No. 196Technology3 minviews 75
Published by GK Library Annex
Your phone wakes up at "Hey Siri." Is it always listening?
Your phone keeps its microphone on, but a tiny recognizer inside the device first filters for the wake phrase alone
To catch a wake phrase like "Hey Siri" ("시리야" in Korean), your phone is always listening to its microphone. At this stage, though, only a very small recognizer that looks for that one short phrase runs inside the device. Apple described this process in detail on its research blog in 2017.
There is a separate tiny ear
According to Apple, the iPhone's microphone turns incoming sound into numbers 16,000 times a second. The phone looks at this sound 0.2 seconds at a time and gives it a score for how much it resembles the wake phrase.
Doing this all day would drain the battery fast. So starting with the iPhone 6s, a low-power helper chip that stays awake even when the screen is off handles this small recognizer. Meanwhile the main processor, the phone's big brain, gets to rest.
It checks twice before waking
Only when the helper chip's score passes a threshold does the main processor wake up and listen once more with a larger recognizer. Siri turns on only if the sound passes this check too, and then the sound goes to Apple's servers, where the full work of understanding it begins.
If the large recognizer on the server listens again and finds it was a similar phrase such as "Hey Seriously" rather than "Hey Siri," it sends the phone a cancel signal and puts it back to sleep.
Why it sometimes turns on by mistake
Because this system decides by score, it is not perfect. Set the threshold too high and it misses you even when you really call it; set it too low and it turns on at TV sounds or similar-sounding words.
Apple added a second, lower threshold in between. When the score comes out borderline, the phone listens more sensitively for a few seconds, so it switches on easily if the user says the phrase again. Having you repeat the wake phrase several times during setup is also a step to tune it to the owner's voice and cut down on mistaken activations.
What about the worry of eavesdropping?
A common misconception is that the phone records every conversation and sends it to a server. By Apple's description, finding the wake phrase is the job of the small recognizer inside the device, and sound goes to the server only after passing several stages.
Still, if it switches on by mistake at a similar phrase, the sound from that moment can reach the server too. If this bothers you, you can turn off listening for the wake phrase in settings and keep only the option of calling Siri by pressing a button.
Your phone is always listening, but a tiny recognizer inside the device first filters for the wake phrase alone.
Notes and references · Apple Machine Learning Research, 2017. "Hey Siri: An On-device DNN-powered Voice Trigger" explains the 16,000 samples per second, the low-power helper chip, the two-stage check and the server cancel signal. It reflects 2017, so details on newer devices may differ
Editorial standardsReport an error
Quick quiz