Welcome to our site! EDAboard.com is an international Electronics Discussion Forum focused on EDA software, circuits, schematics, books, theory, papers, asic, pld, 8051, DSP, Network, RF, Analog Design, PCB, Service Manuals... and a whole lot more! To participate you need to register. Registration is free. Click here to register now.
There is no universal algorithm for feature extraction. The best way is to observe the signal (FFT in Your case) and select the features that are the most correlated with your classes.
If You want to get some advice, You should write more information about Your problem.
Still, there is no information what do You want to detect. Speech recognition? i.e. detect what word You say? Or just appearance of voice in the background noise?
Maybe do You want the select the voiced or unvoiced phonemes?
Basically, the number of feature that can be extracted from fft is unlimited. You can try to calculate the spectral flatness, energy at Your region of interest (or energy ration from different part of regions), or try to find continous peaks along time domain in STFT (using linear prediction or phase compensation). All this features are sensitive to some kind of tonality that can be present in Your voice record (voiced phonems)