognition. The server must expose TCP port 8010 and UDP ports.
- What to watch out for
- Voice input needs your own recognition service
The open edition ships a voice-input page, but recognition is on you: run a FunASR service, or use the browser’s built-in speech recognition. Wake-word interruption, full-duplex conversation and ready-made cloud speech recognition are in the paid commercial edition; the docs give no price and say to contact the author.
Usage guide, voice input (Chinese)Commercial edition page (Chinese)
Last checked: .
If your priority is different
Your industry is not here?
Tell me the industry and what you need solved. I will look at it next.