Speech Enhancement Quality Assessment Based on Aspect-Specific Qualities: A Preliminary Analysis

We propose a methodical framework to develop computational tools for assessing the quality of enhanced speech signals. The central building block of this framework are aspect-specific subjective quality ratings obtained in a listening experiment. We show that mean aspect-specific subjective quality ratings predict overall quality significantly better than objective features of state-of-the-art quality assessment tools. These experimentally obtained aspect-specific subjective quality features can be utilized to determine and tune objective features that predict them and thus indirectly predict overall quality.