Correlation Based Glimpse Proportion Index

The glimpse proportion (GP) index is an objective intelligibility measure (OIM) based on the glimpse model of speech perception in noise. GP uses local SNR as a criterion to identify time-frequency (TF) regions, or glimpses, that are dominated by speech. Although GP has demonstrated high performance in predicting intelligibility in the presence of stationary and fluctuating noise, its application is limited to additive noise conditions. To address this drawback, we propose a correlation based GP (CGP) index that operates in the TF domain similar to GP but can be applied to a wider range of conditions. The proposed measure is optimized and evaluated using 16 subjective datasets involving speech corrupted by modulated noise, nonlinear processing, and reverberation. The results show that CGP has consistent high performance across all degradation conditions and, on average, outperforms several baseline OIMs. Additionally, CGP has low complexity and takes substantially less time to execute compared to baseline OIMs.