The Perception of Spectrally Reduced Prevocalic Stop Consonants

The present paper investigates the recoverability of detailed spectral features such as formant and burst peak frequencies of prevocalic stop consonants in noise excited channel vocoded speech for bandwidths ranging from 250-2000 Hz. It is shown that some formant frequency information is still recoverable up to 1000 Hz. This challenges the claim that listeners must necessarily rely primarily on temporal or global spectral features because such narrowly localized spectral properties are lost in noise vocoded speech. Listeners’ categorization of these stimuli were modelled using empirically measured formant and burst peak frequencies as well as mel-frequency cepstral coefficients in order to evaluate the relative importance of each as correlates to place of articulation perception. It is found that despite a certain level of recoverability for detailed spectral features, spectral shape at burst and voicing onset are better correlates to listeners’ perception of stop consonant place of articulation.