The wake word is the entry condition for engagement
Aliases: hotword · keyword spotting · voice wake · always-on listening
What it is
A wake word (also called a hotword) is the acoustic passphrase that moves a device from low-power listening into engagement. Before a hit, later speech must not be treated as a command; after a hit, a full recognition and dialogue session may open. The phrase is not the command. It is the entry condition. Push-to-talk uses a finger for the same gate; a wake word uses a fixed utterance.
Why it happens
If a far-field microphone streamed every sound into automatic speech recognition, power, bandwidth, and privacy would all fail. Typical stacks therefore split in two: on-device keyword spotting matches a short sliding window against a handful of phoneme sequences; only a score above threshold wakes the heavier front-end, the network path, and language understanding. Audio that misses the gate is usually discarded locally and never parsed as a command. The wake word is thus a state-machine guard: it separates “might be listening” from “is accepting input.” A command that is acoustically clear still must not take effect if it falls outside the guard. Wake-word-free follow-ups inside an open session are a temporary relaxation of the guard, not its disappearance.
Studying it
The headline metric for keyword spotting is false accepts per hour at a target miss rate, not whether a cooperative lab utterance wakes the device once. Independent variables include distance, angle, television and conversation in the background, whether the speaker is enrolled, and whether the phrase must be complete. Latency from acoustic hit to a visible listening state belongs in the same protocol. Quiet-room successes overstate availability: media speech, children mimicking the phrase, and brand names in advertisements are the distribution the guard actually faces.
Where it stops holding
Devices without an always-on microphone do not need a wake word; a steering-wheel button or an earbud tap already supplies the entry condition. Once a session is open, many products allow a few seconds of follow-up without repeating the phrase, so the gate is no longer that utterance. A custom wake word changes the collision surface with everyday speech, but that is a false-trigger statistic, not a question of whether a gate exists. Treating the wake word as part of natural-language understanding leads people to believe any sentence will start the device, while the stack is still waiting for a template.
Applying it
- Draw idle, listening, and processing as mutually exclusive states; execute side-effecting commands only after a hit.
- Test far-field rooms with television and bystanders: with no wake word spoken, no command may run.
- Log keyword hits separately from successful commands, so recognition accuracy is not mistaken for a working gate.