SIDIQ: Computational Quality Assessment of Enhanced Speech Based on Auditory Figure-Ground Segregation, Similarity, and Disturbance

This paper presents a new method to predict the subjective quality of enhanced speech called Similarity Disturbance Quality (SiDiQ). SiDiQ is based on models of auditory figure-ground segregation, similarity to the reference stimulus, and disturbance by background sounds. The models were parameterized in a data-driven way, using subjective ratings on overall quality and single quality aspects. We evaluated SiDiQ on a test dataset and compared it to other quality assessment methods. Despite not reaching the accuracy of PESQ, SiDiQ outperformed simple quality metrics and some of the more sophisticated quality assessment methods. The results show that computational quality assessment based on subjective ratings on single quality aspects is a promising methodical framework.