Improved intelligibility prediction in the modulation domain

sEPSM corr is a recently proposed intelligibility measure that combines a modulation domain auditory model with a correlation similarity metric inspired by the STOI measure. In this paper, we aim at improving the overall performance of sEPSM corr , particularly in modulated noise conditions. We propose to improve its auditory model and use a segmentation scheme and similarity metric adapted from the extended STOI measure that works better than STOI with modulated noise. Performance evaluation against subjective data shows the modified measure is able to predict intelligibility much better than the original version in conditions involving modulated noise, nonlinear processing, and reverberation. The modified sEPSM corr also has better overall performance than several baseline intelligibility measures, especially in modulated noise conditions. Additionally, we study the contribution of the proposed modifications to performance by evaluating variants of the modified measure where subsets of the modifications are in effect.