Extracting vowel characteristics from smoothed spectra