What is a wake word?
A wake word is a short spoken phrase that a voice device listens for continuously and treats as permission to start processing. Until it is detected, audio is scanned locally and discarded frame by frame; once it fires, the device opens a recognition stream and begins acting on what follows.
The detector doing this work is deliberately tiny. A keyword spotting model sized in the low hundreds of kilobytes runs on a phone’s low-power audio core, scoring overlapping audio frames of roughly twenty to thirty milliseconds each, so the main processor stays asleep until the phrase arrives.
How a wake word works
Detection runs as five stages, and each one exists to keep the stage after it from running unnecessarily.
Audio from the microphone lands in a short rolling buffer, usually a few seconds long, which is overwritten continuously so nothing persists. Voice activity detection gates that buffer, letting silence and steady room noise fall away before any model runs. Surviving audio is converted into acoustic features, typically log-mel spectrogram frames that compress the signal into something a small classifier can score.
The classifier is a keyword spotting model trained on one phrase. It scores each sliding window for how closely the audio matches, and a smoothing function plus a confidence threshold turns that stream of scores into a single yes. When it fires, the buffered pre-roll and the audio that follows are streamed to automatic speech recognition, which transcribes the utterance and hands it to the intent or language model behind it.
Types of wake word detection
Fixed-phrase keyword spotting: A model trained on one phrase and shipped with thresholds tuned for it across accents and noise, which makes it accurate and inflexible.
User-defined wake words: The customer enrolls a phrase of their own choosing from a handful of spoken samples, usually at some cost to detection accuracy.
Speaker-verified activation: Detection is paired with a voiceprint check so the device wakes only for enrolled speakers, which matters when the trigger unlocks personal data.
Two-stage verification: A small on-device detector fires, then a larger cloud model re-examines the same audio and cancels the session when the phrase is absent.
Multi-phrase command spotting: Several short phrases are recognized at the wake stage itself, letting simple commands skip full transcription and shave a turn off the interaction.
Wake word vs voice activity detection vs speech recognition vs push-to-talk
These four all decide when a system starts paying attention, and teams specifying a voice product routinely swap one for another in a requirements document. Voice activity detection answers whether any human speech is present in the stream. Automatic speech recognition answers what was said, once something has decided the audio deserves transcribing. Push-to-talk answers nothing acoustically, because a button press supplies the intent directly. A wake word answers a narrower question: did this one specific phrase occur, and should the expensive stages downstream now run.
What it detects | Where it runs | What it triggers | Cost when wrong | Choose it when | |
|---|---|---|---|---|---|
Wake word | One trained phrase | On-device, always on | A full recognition session | False wake or missed wake | The mic is shared with a room |
Voice activity detection | Any speech versus non-speech | On-device or in the pipeline | Downstream processing of a segment | Clipped or padded speech | You need to gate silence cheaply |
Automatic speech recognition | Words in an audio stream | Device or cloud, per session | Intent handling and generation | Wrong transcript, wrong action | Something already opened the stream |
Push-to-talk | Nothing acoustic, a button press | Hardware or UI layer | Capture for the press duration | Missed press, held too long | Hands are free and privacy is strict |
If your product owns a microphone continuously and someone must be able to start a conversation without touching anything, you need a wake word. If a session is already open, such as a phone call, endpointing and voice activity detection are the mechanisms that decide your turn-taking.
Why wake words matter for customer experience
Without a trigger, a voice product has two options and both are poor. It can stream every second of room audio to a server, which costs money and turns the device into a surveillance question. It can also demand a button press, which removes the hands-free property that made voice worth building.
The wake word buys a third option, imperfectly. Every deployment sets a confidence threshold, and that single number trades the two failure modes against each other: lower it and the device wakes for people who never addressed it, raise it and customers repeat themselves until they give up. The two errors do not feel alike. A missed wake makes a product seem deaf; a false wake makes it seem intrusive, and that impression rarely recovers. Once the session is open, barge-in governs the rest of the turn, letting a caller cut through a prompt they have already heard.
How is wake word accuracy measured?
Accuracy is reported as a pair of error rates measured on the same audio. False accept rate counts activations on speech that never contained the phrase, normally expressed per hour of continuous audio, because the device is exposed to time and time is the natural denominator. False reject rate counts genuine attempts the detector missed, expressed as a share of attempts. Sweeping the threshold across its range traces a detection error tradeoff curve, which is the honest way to compare engines, since any single point can be moved by tuning.
Latency is the third number: the gap between the end of the spoken phrase and the moment the recognition stream opens, because a slow trigger clips the start of the command.
Public reference data is thin at this layer. The LibriSpeech ASR corpus gives the recognition stage behind the wake word a shared benchmark, roughly 1,000 hours of read English speech released as 100-, 360-, and 500-hour training subsets with matched clean and noisier evaluation splits, and no wake word corpus carries comparable standing, so detection figures quoted to you come from private test sets.
How AI agents change wake words
Two shifts matter. In telephony the call itself is the session, so a contact-center voice agent has no wake word at all, and the equivalent problem becomes endpointing: deciding when the caller has finished speaking so the agent may reply. On devices, kiosks, and in-app assistants the wake word survives, because the microphone is shared with a room full of other conversations.
Speech-to-speech models change what a late trigger costs. When one model consumes audio and returns audio, the boundary the wake word draws decides exactly what context that model ever sees, so words spoken before the trigger are gone from the input entirely.
Follow-up modes, where the device holds the stream open for a few seconds after replying, cut how often a customer repeats the phrase, and they concentrate false-accept risk inside that open window. Teams comparing call center voice AI tools should ask which activation model a vendor assumes.
What to look for in a wake word engine
Coverage comes first. Ask for detection rates across the accents, ages, and languages your customers actually have, measured in the acoustic conditions you deploy in, since far-field audio in a warehouse behaves nothing like a phone held at arm’s length.
Integration surface is next. An engine shipped as an embedded library with control over the audio front end (gain, echo cancellation, beamforming) behaves predictably; one that only exposes a cloud endpoint moves your always-on audio off the device.
Ownership of the buffer is the governance question. Decide who can retrieve pre-roll audio captured during a false accept, how long it survives, and whether it feeds training. Regulated buyers ask how that retention is evidenced under GDPR, and healthcare buyers ask the same question in HIPAA language, so the answer has to be a written retention policy with a named owner.
The constraint teams underestimate is phrase choice. Once a wake word ships in firmware, packaging, and advertising, replacing it becomes a hardware-cycle decision, and threshold tuning will never repair a phrase that collides with ordinary speech.
Wake words and the rest of the voice stack
The trigger sets the boundary for everything downstream. Speech synthesis shares the same acoustic space, which is why a detector has to ignore the device’s own voice: with no echo cancellation tied to the playback path, a system can wake itself by saying its own name inside a prompt.
Analysis layers inherit that boundary too. Emotion detection only ever sees audio captured after the trigger fired, so the frustrated sigh before a third attempt at the phrase never reaches routing or quality review, and the record looks calmer than the interaction was.
What does a wake word mean in plain terms?
Think of a wake word as your own name in a crowded room. Everyone hears everything, almost none of it registers, and the moment someone says your name you start listening properly. The device behaves the same way: audio passes through constantly and nearly all of it disappears within seconds of arriving.
Take the wake word away and the choice turns stark. Either the device ships everything it hears somewhere else to be understood, or you have to walk over and touch it before speaking, which defeats the point of talking to it from across the kitchen.
The tradeoff is that no phrase is ever perfect. Make it easy to say and it will surface in ordinary conversation; make it distinctive and people fumble it. Every product picks a point on that line and absorbs the complaints from the other end.
Common wake word mistakes
Four patterns cause most of the trouble.
Choosing a phonetically weak phrase is the first. Short phrases, common syllables, and words that hide inside longer words give the classifier little to discriminate on, so the false accept rate is largely set before a model is ever trained.
Tuning the threshold on clean audio is the second. Development recordings are close-mic’d and quiet, and a threshold that looks precise there collapses in a living room with a television on, where the model meets speech-shaped noise it never saw.
Treating pre-roll as ephemeral is the third. The buffer that makes wake words feel responsive also captures whatever was being said just before the trigger, and that audio lands in logs, review queues, and occasionally training sets nobody scoped.
The fourth is testing only with the team that built it. Native-accented adults in a quiet office produce flattering numbers, and the complaints arrive from children, second-language speakers, and anyone in a noisy room, the pattern behind many AI customer service failures that look like model problems and are really coverage problems.
Frequently Asked Questions
What is the difference between a wake word and a hotword?
Wake word and hotword describe the same mechanism in most documentation: a phrase a device listens for continuously to open a recognition session. Hotword is the older engineering term and still appears in keyword spotting literature, while wake word became the consumer-facing label. Some teams reserve hotword for in-session shortcut phrases recognized after the device is already awake.
Wake word vs voice activity detection: which does what?
Wake word detection identifies one specific trained phrase and treats it as permission to start a session. Voice activity detection identifies whether any human speech is present at all, with no interest in the words. Most stacks run both: activity detection gates cheap frames continuously, and the wake word classifier scores only the segments that contain speech.
Can you create a custom wake word?
Custom wake words are supported by many engines, either through user enrollment from a few spoken samples or through a vendor training a dedicated model for a chosen phrase. Enrolled phrases usually detect less reliably than factory-trained ones, because the model has far fewer examples covering accents, distances, and background noise conditions.
Does a wake word mean the device is always recording?
A wake word device is always listening, and listening and recording are different operations. Audio flows through a short local buffer that overwrites itself continuously, with nothing transmitted until the phrase is detected. What leaves the device after a trigger typically includes a few seconds of buffered pre-roll audio, which is why retention policy matters.
Why does a voice assistant activate when nobody said the wake word?
Wake word false activations happen when ordinary speech, television audio, or background noise produces acoustic patterns close enough to the trained phrase to cross the confidence threshold. Similar-sounding names and syllables hidden inside longer words are common culprits. Lowering false activations means raising the threshold, which increases how often genuine attempts get missed.
What makes a good wake word?
A good wake word is three or four syllables long, phonetically distinctive, and unlikely to occur in normal conversation near the device. Length gives the classifier more signal to score. Uncommon consonant clusters help separate it from everyday speech. It also has to be easy for children, second-language speakers, and people in noisy rooms to pronounce consistently.