Statistically Optimum Discrimination of Speech Spectra