Filtering of spectral parameters for speech recognition