Always-on sensing

Also known as: always-on AI

On this page 7

Always-on sensing is the practice of keeping selected sensors and a very small inference engine permanently active on a device, so that it can recognize relevant events such as a wake word, a gesture or a face while the main processor sleeps. In Edge AI, it is the lowest tier of a power hierarchy: a milliwatt-class detector runs continuously and wakes larger, hungrier models only when something worth processing happens.

The approach is also called always-on AI. It underpins voice assistants, fitness trackers, smart doorbells and industrial condition monitors, and it depends on the same model compression techniques used elsewhere in on-device artificial intelligence.

How always-on sensing keeps the device listening

Always-on sensing works by splitting perception into cascaded stages of increasing cost. The first stage runs on a dedicated low-power domain, often a sensor hub, DSP or microcontroller core, that samples a microphone, accelerometer or low-resolution camera and runs a tiny classifier on short windows of data. Only a positive detection raises an interrupt that powers up the application processor, the GPU or an NPU for heavier work.

The economics rest on duty cycling. A main system-on-chip may draw hundreds of milliwatts to several watts when active, while the always-on island typically draws from tens of microwatts to a few milliwatts. Because real events are rare, keeping the expensive silicon asleep for more than 99 percent of the time is what makes all-day or multi-year battery life possible.

The following table shows a typical cascade in a smartphone or smart speaker.

StageHardwareTypical powerTask
Signal gateAnalog comparator or sensor FIFOMicrowattsDetect energy or motion above a threshold
First-stage detectorDSP or microcontroller coreAbout 0.1 to 5 mWKeyword spotting or activity classification
Verification stageLarger DSP or NPU sliceTens of mWConfirm the event with a bigger model
Full processingApplication processor, NPU or cloudHundreds of mW and aboveSpeech recognition, assistant logic, upload
Typical stages of an always-on sensing cascade

Sensors and workloads that run continuously

Always-on sensing is applied to any modality where events are sparse but latency matters. Audio and inertial data dominate because their sample rates are low and their features are cheap to compute, whereas vision requires aggressive reductions in resolution and frame rate before it fits an always-on budget.

SensorTypical workloadInput sizeModel scale
MicrophoneWake word and keyword spotting1 s of 16 kHz audio as spectrogramTens to hundreds of KB
Accelerometer and gyroscopeStep counting, fall and gesture detection50 to 200 Hz windowsA few KB to tens of KB
Low-resolution cameraVisual wake words, presence detection96 x 96 to 160 x 120 pixelsAbout 100 to 300 KB
Environmental sensorsAnomaly or threshold detectionSlow time seriesRules or tiny models
Vibration sensorMachine condition monitoringSpectral features of short burstsTens of KB
Common always-on workloads by sensor type

Models at this tier are built with the toolset described under TinyML: INT8 quantization, pruning, compact architectures such as depthwise separable convolutions, and careful feature extraction. Memory is often the binding constraint, with total SRAM between 64 KB and a few MB shared by weights, activations and audio buffers.

Always-on sensing is a deployment pattern defined by its power envelope, not a model type. It overlaps with several neighboring concepts, and the distinctions matter when engineers choose hardware and budgets.

ConceptPrimary goalRelation to always-on sensing
TinyMLRun ML on microcontrollersSupplies the methods; always-on is its most common use
Edge inferenceRun models locally instead of in the cloudBroader category; always-on is its continuous, lowest-power case
Keyword spottingDetect specific spoken wordsA typical first-stage workload
Visual wake wordsDetect presence of a person in an imageThe vision counterpart of a wake word
Hybrid inferenceSplit work between device and cloudOften the final stage after an on-device trigger
How always-on sensing relates to neighboring concepts

A practical way to separate the terms is to ask whether the model runs when nobody is interacting. Ordinary edge inference starts on demand, for example when a user opens a camera app; always-on sensing runs regardless and decides on its own when to involve the rest of the system.

Designing an always-on pipeline

Good always-on design starts from an energy budget per day rather than from model accuracy. Engineers estimate the active and sleep currents of each stage, the expected event rate and the false trigger rate, then size the first-stage model so that the average power stays within the battery target.

  • Tune the false accept rate of the first stage carefully, because every false trigger wakes expensive hardware and erodes battery life.
  • Keep a ring buffer of recent audio or sensor data so the next stage can see what happened just before the trigger.
  • Compute features in hardware where possible, for example filter banks in the DSP, to avoid waking a general-purpose core.
  • Validate in realistic noise and motion conditions, since laboratory accuracy rarely survives a kitchen, a car or a factory floor.

The decision table below summarizes how common requirements map to architectural choices.

RequirementRecommended choiceReason
Multi-year coin-cell lifeAnalog or comparator gate before any modelKeeps the digital domain off most of the time
Low false wakes in noisy roomsTwo-stage cascade with on-device verificationThe second model filters first-stage errors cheaply
Strict privacy expectationsProcess and discard data locallyRaw audio or images never leave the sensor island
Many supported commandsSmall trigger plus larger on-demand modelVocabulary size grows model size quickly
Design choices for always-on systems

Limitations, risks and common mistakes

The main trade-off is sensitivity against power and false alarms. A detector tuned to miss nothing will wake the system constantly, while a conservative one frustrates users with missed commands. Both errors are usually reported as false reject and false accept rates, often expressed per hour of background audio.

Privacy perception is a second risk. Even when only a tiny local model listens and nothing is stored, users may assume constant recording, so products typically add hardware mute switches, indicator lights and clear data policies. Regulators increasingly treat continuous sensing of people as sensitive processing, which affects consent design.

  • Ignoring the idle floor: leakage of memory and regulators can exceed the model’s own consumption.
  • Oversized first stages: moving verification-grade models into the always-on island defeats the purpose.
  • Neglecting model updates: an OTA model update path is needed to fix accuracy drift without replacing hardware.
  • Testing only clean data: accents, reverberation and device placement degrade accuracy in the field.

Stripped to the essentials

Always-on sensing lets a gadget pay attention all the time without using much battery. A tiny, simple program keeps watching one sensor, such as the microphone, and only wakes the powerful parts of the device when it notices something important.

It works like a night porter in a hotel. The porter sits quietly at the desk and does not wake the manager for every noise, only for a guest who really needs help. The cheap watcher handles the waiting, and the expensive brain works only when needed.

Where it is applied

A maker of wrist-worn fitness trackers needs fall detection for older users without charging the device every night. The engineering team places a gesture and fall classifier of about 20 KB on a low-power sensor hub that reads the accelerometer at 50 Hz, while the main processor sleeps most of the day.

When the classifier sees a pattern resembling a fall, it wakes the main processor, which runs a larger verification model over the buffered ten seconds of motion data and checks whether the wearer gets up. Only a confirmed fall triggers a vibration prompt and, if unanswered, an alert to a caregiver’s phone.

In field trials, this cascade keeps the average sensing power below one milliwatt, so the tracker reaches roughly a week of battery life while monitoring continuously. False alarms drop sharply compared with a single-stage design, because the second model rejects most everyday movements such as sitting down quickly.