EdgeAI @ Burning Man at Google Dev Day

A 10-minute talk for Google Dev Day about the MindControl Helmet: what it takes to run a voice assistant with no cloud at all, on hardware you can wear in the desert.

I didn’t go to Burning Man myself. The helmet was built for someone else to wear on the playa.

The short version

• The desert has no reliable internet, so the AI has to live on the helmet.
• A Raspberry Pi 5 runs the whole pipeline: wake word, speech to text, LLM, text to speech.
• Qwen2.5-Instruct 1.5B at about 7 tokens/s is fast enough to hold a conversation. Bigger models were too slow; a reasoning model spent too long thinking out loud.
• The Hailo-10H NPU is not faster than the Pi 5 CPU at LLMs. Its job is to free the CPU for everything else the helmet does.

Cloud first, then local

The first prototype ran on a laptop and compared three setups: a classic chain (faster-whisper, the Claude API, ElevenLabs), Gemini Live doing audio in and audio out with a single model, and a local Chatterbox voice clone. The cloud setups sounded good but needed internet. The local clone took 10 to 30 seconds per reply. Neither works on the playa.

The local stack

Step	Tool	Runs on
Wake word	openWakeWord, “hey mycroft”	CPU
Speech to text	faster-whisper base.en, int8	CPU
LLM	Qwen2.5-Instruct 1.5B on hailo-ollama	Hailo-10H NPU
Text to speech	Piper, lessac-medium	CPU

Text to speech is nearly free (about 0.03 s of compute per second of audio), so the LLM is the bottleneck. Streaming the reply one sentence at a time means the helmet starts talking before the answer is finished.

NPU vs. CPU

Public benchmarks show the Pi 5 CPU beating the Hailo-10H on every model hailo-ollama offers, for example 11.7 vs. 6.7 tokens/s on Qwen2.5 1.5B (CNX Software). The NPU uses less power and leaves the CPU idle, which matters on a helmet whose CPU also handles speech, the camera stream, the phone portal and the LED display.

Google’s own Gemma Translator makes the same point from the other side: Gemma 4 E2B on LiteRT-LM runs a full offline voice pipeline on a CPU-only Pi 5 at 7.6 to 9 tokens/s. One AI job fits on the CPU; the helmet runs several.

What broke

• hailo-ollama owns the NPU for its whole lifetime, and a second NPU process crashed it. Speech to text moved to the CPU.
• An onnxruntime bug in an optional voice filter crashed every audio frame. The fix was turning the filter off.
• Background loops died silently while the process stayed up, so systemd never restarted them.
• The wake word is “MY-kroft”, not “Microsoft”. That one took a while.

Bonus: the LED display

The 48×12 LED matrix on the back of the helmet came with a phone app. A Bluetooth snoop log, a decompiled React Native app and a JieLi SDK later, it turned out the display accepts commands from anyone in Bluetooth range, no authentication. A small Python library now drives it directly, including live video at 7.7 fps.

Takeaways

• Choose the model for latency, not size.
• Give the LLM its own chip when the CPU has other work to do.
• Stream replies in sentences so speech starts early.
• Plan for silent failures: guard every background loop and let systemd retry forever.

github.com/electronick-co