The acoustic features of speech sounds in a model of auditory processing: vowels and voiceless fricatives

The acoustic features of three classes of complex sounds (complex tones, vowels and voiceless fricatives) were analyzed using a model of auditory signal processing. The model consists of a peripheral cochlear component followed by two central neural networks. At the peripheral stage the asymmetrical shape of the cochlear filters, in combination with the preservation of the fine-temporal structure of their outputs, provide for a robust spatio-temporal representation of speech sounds. The cochlear patterns are subsequently processed by two separate layers of lateral inhibitory networks (LINs) in order to extract perceptually significant features of the input signal. For speech-like signals the LIN output emphasizes the spectral components in the region of the formant peaks. The LIN patterns generated in response to vowels spoken by male and female speakers contain some variability, particularly with respect to the location of formant peaks. However, the relative amplitudes of the LIN peaks (or, more precisely, the weight distribution of the LIN patterns) provide a more stable representation of each of the major vocalic classes. With respect to the voiceless fricatives, the model suggests that the most distinctive acoustic feature is the location of the high-frequency edge of the signal spectrum.