The moment it wakes up for no reason
You're watching a crime drama. A character says something that rhymes vaguely with "hey, portal" and the little puck on your kitchen counter lights up, ready to help. Nobody called it. Nothing close was said. And yet there it is, blinking patiently.
This isn't a glitch in any dramatic sense. It's the system working exactly as designed, just working too enthusiastically. Understanding why means understanding what's actually happening inside that cylinder when it listens.
It's not really listening to you
Here's the part most guides skip entirely: your smart speaker is not, in its idle state, sending audio to the cloud. That would be both a privacy catastrophe and a bandwidth nightmare. Instead, the device runs a stripped-down neural network locally, on a tiny processor, doing one extremely narrow job: decide whether the last fraction of a second of audio contains the wake word.
That network is trained on thousands of hours of people saying the trigger phrase, across accents, distances, and room acoustics. It learns a probabilistic fingerprint, not a perfect template. When incoming audio crosses a confidence threshold, say 85% or higher, the device wakes up and ships the audio off to a server for full processing.
The threshold is everything.
Set it too high and the speaker misses half your legitimate requests, which infuriates users. Set it too low and it wakes for a television commercial. Manufacturers tune this aggressively toward sensitivity, because a device that doesn't respond feels broken, while one that occasionally mishears just feels a little dumb. That is a business decision dressed up as an engineering one, and it's worth knowing that's what it is.
What the network is actually comparing
The audio doesn't arrive as a waveform the model reads like a sentence. It gets converted first into a spectrogram, a rolling visual representation of which frequencies are present and how loud they are at each moment. Think of it like sheet music for every sound happening at once, not just the notes you intended.
The neural network, typically a small convolutional or recurrent model with somewhere between 100,000 and a few million parameters, scans that spectrogram for patterns that match the wake word's acoustic shape. It's looking for the rough contour of formants, the vowel peaks and consonant edges that define the phrase.
Background noise complicates this in a specific way. A television running in the same room adds a constant low-level spectrogram signal. Certain combinations of speech, music, and reverb can produce frequency patterns that loosely resemble the target shape, especially in the mid-range frequencies where human voices live, roughly 300 Hz to 3,000 Hz. When those patterns align with the model's threshold, you get a false positive.
The two-stage trick, and where it breaks down
Most modern devices use a two-stage system to reduce false wakes. The first stage is that tiny on-device model, cheap to run but imprecise. If it fires, a second, more sophisticated model runs, either still on-device or in the cloud, to verify the wake word before full activation begins.
Consider two households with the same device. Maria lives alone in a quiet flat; her second-stage model almost never needs to override the first. Priya has three kids, a television on most of the day, and a partner who works from home on video calls. Her device fires the first-stage model dozens of times daily. The second stage catches most of them, but not all. Priya's device wakes up uninvited maybe three or four times a day. Maria's, almost never. Same hardware, same software, wildly different experience.
The catch: the second stage can only verify audio that's already been captured. It can't correct for the wake word being genuinely absent before the recording started. If the first stage misfires on a TV sound, the second stage gets handed ambiguous audio and has to make its best guess.
What people get wrong about this
The folk remedy that needs to die is the idea that manufacturers are secretly recording everything and the false wakes are proof. The economics and engineering don't support it. Running continuous cloud transcription on hundreds of millions of devices would be staggeringly expensive and would leak within weeks. The actual explanation is both more mundane and more interesting: the local model is just wrong sometimes, because all probabilistic classifiers are wrong sometimes.
The other misconception is that phonetic similarity is the main culprit. It's a factor, yes, but acoustic similarity in the spectrogram is what actually matters, not whether words rhyme or share syllables in written English. A phrase in a completely different language can fool the model if its spectral profile happens to match. Which, frankly, is why some devices have historically woken for phrases in languages they were never trained on at all.
Still, the technology has improved substantially over successive generations. Earlier devices had false-wake rates that made them nearly unusable in loud homes. Newer hardware often includes dedicated neural processing units and multiple microphones running beamforming algorithms, which narrow the device's effective listening cone toward the most likely speaker, improving signal-to-noise ratio before the model even sees the audio.
Microphones in formation
The microphone array is underrated. A single microphone captures sound from everywhere equally. An array of six or seven microphones, spaced around the device's circumference, can compare the tiny time differences in when sound arrives at each one, and that comparison lets the device estimate which direction sound is coming from, amplifying that direction while suppressing others.
This is less like a better ear and more like the device growing a directional attention span. A television two metres to the left gets partially filtered out. Your voice from directly above gets pulled forward. The wake-word model then sees a cleaner spectrogram, and the false-positive rate drops.
Beamforming has limits, though. It works poorly when sound reflects heavily off hard surfaces, which is why tiled kitchens and open-plan living rooms with concrete floors are harder environments than carpeted bedrooms. Reflections arrive from multiple directions simultaneously and confuse the directional estimates.
The honest trade-off
Every smart speaker ships with a sensitivity setting, sometimes a literal dial, more often buried in an app, because there is no right answer. More sensitive means fewer missed commands and more false wakes. Less sensitive means fewer interruptions from the television and more repeated "hey, [device]" in frustration.
So here's the question worth sitting with: what does it tell you about a company when they choose their default for you, without asking?
The companies making these devices have chosen their defaults deliberately, and they've chosen sensitivity. That choice reveals what they've decided is worse: a device that seems unresponsive, or one that occasionally eavesdrops on your crime drama. One of those outcomes makes you call support. The other one you mostly just tolerate.
If your device wakes up uninvited constantly, turn the sensitivity down one notch and live with it for a week. The model isn't broken. It's just optimistic.