Always-On Sensing
Always-on sensing means a device keeps listening, watching, or measuring for a small set of important events even while the rest of the product appears asleep. A voice assistant waiting for its wake word, a leak detector listening for unusual sound, or a wearable watching for a fall all need this capability without draining a battery in hours.
How it works
The device does not run its most expensive AI model continuously. Instead, it uses a low-power sensing pipeline: a microphone, accelerometer, camera, or other sensor produces a stream of data; a small model examines short windows of that stream; and the main processor wakes only when the model detects a likely event. This is like leaving a doorbell active rather than keeping the whole house fully lit.
- A wake-word model checks audio features such as short spectrogram slices for “Hey Siri” or another chosen phrase.
- A vibration sensor on industrial equipment looks for a pattern that signals a failing bearing.
- An accelerometer on a battery-powered tracker recognizes a fall, impact, or unusual motion.
Power is the central design constraint
On a microcontroller, always-on AI must fit within tight RAM, flash storage, and energy limits. Designers reduce cost through quantized models, small input windows, low sensor sampling rates, and duty cycling—briefly waking hardware to sample or compute, then returning it to sleep. A Cortex-M microcontroller running TensorFlow Lite for Microcontrollers can perform a compact keyword or motion classifier locally, without sending raw sensor data to the cloud.
Why the trigger must be dependable
The always-on model is a gatekeeper. Too many false positives wake the main processor repeatedly, wasting power and annoying users. Too many false negatives mean the device misses the wake word, safety event, or equipment fault it exists to catch. Its job is therefore not merely high accuracy: it must deliver useful detection continuously within a fixed energy budget, across noisy real-world conditions.
Always-On Sensing is continuous, low-power monitoring of sensor streams—such as audio, motion, or vibration—by an embedded device, typically using a small model to detect a trigger event. It keeps the system responsive without transmitting or fully processing data continuously. At the edge, it enables wake-word detection, activity recognition, and anomaly alerts within strict battery, memory, and compute budgets.
Think of a smoke alarm: it quietly keeps watch all day and night, ready to react the moment it detects trouble. Always-on sensing gives small AI devices a similar role. A tiny sensor can continuously listen, watch, or measure things such as sound, movement, temperature, or vibration.
The important part is that it does this while using very little power. For example, a smart speaker may listen only for its wake phrase, or a factory sensor may watch for an unusual machine rattle. Most of the time, nothing happens. When something important is noticed, the device can wake a more powerful system or send an alert.