Non-intrusive Speech Intelligibility Prediction of Speech with Additive Noise and Reverberation Using Multiple Deep Learning-Based Speech Enhancement

In the field of non-intrusive speech intelligibility estimation, we have proposed a model that uses speech enhancement as an auxiliary signal for prediction. In this study, we performed intelligibility estimation using multiple speech enhancers, not just a single one, with the expectation that combining multiple enhancers would allow for error interpolation in the presence of various degradation factors such as additive noise and reverberation. As a result, using multiple speech enhancers improved the estimation accuracy compared to using a single model, and we were able to achieve a maximum correlation coefficient of 0.674 for estimation.