Robust speaker-independent speech recognition using non-linear spectral subtraction based IMELDA