Overview
What an avatar is here, and what it costs to put one on a call.
An avatar is a face an agent wears on a call. In avatar mode the worker
publishes a rendered video track alongside its audio, and the person on the
other end sees somebody speaking.
Two ways to get one#
- Your own pipeline. Enrol a recording, LiveGrid renders it on a GPU box you
run.
providerislivegrid. Enrolling a face - A vendor. Ten are supported; the face lives on their side and the worker streams audio to them. Providers and pricing
Both are addressed the same way from a call: an avatar_id passed to
start_call.
The states an avatar moves through#
| Status | Meaning |
|---|---|
pending | the row exists and nothing has been uploaded yet |
training | the recording is confirmed and queued for enrolment |
ready | there is an artifact; the avatar is callable |
failed | enrolment did not work, and status_reason says why in words the customer can act on |
revoked | withdrawn; the likeness is deleted and the avatar can never be called again |
failed carries its reason all the way through to the customer rather than a
generic message, because every realistic failure here is something they can fix —
the face left frame, the take was too short, there were two people in it. A bare
"enrolment failed" turns a two-minute reshoot into a support ticket.
Consent is not a checkbox#
A likeness is a person's face, and in most of the world using one commercially without permission is not a product decision. Consent here is a row, written in the same transaction as the avatar, naming who granted it — and it is queried live on every call path, so a withdrawn consent takes effect the moment it is withdrawn. Consent
What it costs per minute#
Between about $0.04 and $0.35 depending on the vendor, on top of your model and transport costs. Your own pipeline trades that per-minute charge for a GPU box you keep running. The table, with sources and caveats, is on Providers and pricing.