Auditory Processing of Speech

We believe in our group that the only way to describe human speech is to describe not the perception itself but the artificial speech understanding system that is most compatible with the experimental data obtained in speech perception research. For this purpose a multichannel model comprising the cochlea has been built as an analog system. This model is able to detect quick changes in spectrum and intensity. The detection of such irregularities in the speech-induced excitation pattern seems to be an essential part of speech processing. As for spectrum-shape processing, experiments seem to indicate that spectral peaks are extracted at a lower level, whereas an integration takes place at a higher level if the distance between peaks does not exceed 3.0-3.5 Bark.