OpenAvatarChat
A modular digital human chat system: speech recognition, the LLM, text-to-speech and the avatar can each be swapped.
Does it fit your team?
- Consider it if
- You want a conversational avatar where speech recognition, the LLM, text-to-speech and the avatar can each be swapped, and you accept that the default preset sends the LLM and text-to-speech to Alibaba Cloud Bailian.
- Look elsewhere if
- You need many conversations at once: the docs disagree on whether LiteAvatar supports concurrency. Also hold off if you need a settled license for every model before going commercial: the license owner has not yet confirmed whether SenseVoice may be used commercially.
- Can you use it commercially?
- Yes, with two models to watch
The project itself is Apache-2.0. Of the models it uses, LiteAvatar is MIT, LAM and CosyVoice are Apache-2.0, and MuseTalk has MIT code with a README stating that the trained model may be used commercially. Two need care. SenseVoice, used for speech recognition, ships its weights under the FunASR model license, which requires attribution and keeping the model name. Its section 3 (Responsibility and Risk) says the model is “provided for reference and learning purposes only” and does not say whether that limits commercial use. An issue asks exactly that: a maintainer answered that section 3 is a disclaimer and commercial use is permitted, and in September 2026 that answer was marked as not yet the license owner’s final confirmation. The issue was still open on 2026-10-13. The FlashHead model archive includes one component under Open-RAIL-M; with the Pro variant alone, the whole workflow is Apache-2.0.
LicenseMuseTalk license notesFunASR model licenseSenseVoice license discussionFlashHead component licenses
- Is it still maintained?
- Active
Last push on 2026-07-31. The repository is not archived.
- How hard is it to deploy?
- Moderate
It needs Python 3.11 and an NVIDIA driver that supports CUDA 12.8. With uv it is three steps: install dependencies, download models, start. Docker also works once the NVIDIA Container Toolkit is installed; the app is then served at https://localhost:8282.
- What does it cost?
- A GPU, or pay per API call
The software is free. The docs say the 2D avatar is light on hardware and that avatar inference reaches 30 FPS on an i9-13980HX CPU. Moving speech recognition, the LLM and text-to-speech to cloud APIs lowers the hardware needs a lot, in exchange for paying per call. The docs give no figure for GPU memory.
- What else you need
- An Alibaba Cloud Bailian API key, or a local model
The default preset uses Alibaba Cloud Bailian for the LLM and text-to-speech, which needs a DASHSCOPE_API_KEY; speech recognition runs locally with SenseVoice. If you would rather stay off the cloud, the docs say local multimodal models are supported too. For remote access, the bundled coturn can be started alongside it.
- What to watch out for
- The docs disagree on concurrency
The release notes in the README say LiteAvatar supports multiple sessions on one machine, while the FAQ says it does not support concurrency; test it to find out. The stated average response time is 2.2 seconds. The FAQ also notes that personal AutoDL accounts cannot open custom ports, so remote access does not work there.
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.