The first nonlinear technique is used for reducing wideband noise in speech signals. This type of noise includes: magnetic tape hiss, electronic noise in analog circuits, wind blowing by microphones, cheering crowds, etc. Linear filtering is of little use, because the frequencies in the noise completely overlap the frequencies in the voice signal, both covering the range from 200 hertz to 3.2 kHz. How can two signals be separated when they overlap in both the time domain and the frequency domain?http://www.dspguide.com/ch22/7.htm
Here's how it is done. In a short segment of speech, the amplitude of the frequency components are greatly unequal. As an example, Fig. 22-10a illustrates the frequency spectrum of a 16 millisecond segment of speech (i.e., 128 samples at an 8 kHz sampling rate). Most of the signal is contained in a few large amplitude frequencies. In contrast, (b) illustrates the spectrum when only random noise is present; it is very irregular, but more uniformly distributed at a low amplitude.
Now the key concept: if both signal and noise are present, the two can be partially separated by looking at the amplitude of each frequency. If the amplitude is large, it is probably mostly signal, and should therefore be retained. If the amplitude is small, it can be attributed to mostly noise, and should therefore be discarded, i.e., set to zero. Mid-size frequency components are adjusted in some smooth manner between the two extremes.
Due 'd definitely do this process (of noise reduction) perfectly well , but someone with low budget can get a flavor of DSP with standalone Atmega328p also. Especially when narrow-band interference is an issue, than pitch-shifting example could cut-out such annoyance easily.