# Always On Processor **Entity class:** Low-power processor / wake-word inference ## Definition and history Apple's Always On Processor executes the first stage of an on-device neural voice trigger while the main processor remains asleep. [[articles/Intelligence Inside|Intelligence Inside]] uses Apple's engineering account to distinguish a small, continuous acoustic model from the later stages that wake and rescore. The early network maps acoustic input to probabilities for sound classes and uses a dynamic threshold to decide whether to escalate. Model updates supplied from servers belong to a different timescale from instant local inference. ## Mechanism and limits The architecture partitions observation, thresholding, and heavier processing across power domains. It therefore makes a continuously available capability practical on a battery-constrained device. Continuous local recognition should not be conflated with continuous upload of every sampled sound; the published account specifies the wake cascade, while data movement requires its own evidence. ## Relationships **Engineered by:** [[wiki/Apple|Apple]]. **Function:** [[wiki/Inference Layer|Inference Layer]] and [[wiki/Decision Threshold|Decision Threshold]]. **Comparison:** [[wiki/Intel Gaussian and Neural Accelerator|Intel Gaussian and Neural Accelerator]]. **Article route:** [[articles/Intelligence Inside|Intelligence Inside]]. ## Sources [Apple Machine Learning Research, *Hey Siri: An On-device DNN-powered Voice Trigger*](https://machinelearning.apple.com/research/hey-siri); [[articles/Intelligence Inside|Intelligence Inside]].