Train your own openWakeWord wake word (2026 edition)

A working end-to-end kit for training a custom wake word ("hey Robby", "okay Mango", …) with openWakeWord on today's Google Colab — plus, more importantly, the measurement gate that tells you whether the result is safe to deploy.

Why this exists: openWakeWord's official training notebook no longer runs on current Colab (Python 3.12, torch 2.11, reorganized datasets). This kit encodes ~10 verified fixes — see the "Live-run addendum" in FINDINGS.md for the full list — and was proven end to end on 2026-08-01: the resulting model measured 0.00 false accepts/hour over 75 minutes of real living-room audio at 91% recall on the target device (a Reachy Mini robot).

It was built for a robot named Clem; everything except one config cell is name-agnostic.

What's in the box

File	What it does
hey_clem_colab.ipynb	The trainer. Upload to Colab, edit ONE cell, Run all (~1.5–2 h on a free-tier T4).
oww_train_shim.py	Compat shim for openwakeword.train on a modern stack (also written by the notebook itself).
check_model.py	The deployment gate. False-accepts/hour + recall vs threshold, from real audio. Do not skip.
phrase_check.py	Compare candidate phrases (phonemes, clip length) before spending an hour training the wrong one.
slice_positives.py	Splits "me saying the phrase 20 times" into the per-utterance clips the gate needs.
hey_clem.yaml	The training config, as reference — the notebook embeds its own copy.
FINDINGS.md	Findings, phrase research, gate criteria, rollback thinking, and every wall we hit with its fix.

How to use it

1. Pick a good phrase. Short phrases false-accept more. Aim for 3+ syllables / ~8 phonemes ("hello Robby", not "Rob"). If you have espeak-ng locally, phrase_check.py will show you phonemes and (with a Piper checkout) measured clip lengths. Training several phrasings into one model works — we trained "hey / okay / hello Clem" together, and note that n_samples in the config is a TOTAL split across phrases (~25k per phrase is a good density).

2. Edit one cell. In the notebook, the %%writefile hey_clem.yaml cell holds the whole config. Change:

• model_name — yours
• target_phrase — your phrase(s), lowercase
• custom_negative_phrases — near-misses for YOUR name (rhymes, one-sound-off words; ours had "hey clam", "hello glen"…). This is where short wake words win or lose.

Leave the rest alone unless you've read the yaml's comments.

3. Run it. Colab → File → Upload notebook → Runtime → T4 GPU → Run all.

⚠ Budget for Colab Pro (US$9.99/month). The full job is ~2 GPU-hours, and free-tier credit is unpredictable — in our runs it expired mid-job and stranded a nearly-finished model, forcing a rerun. Pro finished the same job comfortably; you can cancel after training. If you do try the free tier, be prepared for the session to be reclaimed partway through. Keep the tab open. Cell 1 fails fast if you didn't get a GPU; cell 2 ends with a preflight so environment problems die in the fast cell, not 30 minutes into training. At the end your browser downloads a zip with <name>.onnx and <name>.onnx.data — you need BOTH (the weights live in the sidecar). To fold them into a single file:

python -c "import onnx; m = onnx.load('hey_clem.onnx'); onnx.save_model(m, 'hey_clem_single.onnx', save_as_external_data=False)"

A ModuleNotFoundError: onnx_tf at the very end of the training cell is expected and harmless — it's a tflite conversion attempt that fires after the ONNX is already written.

4. MEASURE BEFORE YOU DEPLOY. This is the part most people skip and the reason this kit exists. A wake word listens all day; an unmeasured one is a false-accept machine. Record at least one hour of ordinary audio from the room it will live in (any recorder, any sample rate — phone is fine; nobody says the wake phrase), plus ~20 real recordings of you saying it (slice_positives.py splits a single say-it-repeatedly session). Then:

python check_model.py --model your_model.onnx --negatives room_audio/ --positives your_clips/

It reports false-accepts/hour and recall per threshold and recommends the threshold — or refuses, which is also an answer. A reasonable bar: ≤0.5 false accepts/hour at ≥80% recall. If it can't hit both, don't lower the bar — train a longer phrase, or look at openWakeWord's custom_verifier_model (gates detections on your specific voice).

⚠ Score on Linux or on your deployment device — in our runs, macOS scored every model near-zero through the same code path that read 0.89+ on Linux (unresolved platform quirk). The gate's numbers are only as good as the machine they ran on; run them where the model will actually live.

5. Deploy at the recommended threshold, and keep your old wake word one config line away for rollback.

What this kit does — and doesn't — provide

This kit produces and validates the wake-word model. It does not give your app ears: something in your stack must route mic audio through openWakeWord and act on detections. Where that already exists:

• Anything that embeds openWakeWord (Python stacks, Wyoming satellites, Home Assistant's wake-word ecosystem) can load the .onnx directly. Note HA's add-on defaults to tflite models; this kit deliberately produces ONNX.

• Reachy Mini: the stock conversation app has no wake-word support today. Standby-mode support (opt-in STANDBY_ON_SLEEP=1, wake model via WAKE_WORD_MODEL — pretrained name or a custom .onnx from this kit) has been proposed upstream: pollen-robotics/reachy_mini_conversation_app PR #514.

  PR #514 status: open, awaiting maintainer review (last checked 2026-08-02). This README will be updated once its fate is decided — merged (use the stock app) or not (use the PR branch as the reference integration).

Runtime notes for your own stack

Loading the model: openwakeword.model.Model(wakeword_models=["path/to/your.onnx"], inference_framework="onnx"), fed 80 ms chunks (1280 samples) of 16 kHz mono int16. On Python 3.12, install openwakeword with --no-deps plus onnxruntime scipy scikit-learn requests tqdm — a plain pip install openwakeword silently downgrades to a fossil version (no py3.12 tflite-runtime exists).

---

Built with Claude (2026-08-01) for a Reachy Mini named Clem — who woke to his own name on the first try. License: Unlicense — public domain; no attribution required, no rights reserved.