How Do I Train a Custom Wake Word for Raspberry Pi?
Create an offline wake-word model for Raspberry Pi from phrases, then test it with the same USB, I2S, or headset microphone used by your project.
How do I train it?
- Enter one or more wake words or command phrases.
- Select English, Chinese, Spanish, Japanese, German, French, or Hindi.
- Train, evaluate, and optionally upload authorized real recordings to enhance the model.
- Download ONNX or TFLite and run it beside your local Python application.
Can one model recognize multiple keywords?
Yes. A single model can classify multiple labels such as play, pause, and next. This reduces model management for offline Raspberry Pi voice interfaces.
Raspberry Pi deployment notes
Use ONNX Runtime or TFLite Runtime on Raspberry Pi OS. Start with a short audio window and tune the confidence threshold using the final microphone, room noise, CPU model, and sample rate. GPIO actions, media control, and Home Assistant calls should be kept outside the model as application logic.
What should I validate?
Test quiet and noisy rooms, different speakers, distance from the microphone, false activations, and repeated phrases. Upload only recordings you have permission to use.
Train a custom wake word online →