How Do I Train a Custom Wake Word for Android?
Train your phrase online, download an ONNX or TFLite model, and integrate offline keyword spotting into an Android app.
Training workflow
- Define one or more phrases and choose a supported language.
- Train and test with representative voices and microphones.
- Optionally upload authorized recordings to enhance robustness.
- Export the model and connect it to Android audio capture and your command handler.
Multiple keywords in one Android model
One model can recognize multiple fixed keywords and return a label for each command. This is useful for hands-free controls such as open, close, and cancel without shipping a separate model for every phrase.
Android implementation notes
Use ONNX Runtime or TensorFlow Lite with AudioRecord, request microphone permission, and account for sample-rate differences and background execution limits. Tune thresholds on the actual phone models and microphones you plan to support.
Languages
English, Chinese, Spanish, Japanese, German, French, and Hindi are supported. Accuracy varies with wording, speakers, training data, threshold, and environment.
Train a custom wake word online →