RaspiDRVoice assistant on a Raspberry Pi Zero 2 W

Custom wake word, Groq speech, Hermes agent, NeoPixel stage animations

Wake word «хэй пидор»

Custom openWakeWord model trained on a Mac from synthetic and live samples; ~0.65 false triggers per hour.

Full voice loop

Greeting, listening until a pause, Groq STT, streamed Hermes answer spoken chunk by chunk, follow-up without the wake word.

Pi Zero 2 W hardware

WM8960 audio HAT, UPS-Lite battery, NeoPixel ring and rotary encoder — pins, quirks and fixes.

RaspiDR documentation

Voice assistant on a Raspberry Pi Zero 2 W: wake word «хэй пидор» ("hey peedor") → greeting → listen → Groq STT → Hermes agent → Groq TTS, with every stage shown on the NeoPixel ring.

Demo

A video of the speaker in action and photos of the device.

Architecture

How the system works end to end.

Components — Pi, Hermes server, Groq / xAI, Mac for development
Voice loop — states, transitions, ring animations and sounds per stage
- Response pipeline — STT → Hermes stream → chunking → TTS → player
Modules (src/)
Audio on the Pi — dmix/dsnoop, PulseAudio, saved mixer state
Configuration — .env keys and assistant.py flags
Development and deployment
Measurements (2026-09-26) — CPU, RAM, latencies
Known issues and next steps

Power efficiency experiments

Ideas to measure before changing the architecture: audio capture, wake-word gating, GPIO, native components, codec and amplifier power management, with an ordered A/B test plan.