MIUPOD
A runtime for AI companions that decide when to speak.
Overview
A MIU (Multimodal Interaction Unit) is an AI companion for a stream or a voice chat. It listens to the streamer and the chat, follows the game, and answers out loud. Most of the work goes into deciding whether to answer at all.
Decision first
A separate learned model decides whether and when to talk. The reply is written only after that.
Silent by default
If the model is unsure, or any step fails, the MIU says nothing.
No canned phrases
Each reply is written for the moment it answers.
Emotion in the voice
The voice model gets an emotion with each line and speaks it that way.
Architecture
Each stage does one job, from left to right. The decision is a separate block from the model that writes the reply. Next to it, the pacing and safety policy can cancel a reply.
Anatomy of a moment
The same stages, drawn as modules on one rail. An event travels along the bus. Most events stop at the decision unit, and a few go on to the voice. Hover over a module to read what it does.
The life of one moment
An event goes through seven steps before the MIU says anything, and any step can end in silence. The MIU updates its memory and mood either way.
Deciding
The decision has three outcomes: reply, wait or stay silent. The MIU can also speak first when something happens in the game or on the stream. A wait is short and ends in a new decision. If nothing happens, the MIU stays quiet.
In numbers
The numbers are rounded. Exact limits and timings are private.
Where it stands
It runs in the owner's own testing. A public release with a subscription is planned.
Works today
- Reactive conversation
- The decision gate
- Reactions to the game and the stream
- Memory of people and talk
- Plugins and the control panel
Next
- The next training round of the decision module
- A new local voice, in acceptance
- A public release by subscription, in credits