Improvements to Non-Intrusive Intelligibility Prediction for Reverberant Speech

Speech intelligibility prediction (SIP) allows the prediction of intelligibility without time-consuming subjective evaluation and is being actively pursued since it has become essential to constantly monitor the intelligibility of ubiquitous speech communication. We propose a non-reference SIP method by predicting clean speech from reverberation degraded speech. Speech intelligibility was predicted from the difference between degraded and estimated clean speech. We were able to predict intelligibility with Root Mean Square Error (RMSE) between true and predicted intelligibility of 0.09, and Pearson correlation coefficient of 0.75 with the proposed method. This prediction was done using the whole sentence speech, where test words were embedded in key phrases since we are dealing with reverberations. However, intelligibility is decided by how well the keywords themselves can be differentiated. The rest of the phrase does not contribute but rather averages out the acoustic difference between the sentences. Thus, we also attempted to predict intelligibility from keyword speech only, excised from the sentence speech. The RMSE decreased to 0.07, and the correlation increased to 0.82. This is more accurate than other SIP models, such as SRMR. We further plan to expand our model to speech degraded with additive noise and reverberation.