A wake word is a short spoken phrase that a voice assistant continuously listens for on the device and that, once detected, opens a session in which the following speech is recorded and interpreted as a voice command. Among voice assistants, phrases such as “Alexa”, “Hey Google” and “Hey Siri” are the most familiar examples, and they act as the gate between passive listening and active processing.
The wake word exists for both technical and privacy reasons. Recognizing full speech requires substantial computing power, usually in the cloud, while detecting one fixed phrase can run on a tiny, low-power chip. Only after the phrase is detected does the device begin streaming audio for full recognition.
How wake word detection works
Wake word detection is a specialized form of keyword spotting. The device keeps a short rolling audio buffer, typically a second or two, and passes small frames of sound through a compact neural network trained to recognize one phrase. Everything else is discarded continuously without leaving the device.
Most smart speakers use a multi-stage pipeline. A very low-power first stage runs constantly; a more accurate second stage confirms the detection; and some assistants add a cloud verification step that rechecks the audio before acting, to cancel false triggers.
| Stage | Where it runs | Purpose |
|---|---|---|
| Audio front end | Digital signal processor | Beamforming, echo cancellation, noise reduction |
| First-stage detector | Always-on low-power core | Cheap, sensitive screening of every frame |
| Second-stage verifier | Main processor on the device | Rejects near-misses with a larger model |
| Cloud verification | Assistant service | Final check before the request is executed |
| Speaker identification | Device or cloud | Matches the voice to a profile, as in Voice Match |
The audio front end matters as much as the model. A far-field microphone array with beamforming focuses on the speaker’s direction, while acoustic echo cancellation lets the device hear its wake word even while playing music loudly.
Accuracy measures and design trade-offs
Wake word performance is described by two error rates that pull in opposite directions. Raising the detection threshold reduces accidental activations but makes the assistant ignore genuine requests more often.
| Metric | Definition | Effect on users |
|---|---|---|
| False accept rate | Unwanted activations per hour of background audio | Unexpected listening, privacy concerns |
| False reject rate | Share of genuine wake words missed | Repeated phrases, frustration |
| Detection latency | Time from end of phrase to response cue | Perceived responsiveness |
| Power consumption | Energy used by always-on listening | Battery life on portable devices |
Designers tune the operating point for the setting. A device in a kitchen with a television nearby needs a stricter threshold than a phone held close to the mouth. Phrases with three or more distinct syllables are easier to detect reliably than short single words, which is why many assistants use two-word phrases.
Wake word compared with other activation methods
The wake word is one of several ways to start a voice interaction. Each method strikes a different balance between convenience, privacy and reliability.
| Method | Hands-free | Privacy profile | Typical use |
|---|---|---|---|
| Wake word | Yes | Microphone always on, local detection | Smart speakers, displays, phones |
| Push-to-talk button | No | Microphone off until pressed | Remotes, cars, some wearables |
| Continued conversation | Yes, for a short window | Listens briefly after a reply | Follow-up questions |
| Gaze or proximity detection | Yes | Uses camera or sensors | Some smart displays |
| Physical mute switch | Not applicable | Disconnects microphones | Privacy control on speakers |
Many products combine methods: a wake word for convenience and a hardware mute switch for moments when no listening should occur. Remote controls with a microphone usually rely on push-to-talk because their microphone is too close to the speaker for far-field processing to be needed.
Custom wake words and local voice control
Commercial assistants offer only a small set of fixed wake words, because each phrase requires a separately trained and tested model. Some allow a choice among a few alternatives, such as changing “Alexa” to another approved name.
Open and local voice control systems for the smart home allow custom wake words trained from synthetic speech samples. These can run on a small computer or a satellite microphone in each room, keeping all audio inside the home. The trade-off is that custom models are usually less robust to accents and noise than heavily tuned commercial ones.
- Choose a phrase that is uncommon in everyday conversation and television.
- Avoid words that rhyme with names of household members.
- Test with different speakers and distances before relying on the phrase.
- Keep a push-to-talk fallback for noisy rooms.
Privacy, false triggers and common misconceptions
The wake word is central to the privacy model of voice assistants. Audio before the phrase is meant to stay on the device, but a false accept causes a short recording to be sent for processing, which is the source of most unexpected recordings.
| Misconception | Reality |
|---|---|
| The device records everything | Only a short local buffer exists until the wake word is detected |
| The wake word is understood by full speech recognition | A dedicated small model spots one phrase |
| False triggers mean the device is faulty | Every detector has a nonzero false accept rate |
| Mute only silences the speaker | On most smart speakers it disables the microphones |
| Any word can be the wake word | Commercial devices support only trained phrases |
Users who want stronger control can review and delete voice history, enable automatic deletion and use the mute switch. Lowering wake word sensitivity, where offered, reduces false accepts at the cost of more missed requests.
Put simply
A wake word is the name you call to get a voice assistant’s attention, such as “Alexa” or “Hey Google”. The speaker is like a person half-asleep in a room: it hears noise all the time but only wakes up when it hears its own name.
Until it hears that word, the device is designed to forget what it heard after a moment. Once it hears it, it starts listening properly to your request. Sometimes it mishears a similar word on television, which is why a speaker occasionally lights up when nobody spoke to it.
In practice
A household with a smart speaker in the living room notices that it lights up several times each evening during television programs. The voice history shows short recordings of dialogue that sounded similar to the wake word, a classic pattern of false accepts caused by media audio.
The household changes the wake word to a less common alternative offered by the device and lowers the wake word sensitivity one step in the app. They also move the speaker about two meters away from the television loudspeakers, which reduces the energy of the interfering sound reaching the microphones.
Over the following weeks, accidental activations drop to roughly one per week, while genuine requests from across the room still work on the first try in most cases. The household also enables automatic deletion of recordings, combining better detection with stronger privacy.