ORBIS

Appearance

Train your own wake word ​

ORBIS detects wake words with openWakeWord — small, on-device models that run before the main speech pipeline. The picker in Settings → Voice → Activation ships "Hey Orbis" plus the stock openWakeWord set, but you can train a model for any phrase you like ("Hey Computer", your own name, …) and integrate it into a development build.

A wake model is tiny (~200 KB–1.5 MB) and trains on synthetic speech, so you don't record anything yourself — a free Colab GPU does it in under an hour.

1. Train the model ​

openWakeWord ships an automatic training pipeline. The easiest path is its Colab notebook:

  1. Open openWakeWord → automatic model training (the automatic_model_training.ipynb notebook — "Open in Colab").
  2. Set your target phrase (e.g. hey computer). Keep it 2–4 syllables and distinct from everyday speech — short or common phrases false-fire.
  3. Run the cells. It synthesizes thousands of positive samples (via Piper TTS) plus hard negatives, trains a small classifier, and exports an .onnx.
  4. Download your_phrase.onnx when it finishes (~30–60 min on the free GPU).

Want the exact recipe we used for "Hey Orbis"? It's at protoLabsAI/hey-orbis-wakeword.

2. Integrate it into ORBIS ​

The shipped picker accepts the pinned built-in catalog. Dropping an arbitrary file into the models directory does not register it or make it selectable. Custom model import is not currently an app feature.

For a development build, add an entry to voice/wakeword_catalog.json with the classifier's filename, immutable download URL, SHA256, input embedding window and output scoring offset. Python downloads and Rust inference read this same manifest. Keep the shared mel-spectrogram and speech-embedding dependencies. Verify the input/output shape and Rust/Python score parity before exposing it in the picker, then run spoken positives and hard negatives to tune its threshold. Build the native app with the updated catalog.

3. Select + tune it ​

Follow Enable wake-word activation: download and select the catalog entry, enable Wake word, and relaunch. Sensitivity changes also apply on the next launch. Lower thresholds can increase false triggers.

Troubleshooting ​

  • Never fires. First check your mic input level — System Settings → Sound → Input → raise Input volume. openWakeWord needs a real signal; a near-silent mic (the meter barely moving) won't trigger it no matter the threshold. See Voice isn't working.
  • False fires. Raise the sensitivity threshold, or pick a longer / more distinctive phrase and retrain.
  • Won't load. Make sure both shared models are present in ~/Library/Application Support/studio.protolabs.orbis/models/wakeword/, the catalog checksum matches, and the classifier's input/output shapes agree with its catalog entry.