Macroscopic models predict the influence of energetic maskers on speech intelligibility surprisingly well. However, when comparing confusions for individual noisy speech tokens (i.e. microscopic modelling), the listener-model fit is often very poor. It is likely that incorrect allocations of foreground and background audible components is responsible for the poor fit.
Goal: This project aims to develop models of audible signal component allocation based on (i) simulations of energetic masking incorporating level adaptation and (ii) model-based decoding of foreground and background components.
Relevance: Detailed intelligibility models have numerous applications in predicting the effect of noise on listeners and in measuring the intelligibility of generated speech in realistic conditions.
Main host institution: Universidad del Pais Vasco
Second host institution: University of Sheffield
Industry partner: Oticon Research Centre Eriksholm