qwertygin:
Maybe if ELM Chan's FFT Library for Arduino was used it could work?
http://arduino.cc/forum/index.php/topic,56331.0.htmlthe dull sound would have less high frequency energy compared to the high pitched frequency components of the normal sound...
correct me if i'm wrong,but sampling should also be very fast to accurately record high pitched components, but maybe...
This has potential. Offloading the spectrum analysis to the MSGEQ7 would help reduce CPU cycles spent sampling the audio and free them up for the actual fingerprint analysis. Essentially, you'd need to record as many example sounds as possible, play them into the device, note the spectrum spread, convert this into percentage values for each spectrum as compared to the others, keeping in mind volume variations and allowing a certain percentage of variation.
Test, repeat, repeat (repeat......) until you have a reasonable table of sounds to values.
It's more and more possible the more I think about it, though not easy at all. Decay of the sound will also play a big part when testing against certain sounds, as well. The MSGEQ7 has variable delay by just resetting and throwing away samples. Play with that as well... Keep the expected duration of your event in mind, and sample continuously, then go back by the appropriate amount of time and process your array then. This let's you do nothing but sample for most cycles then do the hard work only when a threshold is hit.
Have fun out there