How Do I Train a Custom Wake Word for Web Apps?

Create a custom offline wake-word model for browser applications, with one model able to recognize multiple keywords.

From phrase to browser model

1. Enter your wake word or several command phrases.
2. Choose English, Chinese, Spanish, Japanese, German, French, or Hindi.
3. Train and evaluate the model, then optionally upload authorized recordings for enhancement.
4. Export ONNX or TFLite and run it locally with a browser-compatible runtime.

Multi-keyword recognition

A single model can emit different labels for multiple phrases, allowing a web interface to distinguish commands without sending microphone audio to a server.

Web deployment notes

Use ONNX Runtime Web or a compatible TFLite Web runtime with WebAssembly or WebGL where appropriate. The browser still requires microphone permission, and performance depends on the browser, device, audio sample rate, and CPU. Always test the production browser matrix and tune the threshold for false activations.

Train a custom wake word online →