Analyzing sound data

[~Spaceaholic]

I am working on android devices, but I do not think this questions needs to be android specific so i am posting it here...
I would like to detect a user blowing into the microphone of my phone so I can create something similar to the candle apps that let you blow them out.
currently I have captured the audio and get the max amplitude for a given sample, if the amplitude is high enough I consider it a blow. but this works the same for any loud noise. how would I go about filtering out everything but a blowing noise on the mic? Thanks.

[aazimon]

You may want to add a maximum amplitude, to cut out sounds that are too loud. Another thing would be to look at the sound's wave pattern. (You should be able to view that through any decent sound editor), and then try to match it with the sound coming in from the mic. You will need to have a variance, in which close enough is good enough.
This is just a theory. I hope it helps. 

[philfrei]

I think you are going to have to get into some fancy filtering algorithms. Most loud noises will have some strong pitches in them, or sharp amplitude peaks and valleys. The puff of air sound is more like white noise (and might not be distinguishable from someone speaking "shhh" or "ffff" or "sss").

Have you looked into the DFT (discrete fourier transform) or Fast Fourier algorithms? This is not a trivial problem, computationally.

Games published by our own members! Check 'em out!

[krasse]

My first approach would be to use a couple of bandpass filters that are reasonably spread out over the frequency.
If all of the filters get "sufficiently much" simultaneous signal energy out of them when you send in your data, you can assume that you have a wide energy spectrum (which is one of the characteristics of white noise).

If that wouldn't work, I would use a DFT like philfrei suggested. Note that you have to split up the data in smaller chunks where you smooth the edges. The bandpass filter approach is probably cheaper though but if you use many filters it may not.

You can probably use the filter part of a vocoder as well, if I am not mistaken

[Nate]

As mentioned, blowing into the mic will look to an FFT like significant energy is produced at many different frequencies, varying rapidly. Note libgdx has code to do FFTs.

[nsigma]

My first approach would be to use a couple of bandpass filters that are reasonably spread out over the frequency.

If you're looking for bandpass filter code, you could try this class - http://code.google.com/p/praxis/source/browse/audio.ops.impl/src/org/jaudiolibs/audioops/impl/IIRFilterOp.java  It's ported from Gervill / OpenJDK to this interface in order to remove any other dependencies.  Type.BP12 is the bandpass.

[~Spaceaholic]

Thanks for all the responses
I would like to try and use the FFT which will require that I convert my audio file into a float[], any pointers as to how I should go about doing this?

[krasse]

http://www.jsresources.org/faq_audio.html#reconstruct_samples