try it live: tap the bubble and say “hey outspoken”

Train your own wake word.

Type a phrase, get a small file that runs on a Raspberry Pi, in Home Assistant, in the browser or in your app. Offline, no cloud, no ML expertise.

Train your wake word
NO CLOUD

Listens on the device, not in the cloud

Your audio never leaves the device. No server to pay for, and it keeps working when the wifi doesn't.

No cloudNo internet neededPrivate by design

Built for tinkerers. Ready for products.

Home Assistant and hobby projects

Pay once per model and use it in as many hobby projects as you like. Drop the .tflite into the openWakeWord add-on and Assist starts listening.

  • ✓Works with the openWakeWord add-on
  • ✓Dutch, German and French too
  • ✓Pay once, unlimited hobby use
Home Assistant setup guide →

Shipping a product?

A wake word plus fixed commands like “stop timer”, in one on-device model. Ship it in your app or device without per-user fees.

  • ✓Subscribe to Studio only when you run it commercially
See Studio pricing →

Pay per model. Subscribe only to ship.

A Balanced model costs 1 credit, Premium costs 2. Credits never expire. Hobby use needs no subscription; commercial use needs Studio.

Try it first
€0

Test any model in the playground. No account needed.

Starter
€9 3 credits

Three Balanced models, or one Premium and one Balanced.

Builder
Best value
€29 10 credits

€2.90 per model. Enough to iterate on a wake word and its commands.

Running it commercially? Studio, €19/month →

Questions, answered.

How much does it cost to train a wake word model?

Credits are sold in one-time packs starting at €9 for 3 model credits, or as part of the €19/month Studio license subscription. Credits never expire.

Are there subscriptions or per-device fees?

The €19/month Studio license is a subscription that includes 5 model credits every month, unlimited benchmark runs, and commercial deployment rights. One-time credit packs remain available for personal, hobby, and prototype/internal evaluation use without a subscription.

What format are the trained models?

Models are exported in both ONNX and TFLite formats. ONNX runs anywhere ONNX Runtime is supported — Home Assistant, Python, React Native, C++, browsers. TFLite covers Android, microcontrollers (TFLite Micro), and Coral TPU.

Where does audio processing happen?

All inference runs entirely on-device. Audio never leaves the device — no cloud calls, no latency, no privacy concerns.

Can I use trained models in commercial products?

Personal, hobby, and prototype use are included with credit packs. Commercial venue, fleet, resale, or customer deployments require either the self-serve €19/month Studio license or, for fleet/enterprise volume and custom terms, a separate negotiated deployment license.

Say the word.

Your model is ready in about 45 minutes.

Train your first model