Fricative discrimination based on human audition

As a continuation of our study of phoneme recognition based on human audition, we report on an attempt to recognize fricatives. Our system uses a bank of 13‐octave bandpass filters to model to a first approximation the tuning curve data of Kiang, as well as the critical band data, over the speech frequency range. Five hundred and ten isolated fricative‐vowel pairs, spoken by 19 different speakers: both male and female, were spectrally analyzed with the 13‐octave system. Following spectral analysis 54 spectral features were quantified for each utterance. These features included voice onset time, time averaged spectra, and gross spectral energy distributions. The features from half the data set were used to train a decision algorithm, discriminant analysis. The system achieved 89% correct overall classification of the training set data. This classification system was then applied to an “unknown” set of utterances of ten different voices. This prediction experiment yielded 77% overall accuracy. In order to p...