What is Wake Word?
Also called hotword, trigger word, keyword spotting.
A wake word is a short spoken phrase that a device or application listens for in order to activate its full voice pipeline. A small always-on detector runs continuously on a local audio buffer and only opens the main recognition path once the phrase matches, which keeps compute low and limits how much audio ever leaves the device.
Wake word detection is usually implemented as keyword spotting, a narrow classification task rather than general transcription. A compact acoustic model scores short overlapping windows of audio against the target phrase and fires when confidence crosses a threshold. Because the model is small, it can run on a microcontroller, a phone's low-power audio processor, or a browser worker without the battery cost of a full speech recognizer.
Designers tune the threshold along a tradeoff between false accepts, where the system wakes on unrelated speech, and false rejects, where a genuine invocation is missed. Longer and phonetically distinctive phrases reduce false accepts, which is why most commercial wake words have three or more syllables and uncommon sound sequences. Accents, background noise, and playback of the phrase from television audio all complicate detection.
Privacy expectations shape the architecture. The common design keeps a rolling buffer of a few seconds in memory, discards it continuously, and only streams audio to a server after a match, sometimes including a short pre-roll so the first words of the request are not clipped. Vendors differ in whether detection audio and false accepts are retained for model improvement, and this remains a contested area of consumer trust.
In agent deployments a wake word is one of several activation strategies. Kiosks and hands-free settings favor it, while headsets and desktop applications often prefer an explicit control. Some systems combine both, allowing a wake word for hands-free use and a button for noisy environments.
Key points
- Small always-on model gates the expensive speech pipeline
- Threshold tuning trades false accepts against missed activations
- Longer, distinctive phrases detect more reliably
- Pre-roll buffering prevents clipping the start of a request
- Local detection limits how much audio is transmitted
In practice
A warehouse tablet runs a voice assistant that stays silent until it hears its two-word activation phrase. A detector scores the microphone stream locally at all times, and forklift noise never crosses the threshold. When a worker says the phrase, the tablet plays a short tone, streams the following utterance to the recognizer along with half a second of buffered pre-roll audio, and shows the transcribed stock query on screen.