Application of auditory image model for speech event detection

To provide an appropriate model for perception of temporal structures of speech, we applied a comprehensive computational model of the human auditory peripherals to detect changes in speech signals that potentially indicate arrivals of new events. In each tonotopic sub-band, an increase in the activation level was taken into account for the plausibility of a new event, while a decrease was ignored. The total contour obtained by integrating the sub-band information exhibited sharp peaks and dips compared to the loudness contour. A quantitative evaluation to estimate the speaking rate of natural speech also demonstrated that the event-plausibility model performs better than the loudness model.