How to Train a Custom Wake Word
There are two practical ways to get a custom wake-word model: build the training pipeline yourself, or use an online platform that generates and lets you test the model before download.
Self-training: maximum control, maximum setup
A self-hosted workflow normally includes phrase selection, synthetic positive data, real recordings, negative audio, augmentation, training, export, threshold tuning and evaluation on the target microphone. With openWakeWord, the official project provides an automated Colab path and a more customizable notebook. Its documentation recommends several thousand positive examples and substantial negative data; the current official README also says the main project is focused on English.
| Strengths | Costs and risks |
|---|---|
| Control over code, data, checkpoints and evaluation | Environment and dependency setup can be difficult |
| Can build a reproducible internal pipeline | You own data curation, hard negatives and false-trigger testing |
| Useful for research or specialized constraints | Multilingual data and deployment packaging become your responsibility |
Online training: faster validation and delivery
An online platform hides most of the training infrastructure. You enter a phrase, generate a model, test it in the browser, and download a runtime package. Picovoice Console can train a typed phrase in seconds and produces platform-specific keyword files for its SDK; runtime use requires its AccessKey and vendor runtime. Voicute is designed for online training and pre-download testing with standard ONNX/TFLite delivery and open offline inference examples.
Which option is better?
| Situation | Better starting point | Why |
|---|---|---|
| You need a multilingual product prototype quickly | Online platform | Less setup and faster phrase-to-device validation |
| You already use a Picovoice SDK | Picovoice Console | Custom phrase and platform SDK fit are convenient |
| You need open model files and no runtime key | Voicute or a self-hosted pipeline | Compare model format, runtime and commercial terms |
| You have ML engineers and unusual data requirements | Self-training | Maximum control, with more maintenance |
What “better” must be measured against
Do not compare only a demo score. Test the same phrase, speakers, microphone, enclosure, distance and background audio. Record recall, false triggers per hour, latency, power and behavior after repeated commands. A good online platform is the one that reaches a trustworthy target-device result with the least work for your team.
Recommended path for a small product team
- Choose a short, distinctive phrase and decide whether it is a wake word or a fixed command.
- Generate and test it online in each target language.
- Run the downloaded model on the closest target platform.
- Only build a self-training pipeline if the online result, data ownership or customization requirement justifies the ongoing engineering cost.
Primary references
- openWakeWord training documentation
- Picovoice Porcupine custom wake-word documentation
- Picovoice model API and supported languages