Wake word and hotword detection
In one sentence
A wake word is the phrase that activates a listening device, and hotword detection is the always-running, low-power on-device process that recognizes that one phrase and rejects everything else.
Definition
A wake word is the phrase you say to switch a listening device on, like the name you call out to a smart speaker.
Hotword detection is the small, always-running process on the device that recognizes that one phrase and rejects everything else.
Wake word detection exists to solve one problem. A device that has to be ready to listen at any moment cannot stream all of its audio to a server around the clock, for reasons of cost, bandwidth and privacy. So it listens locally for a single trigger, and only that trigger opens the connection.
How it works
- A very small model runs continuously on the device, processing audio locally.
- It is trained to recognize one specific phrase and to reject everything else.
- Only when that phrase is detected does the device begin streaming audio onward.
- Audio from before the trigger is generally discarded on the device, though a short rolling buffer is usually kept so the start of your next sentence is not lost.
The constraints that make it hard
- It has to run on a tiny power and compute budget, indefinitely, without draining a battery.
- It needs a very low false accept rate. A device that wakes on ordinary conversation is unacceptable, both in practice and for trust.
- It also needs a reasonable false reject rate. A device that ignores its own name is useless.
- Those two error rates trade against each other, and where you set the balance is a product decision, not a technical optimum.
- It has to hold up across accents, distance, background noise, and even the wake word being said on a television.
The privacy dimension, which is what the public actually cares about
- The technical position is that audio stays local and is not transmitted until the wake word fires.
- Public trust in that claim has been damaged by incidents of accidental activation and human review of recordings.
- The gap between the technical reality and what people believe is itself a commercially significant fact for anyone shipping always-listening hardware.
Custom wake words, and where they go wrong
- Businesses sometimes want their own brand name as the wake word. It requires training and carries real risk.
- Short or common phrases produce high false accept rates, which is why wake words are usually multi-syllable and phonetically distinctive.
Where it simply does not apply
- A voice widget on a website has no wake word at all. It activates on a click or a behavioral trigger, and the microphone only opens after an explicit user gesture, which is a browser requirement as much as a design choice.
- That removes the always-listening question entirely, which is a genuine privacy advantage worth stating rather than assuming a buyer will notice.
Common misconception
That a wake word device records everything, all the time. The design is built specifically to avoid that. Public skepticism is still understandable given past incidents, and waving it away is a mistake rather than a defense.
Why it matters commercially
Wake words define the always-listening product category and the privacy argument that comes with it. For voice that lives on a website, having no wake word takes that argument off the table completely, and that is worth spelling out rather than assuming a buyer will spot it on their own.
In voice specifically
A text widget has nothing to wake. There is no microphone to gate and no always-on stream to worry about, because typing is discrete and only happens when the visitor chooses. The entire wake word problem exists because audio is continuous and something has to decide when to start paying attention.
Where AsqVox fits
Not applicable. The Orb activates on a click or on behavioral intent signals, and microphone access always requires an explicit user gesture plus browser permission. Nothing monitors audio continuously, which is a meaningful distinction for a privacy-conscious buyer.
Visual
The small model that never stops running
A tiny model listens locally for one phrase. Nothing leaves the device until that phrase fires.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
Wake word models run on-device on a very small compute and power budget, trained to recognize a single phrase while rejecting all other audio. Their performance is described by a false accept rate and a false reject rate, which trade against each other.
False accept vs false rejectindustry rangeWake word engineering convention, 2026 - A design description rather than a measurement. Wake words are usually multi-syllable and phonetically distinctive precisely to keep the false accept rate down.
Silero VAD scores a 30 millisecond audio chunk in under 1 millisecond, showing that lightweight, always-running audio processing is computationally cheap. Wake word models operate in a similar efficiency regime.
Under 1 ms per 30 ms chunkvendor claimSilero VAD, 2026 - Reported by the project, and cited here only as context for how little compute an always-on listener needs, not as a wake word benchmark.
There is no published independent benchmark for wake word false accept and false reject rates across commercial implementations, and vendors do not publish comparable figures.
-no reliable figureThe two error rates are the whole performance story, and neither can be looked up across products. Any comparison has to be run on your own device and phrase.
There is no reliable public data on how often devices activate accidentally in real household conditions, despite accidental activation being the central privacy concern.
-no reliable figureThe absence is notable, because the one number the public most wants is the one nobody publishes. State that plainly rather than borrowing an anecdote.
Examples
In practice
A business asks for its two-syllable brand name as a custom wake word. Testing turns up an unacceptable false accept rate, because the name sits phonetically close to everyday speech. Extending the wake phrase with a preceding word makes it longer and more distinctive, and the false accepts drop to an acceptable level without hurting detection.
The everyday version
A wake word is the name you say to a smart speaker to get its attention. A tiny program on the device listens for that one phrase, ignores everything else, and sends nothing anywhere until it hears it. A voice widget on a website works the other way around: nothing listens until someone taps it and the browser asks permission first.
Usage
Who says it
- Device and embedded audio engineers.
- Privacy specialists, in relation to always-listening products.
- Consumers, though usually by naming the specific phrase rather than the concept.
Where it turns up
- On a spec sheet it sits next to the activation method, on-device processing, false accept rate, custom wake word support and microphone permission handling.
- For always-listening hardware it is where the privacy questions cluster, which is exactly where on-site voice has nothing to answer.
Common misuse
- Claiming a device records continuously, which misdescribes how the design actually works.
- Assuming any short brand name will function as a wake word.
- Applying always-listening privacy concerns to click-activated or gesture-activated voice, where they do not apply.
Questions people ask
Do smart speakers record everything you say?
No, and the design specifically avoids it. A small model on the device listens locally for one phrase and rejects everything else, and nothing is transmitted until the wake word fires. Public skepticism is understandable given past incidents of accidental activation and human review of recordings, but continuous recording is not how the technology is built.
What is the difference between a wake word and hotword detection?
A wake word is the phrase itself, the name you say to get a device's attention. Hotword detection is the always-running, low-power process on the device that recognizes that phrase and ignores all other audio. One is the trigger, the other is the mechanism that listens for it.
Can I use my brand name as a custom wake word?
Sometimes, but it carries real risk. Short or common phrases produce high false accept rates, so wake words need to be multi-syllable and phonetically distinctive. If a brand name is too close to everyday speech, extending it with a preceding word makes it longer and more distinctive and usually brings the false accepts down without hurting detection.
Does a website voice widget have a wake word?
No. A voice widget activates on a click or a behavioral trigger, and the microphone only opens after an explicit user gesture and browser permission. There is no continuous listening and no always-on model, which removes the always-listening privacy question entirely.
Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.