How does a voice assistant decide you said its name?
A voice assistant decides whether a sound it hears is its name, called the wake word. It compares the sound with the pattern it is listening for and gives it a match score. If the score is over a line, it wakes up and starts recording. Sounds that are close to its name can wake it by mistake.
In one line: A program in a speaker or phone that listens for its name, then records what you say next and turns it into words.
Last reviewed: 25 October 2026 3 min read
What decision does it make?
Before it does anything else, a voice assistant makes one yes-or-no choice, over and over: was that my name? Only after a yes does it record what you say next and work out the words.
| Decision | Possible answers |
|---|---|
| Did someone just say my name? | Yes: wake up and start recording. No: keep waiting. |
What does it look at?
- The sound pattern. It listens for one pattern of sounds, the wake word, and checks how closely each new sound matches it.
- How close the match is. A sound that is only a little like its name gets a lower score than a clear, close match.
- The words after its name. Once awake, it records what you say and works out which words you most likely said.
How does it learn the rule?
- Collect examples. People gather many recordings of the name being said, and many recordings of other speech and sounds, such as a TV.
- Learn the pattern. The program learns which sound patterns belong to the name and which do not.
- Score new sounds. Every new sound gets a score for how much it sounds like the name.
- Set the line. People choose how high the score must be to wake up, then check the mistakes and adjust.
Where can it go wrong?
- It wakes up by mistake. A word that sounds like its name, even from a TV show, can make it start recording.
- It misses its name. A real call can still score under the line, so it is ignored.
- No line is perfect. Move the line down and it wakes up by mistake more often. Move it up and it misses more real calls.
Try it
This toy assistant is named Pip. Move the line so it wakes up for its name and stays asleep for everything else. Can you get zero mistakes?
Toy example: ten sounds, the made-up score a voice assistant gave each one, and whether it really was its name.
| Item | Values | Right bin |
|---|---|---|
| "Hey Pip!" said clearly | Sounds-like-its-name score: 95 | Wakes up |
| "Hey Pip" across a noisy room | Sounds-like-its-name score: 75 | Wakes up |
| "Hey Pip, what time is it?" with music on | Sounds-like-its-name score: 70 | Wakes up |
| "Hey, pick it up!" | Sounds-like-its-name score: 65 | Stays asleep |
| A TV character says "Hey Kip" | Sounds-like-its-name score: 60 | Stays asleep |
| "Hip hip hooray!" | Sounds-like-its-name score: 50 | Stays asleep |
| "Pip" said softly in the middle of a yawn | Sounds-like-its-name score: 45 | Wakes up |
| "Hey Pip" whispered from far away | Sounds-like-its-name score: 40 | Wakes up |
| "Pass the dip" | Sounds-like-its-name score: 35 | Stays asleep |
| A dog barks | Sounds-like-its-name score: 10 | Stays asleep |
Check it
What does a voice assistant decide first?
Answer: Whether a sound was its name. It keeps checking one thing: was that its name? Only then does it wake up.
You move the line lower. What happens?
Answer: It wakes up by mistake more often. A lower line lets more sounds through, including ones that only sound a bit like its name.
Why might it miss its name?
Answer: The sound scored under the line. A real call can still get a score under the line. The machine has no feelings.
For older readers
Spotting the wake word is a classification problem: a model trained on labeled audio outputs a score, and a decision threshold turns it into wake or stay asleep. Waking for the wrong sound is a false positive (often called a false accept); missing the real name is a false negative (a false reject). After waking, speech recognition turns the recording into words, using pattern classification together with a language model that estimates which word sequences are likely.
Questions
It listens for one sound pattern, its name. When it hears something that sounds like its name, it wakes up and records what you say next.
Voice assistants usually send recordings to the company that made them. A grown-up can check the settings to see, delete or stop saving recordings.
Something it heard scored over the line, such as a word on TV that sounds like its name.
Many devices have a button or switch that stops the device from listening. Ask a grown-up to show you.
Sources
- FTC: How to secure your voice assistant and protect your privacy (opens another website)
- University of Washington CSE 416: Classification (confusion matrix and changing the threshold) (opens another website)
- University paper: wake word detection, false accepts and rejects (opens another website)
- MIT OpenCourseWare: Automatic speech recognition, syllabus (opens another website)