Briefing
The Pi wakes on a custom "Hey IRIS" phrase detected on the device, sends the audio to faster-whisper on the workstation and routes commands such as stop and volume past the language model. Replies stream sentence by sentence into Kokoro speech; the README gives about 1.5 seconds to first audio, against about 23 seconds for a full reply. An emotion tag on each reply drives the eyes, mouth and LEDs at once.
The author checks every change on the physical robot himself; the README says an AI cannot tell whether the eyes track or the voice sounds right. One bug surfaced only because IRIS kept losing a hand-reading game: the camera had never focused, because capture skipped the autofocus cycle. The wake-word model carries a non-commercial CC BY-NC-SA restriction from its training data.
Written by the Boards & Bots curator
“the AIs plan, implement, and verify at the code level; I own physical reality.”
From the GitHub API
Repository facts
| Repository | maestro8484/ |
|---|---|
| Stars | 2 |
| License | AGPL-3.0-or-later Copyleft |
| Language | C++ |
| Last push | 27 Jul 2026 3 mo ago |