Wake Word

A wake word is a specific word or short phrase that a device listens for. Well-known examples include “Alexa,” “Hey Siri,” or “OK Google.” As long as this word isn’t spoken, the speaker or phone doesn’t react to what’s being said in the room. For this to work, the microphone does listen continuously, but it only checks a single question: Did this one word just occur? You can think of it like a guard who waits for a password and ignores everything else. Only after the password does the actual work begin.

What the wake word has to do with privacy

Many people distrust voice assistants because a microphone is switched on around the clock. The wake word is manufacturers' technical answer to this distrust. The idea: the continuous check happens directly on the device, without an internet connection. Only once the wake word has been recognized does an audio recording go to the provider’s data centers.

In practice, this doesn’t work perfectly. Devices regularly trigger on similar-sounding words, such as “Alexander” instead of “Alexa.” Such false triggers are known in technical jargon as False Accepts. Studies have shown that common assistants start up accidentally several times a day and then transmit snippets of conversation. Conversely, there are False Rejects: the device doesn’t respond even though it was called.

For companies, the wake word is also a piece of brand policy. Whoever turns their company’s name into a wake word gets millions of people to say that name every day. That’s why wake words are rarely freely choosable, but instead limited to a few variants predetermined by the manufacturer.

How a device recognizes its own wake word

A very small computational model runs on the device, with only one task. It breaks the sound down into short snippets of a few milliseconds. For each snippet, it calculates how strongly it matches the sound pattern of the wake word. If a matching pattern emerges across several snippets, the wake word is considered recognized.

This model is tiny compared to the systems that later understand the actual question. It often needs less than a megabyte of storage and runs on a cheap chip. That’s by design: it has to work continuously without draining the battery. Running a large language model continuously would be technically and energetically impossible.

One detail surprises many people: most devices store the audio from the last one to two seconds in a small buffer. This way, the beginning of the sentence isn’t lost if you keep talking right after the wake word. This buffer is continuously overwritten as long as nothing is recognized. But it is precisely this buffer that’s the reason why, in the case of a false trigger, words spoken before the supposed wake word also get transmitted.

From smart speaker to car

Wake words are best known from speakers like the Amazon Echo or Google Nest. But smartphones, headphones, TVs, and cars work with them too. In newer vehicles, a wake word is increasingly replacing buttons for navigation or the air conditioning. Even some kitchen appliances and fitness watches now listen for a fixed command.

In the news, the term usually comes up in connection with privacy disputes. It has repeatedly come to light that manufacturers' employees listened to accidentally recorded conversation snippets in order to improve recognition. Such reports have led to fines and changed default settings. With most assistants today, you can check in the settings when the device was triggered, and delete recordings.

A common misconception is that a wake word is the same as speech recognition. It’s only the door opener. The conversion of speech into text and the understanding of the request happen only afterward, usually on remote servers. Anyone who keeps these terms separate will also better understand why a device without internet can wake up, but still can’t respond.