How Do I Train a Custom Wake Word for Linux?

Train a custom offline wake-word model and integrate it with Linux applications, scripts, kiosks, or local voice automation.

Training workflow

1. Enter one or more keywords or command phrases.
2. Choose English, Chinese, Spanish, Japanese, German, French, or Hindi.
3. Train and test, with optional authorized recordings to enhance the model.
4. Export ONNX or TFLite and run inference in your Linux application.

Can one model recognize several keywords?

Yes. A multi-keyword model returns a label for each phrase, allowing one local service to handle several commands instead of loading separate wake-word models.

Linux deployment notes

Use ONNX Runtime or TFLite Runtime with ALSA, PulseAudio, or PipeWire audio capture. Validate device selection, sample rate, CPU usage, service restarts, and thresholds under the same noise conditions as deployment. The model can remain local for privacy-sensitive applications.

Train a custom wake word online →