How to Turn Your Product Name into a Custom Wake Word
A smart light, robot, appliance or app may already have a product name. The next question is whether the device can actually listen for that name locally and start the right action.
A product name is not automatically a good wake word
A brand name is designed for people. A wake word is designed for a detector. A useful wake word should be easy for target users to pronounce, long enough to be distinguishable, and uncommon enough to avoid frequent false triggers.
“Turn on the living-room light” is a fixed command. “Hey Taro” could be a wake phrase. If the product only needs a short list of actions such as on, off, brighter and dimmer, a multi-keyword model may be more appropriate than a separate wake-word and full-ASR pipeline.
Choose KWS or ASR from the product requirement
| Product need | Good first option | Why |
|---|---|---|
| Wake a device or app | Single-keyword KWS | It detects one known phrase with a simple response path |
| Recognize 2–10 fixed actions | Multi-keyword KWS | One model can map several phrases to actions |
| Understand arbitrary sentences or numbers | ASR plus intent handling | The system needs open speech converted into text |
| Verify who is speaking | Speaker verification | That is an identity problem, not a wake-word problem |
If the product only needs a wake phrase or a few fixed commands, starting with KWS can avoid the memory, latency and integration cost of a full speech-recognition stack.
How to evaluate a product name
- Read it aloud instead of judging the spelling alone.
- Avoid very short, common or easily confused phrases.
- Ask several speakers to say it at a natural speed.
- Check accents, languages and pronunciation differences in the target market.
- Test both detection and false triggers from ordinary conversation.
Browser testing is useful for rejecting obviously weak candidates. It is not a substitute for testing the final microphone, enclosure, distance and noise environment.
From a product name to an offline model
- Enter the product name or fixed command.
- Select the training language and model option.
- Generate the model and test it with real voices.
- Download a standard ONNX or TFLite model when the result is promising.
- Use the inference examples with Android, Web, Python, desktop or embedded targets.
- Measure recall, false triggers, latency and power on the final device.
When real-voice enhancement helps
Real-voice data is useful when a product targets a specific accent or user group, or when synthetic-voice tests look good but natural speech is less reliable. It complements, rather than replaces, real-device validation.
Where Voicute fits
Voicute moves the question “can my device hear its own name?” before model download. A product team can enter its own product name, train and test a Basic model, then decide whether it needs a downloadable model, real-voice enhancement, multiple keywords or further false-trigger optimization.
The downloaded model can run locally without cloud inference, an AccessKey or a per-device runtime fee. The final result still depends on the phrase, language, microphone, enclosure and target device.