Say "clip that" — and the last 30 seconds of audio are saved and transcribed.

clip-thing is a small, always-listening recorder for meeting rooms and desks. It keeps a rolling 30-second buffer of audio in memory; when it hears the wake word it writes that buffer to a clip, transcribes it, and lists it in a self-hosted web UI you can play and search.

It runs two ways from one codebase:

• Laptop mode — pip install, run, uses your built-in mic. Zero hardware.
• Device mode — a Raspberry Pi 4/5 + ReSpeaker mic that sits in a room.

Everything is open source — the code, the hardware build guide, and the wiring schematics.

mic → rolling 30s buffer (in memory) → wake word ("clip that")
        → save WAV → transcribe → SQLite → web UI (play / read / search)

• Wake word: openWakeWord — open source, no API keys, runs on a Pi.
• Transcription: on-device by default (faster-whisper); an optional cloud backend is one config line away.
• Self-contained: each unit captures, transcribes, stores, and serves its own UI. No central server, no account.

clip-thing is an always-on microphone. Be transparent with anyone in the room.

• The rolling 30-second buffer lives in memory only and is continuously overwritten. Audio is written to disk only when the wake word fires.
• With the default local transcription backend, audio never leaves the device — capture and transcription are entirely on-device.
• The web UI has no authentication. It is built for a trusted LAN: anyone who can reach the address can play and read every clip. Do not expose it to the public internet. Set web.host: 127.0.0.1 to keep it local-only.
• Switching to the cloud transcription backend sends clip audio to a third-party API. That is opt-in.

Requires Python 3.11+.

git clone https://github.com/yoelgal/clip-thing.git
cd clip-thing
python -m venv .venv && source .venv/bin/activate
pip install -e ".[local]"

clip-thing run

Open http://localhost:8080. Say the wake word near your mic — within a few seconds a transcribed clip appears.

First run downloads the wake-word model (~a few MB) and, on first transcription, a Whisper model (~75 MB for tiny).

Full walkthrough: docs/setup-laptop.md.

Build a Raspberry Pi + ReSpeaker 2-Mic HAT unit and have it autostart on boot:

• Hardware build, bill of materials, and wiring: hardware/README.md
• Pi software setup and autostart: docs/setup-pi.md

openWakeWord ships no stock "clip that" model, so:

• Out of the box clip-thing uses a stock model (hey_jarvis) — the pipeline works immediately, with no training and no setup.
• For the actual "clip that" phrase, train a model once with clip-thing train-wakeword "clip that", then point wakeword.model_path at the resulting models/clip_that.onnx.
• Want a different phrase? Same command with your phrase — see docs/training-wakeword.md.

Copy config.example.yaml to config.yaml and edit. Every option is documented in docs/configuration.md.

pip install -e ".[local,dev]"
pytest

Documented but intentionally not built yet, to keep the first build bare-bones:

• A physical "clip that" button as a manual backup to the voice trigger.
• A small OLED/e-ink display showing last-clip status or a QR code to the UI.
• whisper.cpp as an alternative on-device transcription backend for slower Pis.

MIT — code and docs. The hardware build uses off-the-shelf parts (no custom PCB), so the wiring guide is documentation under the same license.