FAQ
Getting it running
The page says the microphone is unavailable. Browsers only expose the microphone on a secure origin. Use the HTTPS port (8443 / 8455), accept the self-signed certificate or install a real one (Deploy → TLS). http://127.0.0.1 also counts as secure for local testing.
"engine busy" / "数字人正忙". One avatar worker serves one session. Either wait, add workers (本地服务 → engine pool), or raise capacity with a bigger GPU. The playground competes with visitors for the same workers.
Where is the admin password? Printed once in the cored log when the database is first initialized: docker compose logs cored | grep "initial admin". You must change it on first login.
docker compose up aborts naming a variable. Filesystem paths in .env have no defaults on purpose. Set the named variable to a real directory on your host.
GPU not visible in the container. Enable CDI: sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml, re-run after adding or removing GPUs.
EdgeTTS says nothing / errors. It needs outbound internet to Microsoft. Air-gapped? Use Qwen3-TTS locally.
Model downloads are slow or blocked. Defaults use China-friendly mirrors (HF_ENDPOINT=https://hf-mirror.com, ModelScope). Outside China set HF_ENDPOINT=https://huggingface.co.
Using it
How many people can talk to it at once? As many workers as fit in VRAM: about one per 2 GB with wav2lipLS, one per 7–8 GB with MuseTalk or FlashHead, plus ASR. A 24 GB card runs 2–3 concurrent quality sessions. The console does the arithmetic and enforces channel caps.
Can it run with no GPU? cored and the console can. The avatar worker needs a GPU somewhere; rent one by the hour and point AVATAR_ADDR at it (Quickstart, path C).
How do I give it my company's knowledge? Upload .txt / .md to a knowledge base and enable RAG. It answers from your documents; it can still be wrong, so test with 检索测试 and keep the system prompt honest about limits.
Does it work on phones? Yes, in the mobile browser, no app. Microphone permission is asked on first use.
Can I put it in my own website? Yes: iframe, SDK, or custom element. See Embed.
Can it make gestures? Avatars can carry baked one-shot clips (wave today). Visitors get a button, the SDK and agents get an API.
Can I interrupt it? Yes. Start talking; the turn detector notices and the avatar stops. From code, interrupt().
Licensing and business
Is it really free for commercial use? Apache-2.0 for the code. Default engines are chosen for commercially usable licenses (FlashHead Apache-2.0, MuseTalk MIT, self-trained wav2lipLS; InsightFace replaced by MediaPipe). Non-commercial components were removed. See LICENSING.md.
Is there a watermark or a usage phone-home? No.
Can I train an avatar of a real person? With their consent. The console asks for it; the policy is in Compliance.
How does the project make money? Deployment, avatar production, knowledge base integration and support contracts. The code stays open; if the maintainers vanished tomorrow you would still have everything.
How does this relate to LiveTalking? LiveTalking is an excellent engine framework; Mynah's author worked with its author on OEM projects. Mynah open-sources the layer that usually sits above the engine: console, channels, knowledge base, concurrency, engine pool. Different layer, no conflict.