Non-intrusive speech intelligibility estimation using deep learning with speech enhancement and convolutional layers

In a previous study, we proposed a non-intrusive intelligibility estimation method using speech enhancement for reverberated speech and showed that the accuracy of intelligibility estimation can be improved by adjusting the filter bank processing and SNR weighting parameters in the feature calculation using the enhanced speech along with the reverberated speech. However, since the selection of the optimal filter bank and the associated search for the optimal parameters are not trivial and time-consuming, we attempted to improve intelligibility estimation accuracy by performing optimal feature extraction in a data-driven manner by applying a convolution filter with a limited receptive field in the frequency direction to the spectrogram. The results of training and estimation for degraded speech and enhanced speech using a shared filter and a separate dedicated filter showed that subjective intelligibility could be estimated with the highest accuracy using separate filters with a linear correlation coefficient of 0.840, suggesting that the convolutional layer can be effective in extracting features that can lead to higher accuracy intelligibility estimation.