How do voice assistants process commands?

12 Tem 00:42

Voice assistants basically split the work between local chips and the cloud. Take Siri or Google Assistant for example—they pre-process on-device with neural engines (like Apple’s A13/A14/A15/A16 Bionic or Qualcomm’s Hexagon in Snapdragons) to do a quick keyword spotting and noise suppression. That first 100–150 ms latency you notice is usually the local step; it filters out the “Hey Siri” so the rest of the query can be burned off to the cloud without wasting bandwidth. I’ve seen off-line commands hit ~92 % accuracy for native speakers on my Pixel 6, but once you drift into broad Glaswegian or thick Boston accents you’re talking mid-80 % unless you’ve trained a profile. Accuracy for non-native or heavy accents depends on how much the cloud model has ingested. If you drop a phrase like “set an alarm for 5:30 AM” it’ll still pick up the numbers and time intent, but plural nouns (“five thirty”) can poison it. For accents I’ve tried an easy trick: before speaking, switch the assistant language to the one that matches your phonetic roots—e.g., set Assistant to English (UK) when you’re rocking a Scouse burr; the cloud model then weights the linguistic pipeline toward British English acoustics. You won’t hit 99.9 %, but you’ll claw back 3–5 % points without extra cloud cycles.