Automatic discrimination of fricative consonants based on human audition

A phoneme discriminator, designed to model the human auditory system, has been tested in a fricative discrimination task. Nineteen speakers, ten male and nine female, generated a data set comprising the nine isolated fricative consonants, each followed by three different vowels. The system was first trained on the utterances of nine of the voices and achieved 89% correct classification of the nine fricatives. Then this classification system was applied to an "unknown" set of utterances of the remaining ten voices. This prediction experiment yielded 74% accuracy. The system design, based on auditory processing, involves a spectral analysis by means of a bank of 1/3- octave bandpass filters. Acoustic features derived from this spectral analysis include voice onset time, time averaged spectra, and gross spectral energy distributions.