Jev 101
Start Guides
Search
中文 EN
Realtime outlineExperimental — needs a realtime voice model plus Jev; calibrate on your own sessions

Live avatar gesture select

Who: Builders of voice agents with an on-screen character. Steps: Send the live transcript fragment as state; ask Choice questions for face, gesture clip and reaction, plus a Noul for whether an action was promised; act only above a confidence gate and fall back to local rules. Expected effect: an avatar that reacts in about a tenth of a second without a second chat model.

Source: https://github.com/rehan-remade/open-annie ↗ Docs / source ↗
1

Treat this as an outline — adapt state and questions to your data.

2

Implement in code — call System One / Jev; compose answers yourself.

3

Gate on confidence — act, confirm, or escalate before side effects.

Outline sketch

Prompt 
# Live avatar gesture select
State: the latest transcript fragment and who is speaking.
Ask: Choice for face, gesture clip and reaction from installed assets only; Noul for whether the character committed to an action.
Then: play the clip only above a confidence gate; refusals veto motion; on timeout or error use local rules.

Needs access to: gpt-livethree.js

Who it's for

Builders of voice agents with an on-screen character

Steps / how it's set up

Steps (experimental)

  1. Stream the conversation transcript in short fragments.
  2. Ask Jev a few Choice / Noul questions per fragment, limited to the clips you have installed.
  3. Gate every answer on confidence in code, and fall back to local rules when Jev is slow or unavailable.

Sources

This pattern is experimental. Author-reported behavior — calibrate on your own sessions.

Expected effect

Fast, bounded motion choices for a talking character

Unofficial outline for learning. Paraphrased from public docs and tutorials — not a production recipe. Review sources before you automate anything.