Hearing aids were the first device that could not afford a cloud round trip. Security cameras are the next one.

It means the AI that classifies the soundscape and enhances speech runs on a chip inside the hearing aid itself, not on your phone and not in a cloud datacenter. Starkey's Edge AI line, for example, ships with a G2 Neuro Processor that has a fully integrated NPU (neural processing unit) on-die, dedicated to running deep neural networks for noise reduction and speech enhancement in real time. The 'edge' in edge AI is not a marketing term here. It is a physical place: the silicon behind the receiver, behind the wax guard, behind your ear.

Because of latency physics. The human auditory system fuses the direct acoustic path through your ear canal with the processed path from the hearing aid; if those two paths arrive more than roughly 10 milliseconds apart, you perceive an echo or a smear that defeats the purpose of the device. A round trip from the hearing aid to a phone over Bluetooth and back is on the order of 30 to 100 ms even on best-in-class transports. A round trip to a cloud server is hundreds of milliseconds. Neither is fast enough. The DNN has to live on the same chip as the receiver, draw less than a watt, and finish each inference inside the perception window.

Three ways. First, the workload is steady-state and known: a small set of audio models running continuously, not a parade of different mobile apps each demanding their share of compute. That lets the chip designer co-locate the NPU with audio I/O so there is no round trip across a system bus. Second, the power envelope is brutal. A hearing aid runs ~20 hours on a battery the size of a watch button; the NPU must contribute less than a few hundred milliwatts. Third, there is no fallback. A phone NPU can punt to the CPU or the cloud. A hearing aid NPU cannot. If it can't classify the room in 10 ms, the user hears nothing useful.

The category is anything that has to make sense of a continuous sensor stream within a perception or action window, where shipping the stream to the cloud is impossible or unwise. Hearing aids do it for ~10 ms speech enhancement. Smart cameras (Cyrano, Wyze, Eufy, Reolink) do it for second-scale event detection on continuous video. Voice-activated wake word chips (Amazon Echo, Google Nest) do it for ~100 ms keyword spotting. Wearables doing fall detection, pacemakers doing arrhythmia classification, automotive radar doing collision-avoidance: all of them share the same constraint pattern. The latency budget changes (10 ms for hearing, 30 seconds for security action), but the architectural pressure is the same.

The same engineering constraints. A security camera doing real-time event detection cannot stream 24/7 video to the cloud (bandwidth, cost, privacy), cannot tolerate cloud latency on a developing incident, and cannot trust an internet connection at the moment a power-cut intruder kicks the panel. So it has to run inference locally. The latency budget is wider than hearing (you have ~30 seconds, not 10 ms, before a human cannot meaningfully act on an alert), but the no-cloud-round-trip rule is identical. Cyrano's edge AI device for security cameras applies the hearing aid pattern to commercial video: an NPU on a small box, fed continuous video, classifying events, no cloud streaming required.

A single $450 device that ingests up to 25 camera tiles in parallel. The novel input path: instead of pulling RTSP streams or replacing analog cameras, Cyrano plugs into the HDMI output of the existing DVR or NVR (the cable that drives the wall monitor in the leasing office). It captures that HDMI signal, splits it back into the multiview tile grid, and runs per-tile inference on each camera pane. Median capture-to-index latency is roughly 7 seconds. Every event writes a 9-field record (tile.label, tile.index, tile.coords, property, layout_id, overlay_mask, event_class, iso8601_ts, latency_ms). No DVR credentials are used. No cloud round trip happens for the inference itself.

It is slow on the speech timescale. It is fast on the human-action timescale. A hearing aid has to beat the brain's auditory fusion window, which is biological and non-negotiable. A security alert has to beat the time it takes for a monitoring agent to notice, classify, and dispatch a response (typically 30 to 90 seconds end-to-end on a well-run service). Inside that window, 7 seconds of edge classification leaves a comfortable margin. The lesson from edge AI hearing aids is not 'all edge AI must be 10 ms,' it is 'edge AI must be inside the perception or action window for its own domain.' Cyrano's window is human security response. Starkey's window is auditory fusion. Both are edge.

Yes, two. First, model size. Hearing aid DNNs are tiny (single-digit megabyte) because power is the binding constraint; security camera edge AI can spend tens to hundreds of megabytes because the appliance plugs into wall power. Second, model freshness. Hearing aid firmware updates ship monthly to quarterly because audiology is a stable domain; security event classifiers benefit from more frequent updates because the 'what does a tailgate look like' question is influenced by recorder UI changes, lighting, and seasonality. Cyrano's overlay_mask library, which lets one event schema work across 14+ DVR brands, is the security-side answer to a problem hearing aids do not have because they own the entire signal path.

It is the most prominent. Starkey shipped Edge AI in 2024 with the G2 Neuro Processor, the first hearing aid with a fully integrated NPU. Starkey claimed roughly 100x DNN throughput vs. their prior Genesis AI platform. As of January 2026, a firmware update added Auracast assistant support. Edge AI was succeeded as the flagship by Starkey Omega AI in October 2025, but the underlying NPU pattern is now industry-standard; competitors like Phonak and Oticon have parallel chips with on-device DNN paths. The pattern (integrated NPU, sub-10ms inference, no cloud) is no longer novel within hearing aids; it is now the baseline.

Five things, in this order. (1) Inference latency at the receiver, in milliseconds, not the marketing 'real time.' (2) DNN throughput on the integrated NPU, typically expressed as 'X times the prior generation.' (3) Battery life on continuous DNN load (most spec sheets quote idle). (4) Whether the device has a fallback path if the DNN fails (it should not need one, but knowing the failure mode matters). (5) Frequency of model updates and how they ship (over-the-air vs. clinic visit). The Starkey product pages cover #1 through #4 well, are quiet on #5. The cross-domain 'why edge' framing in this guide is meant to give you the vocabulary to ask the audiologist the right questions, not to replace their fitting.