Intelligibility Prediction of Single-Channel Noise-Reduced Speech

Abstract:
In general, single-channel noise-reduction algorithms do not improve the speech intelligibility for normal-hearing listeners. A reliable objective intelligibility measure is therefore of great interest. It could be used for analysis and/or optimization of noise-reduction algorithms. For these applications it is important that the objective measure can correctly predict the difference in intelligibility before and after noise reduction. Typically, existing studies do not evaluate objective measures for this property. Twelve objective measures are evaluated in order to let them predict the intelligibility before and after noise reduction. Best performance was obtained with a recently developed intelligibility predictor called STOI. Modest results were obtained with WSS and NSEC. The remaining measures significantly overestimated the intelligibility of the noise-reduced speech.