I'm not sure either, but I want to say that since the sound will be at a given frequency in a given pattern, you could probably pick it out relatively reliably...
Also, unlike fire alarms, I think I'd be less worried about false positives or the thing not working -- it would be an annoyance, not something that could potentially endanger my life.
Unsupervised learning to extract high level features might work. Then get examples of the kinds of sounds people are training it on and use that to extract even better features.
Also, unlike fire alarms, I think I'd be less worried about false positives or the thing not working -- it would be an annoyance, not something that could potentially endanger my life.