Run
FAQ
Direct answers to the questions that come up in nearly every evaluation.
How fast does it respond? Around two seconds voice-to-voice with the GPU co-located with the user. Distance and language model choice are the two things that move that number.
Which language model can we use? Any that speaks HTTP. OpenAI, Anthropic, Google, Mistral, your own self-hosted model, or ours. Switching is a configuration change.
Can it run entirely on our infrastructure? Yes — the on-prem and kiosk shapes exist for exactly that. With a self-hosted language model, no conversation content leaves your network. Fully air-gapped operation, with no periodic licence check at all, is not supported yet.
What happens if the internet goes down at a kiosk? Session state lives on the local network, so a LAN deployment keeps working for locally served conversations, and the screen has a defined fallback. If your language model is a hosted API, that part obviously stops.
How many languages? Over 50, both directions. Arabic, English, Russian, Mandarin, Spanish, French, Hindi, Turkish and more. For kiosks we usually pin the expected languages, which improves recognition accuracy.
Can the avatar be a real person from our organisation? Yes, with their documented consent. A short controlled recording session produces a 2D photoreal avatar of that person.
Can we run it on a phone? A 3D Unreal Engine avatar can render on the device, including smartphones. 2D photoreal avatars render on a GPU and stream to the phone.
What GPU do we need? RTX 3060 as an absolute minimum for 2D rendering alone. RTX 5070 Ti Mobile is the recommendation when one machine has to run everything — recognition, retrieval, the model, speech, and the avatar. Details in Hardware requirements.
Can users interrupt it? Yes. Speech is full-duplex — start talking and the avatar stops and listens.
Does it work in a noisy environment? With the right microphone, yes; that is what the hardware recommendation is for. Noise handling and echo suppression run server-side, but no amount of software rescues a badly placed microphone.
What does it cost? Quoted per deployment shape rather than per seat, because a single kiosk and a fleet across thirty service desks are not the same engagement. See pricing.
How long until we are live? Days on hosted SaaS. Two to three weeks for on-prem. Three to four for kiosks, including hardware. See Getting started.
Can we try it before committing? Yes — the live demo needs no signup and runs the same stack as a production deployment.
Something is not answered here. Talk to us. Questions from evaluations end up on this page.