That’s plan B if I fail. I need to determine direction of sound source that’s why I need 6 mics.
@ Mike
I don’t see the sampling time spec-ed anywhere. And you can tell the ADC to convert the cannels one after the other automatically, in a burst.
Either way, one mic or 6, ASM or C++ will be required, and will push the chip to it’s limits…
Also, my calcs assumed 6 mics, and 6 frequencies. If you want directional then you must test for each frequency on each mic, ie, 6 mics * 6 frequencues each=36 frequency calcs, WAY more than this chip can handel.
Don t bother with such complex software (netmf is not made for that). All can be done with simple electronics, an operational amplifier with a high pass filter will result in an simple analog circuit you can directly connect to your ADC input. With some calibration (linearity), you will be able to have for an example, input reading of 0 = 0hz, 3,3v = 20Khz. Then, no fft to do on the fez, just read the 6 ADC for the 6 directions, and reacts to the input value directly ?
Software is great - but sometimes electronics allow to make more and easier 
Wtith the current requirements, the Panda is not the right platform for this application. If it is/was possible, the time spent implementing it would far exceed the cost of starting over with the appropriate hardware.
Nicolas
The OP has repeatedly stated that no more external hardware can be used.
Aah… A bionic ear. Intruiging 
The Panda is not going to cut it. Matching to a known frequency is one thing but matching a
distinctive sound another.
This sound like a job for the more powerful EMX or even a step higher with an embedded PC running Windows CE or Android. Certainly Android smartphones are capable of voice and sound recognition so it should be a better platform for now.
If you want to stay with the small chips you will need external DSP help. Maybe look at a voice recognition chip. It should have no problem with specific sounds.