Continuously listening for a wake word on the main A-series application processor would drain the battery in hours. So Apple splits the job across two chips with wildly different power budgets. On iPhone 6S and later, the Always On Processor (AOP) — the same low-power embedded motion coprocessor that counts your steps — taps directly into the microphone signal and runs a tiny acoustic model day and night. That first-pass detector is a deliberately small deep neural network: 5 layers of just 32 hidden units. It slices incoming audio into roughly 0.01 second frames, converts each into a spectral feature vector, and outputs a probability distribution over speech sounds. An HMM decoder stitches those frame-by-frame predictions together with dynamic programming into a single confidence score for the phrase. The clever part is what it does NOT do. It does not understand language. It does not transcribe. It answers one question cheaply: did that sound like “Hey Siri”? When the score crosses...