The perception of noise vocoded prevocalic stop bursts and formant transitions

Abstract

Theories of prevocalic stop consonant perception differ with respect to the amount of spectral resolution required by the model, e.g., gross spectral shape theories predict that human speech perception is more robust against decreased frequency selectivity where detailed cues such as vocalic formant transitions break down. Previous results in the perception of noise vocoded speech [M. Kiefte, Int. Conf. Phon. Sci. (1999)] showed that, although listeners’ performance worsened significantly at bandwidths as low as 500 Hz, a spectral shape model using cepstral coefficients was better able to predict subjects’ responses. However, a simulation which showed that release bursts require much less spectral resolution than vocalic formant transitions complicates the interpretation. Therefore, in this experiment, subjects were asked to identify the place of articulation for gated release bursts independently of vocalic portions for both unprocessed and noise channel vocoded speech. Preliminary results show that correct identification of stop place from only the release burst does not change significantly up to a channel bandwidth of 2000 Hz. Results suggest that responses from the previous experiment can be explained by the robustness of the burst in those stimuli. Gross versus detailed cue models are also evaluated with respect to responses to formant only stimuli. [Work supported by SSHRC.]