Similarly glass-break detectors for home security systems have done this for years. The advantage of an ML model is that you can do multi-class prediction for

Similarly glass-break detectors for home security systems have done this for years. The advantage of an ML model is that you can do multi-class prediction for a sound clip with somewhat arbitrary complexity and the cost of execution is more or less the same even if you add an extra class or two. It's just an extra row in your output prediction vector. By complexity I mean signals that don't have an obviously characteristic spectrum, like glass smashing. A typical CNN backbone is capable of classifying hundreds if not thousands of classes with high accuracy - even the edge architectures. Always-on detection tends to use very compact networks (kB in size) that will run on low-power ARM cores, or even specialist ASICs, but even so 20-30 common audio event types seems very feasible. For people worrying about sending data back, there's no reason why you'd do this off-device. The exception might be that you feedback to Google that there was a false alarm, so they can use your sound clip as a negative training example. Just a guess there, but Tesla does this extensively for Autopilot - they deploy models in your car and specifically ask it to capture images of rare events (Andrej Karpathy gave an example of tree-occluded stop signs).