RaspiDR

Voice assistant on a Raspberry Pi Zero 2 W: wake word ("hey peedor"; custom openWakeWord model) → greeting → utterance recording → Groq STT → Hermes LLM agent → Groq TTS, with each stage animated on a NeoPixel ring.

Documentation: flopsstuff.github.io/raspidr (sources in docs/) · Demo video: flopsstuff.github.io/raspidr/demo

docs/index.md	documentation contents — start here
docs/demo.md	demo video and photos of the device
docs/architecture.md	how the system works: loop, modules, audio, configuration, deployment, measurements
docs/hardware.md	the speaker hardware: pins, WM8960, UPS-Lite, ring, encoder, known issues
docs/wakeword_training.md	how the wake word model was trained and how to retrain it

src/            speaker code (deployed to the Pi): assistant.py is the entry point
src/tools/      hardware checks and utilities (hwtest, micmeter, micprobe, record_samples, make_sounds)
training/       wake word training scripts (datasets in training/data are downloaded/generated, not in git)
models/         hey_peedor.npz — wake word model (Git LFS)
sounds/         greetings, «секунду…» ("one sec…"), waiting sound (Git LFS)
docs/           documentation, published with VitePress
deploy.sh       deploy to the Pi; requirements.txt — Pi dependencies; .env.example — configuration template

Voice recordings and phrase samples (hey-peedor/, recordings/) are personal and gitignored — docs/wakeword_training.md explains how to generate and record your own.

Quick start

git lfs install && git clone [email protected]:Flopsstuff/raspidr.git && cd raspidr
cp .env.example .env                                       # fill in the Pi host, Hermes URL and API keys
./deploy.sh --install --restart                            # sync to the Pi, install deps, start the assistant
./deploy.sh --logs                                         # follow the log; say «хэй пидор»

python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
.venv/bin/python src/assistant.py --text "Привет, кто ты?"  # on the Mac, no microphone

License

Code and documentation: MIT.

The wake word model models/hey_peedor.npz is trained on openWakeWord's negative feature set, and openWakeWord distributes its pre-trained models under CC BY-NC-SA 4.0 because of its training data. Treat this model the same way: non-commercial use only. Retrain it yourself (see docs/wakeword_training.md) if you need different terms.