WHERE THIS TEST ENDED
The wake word and a spoken reply work. The intended speaker route is still unresolved in this episode.
- Wake word
- Question processing
- Generated speech
- Selected speaker
Follow the voice through the system
I test the wake word, ask simple questions, and listen for where the answer comes from. The Atom Echo is the small listening device. The projector is the intended speaker for the desk experience.
A projector test beep succeeds, but Fletcher still answers from the Echo at the end. That distinguishes a working HDMI audio output from a complete assistant configuration using that output.
Give tracking and language processing room to run
I also describe the Pi working harder when hand tracking runs alongside the assistant. The corresponding build notes record contention between the tracker and local language-model work.
The software work separates their CPU allocation and limits the language-model thread count. This is about making simultaneous jobs fit on the machine; installing an accelerator or SSD alone does not demonstrate that every part of the voice path is accelerated.
What the next test needs to prove
The next useful test is a real question with the microphone listening and the reply coming from the projector. The follow-up video returns to that route and then tests the on-screen assistant.
The episode is useful because the failure is specific: the system can wake and speak, but it is speaking through the wrong device for the intended desk setup.
About this breakdown
Matched to this original TikTok, including its spoken captions. Technical context also comes from the corresponding Fletcher build notes. The result above describes this episode; later work is identified separately.
Breakdown checked September 6, 2026.