Binaural intelligibility prediction for noisy and non-linearly processed speech
Speech intelligibility prediction is becoming an increasingly popular tool within the speech processing community, as an alternative to time consuming and costly listening experiments. The short-time objective intelligibility (STOI) measure has enjoyed particular popularity, due to its simplicity and high performance in a range of key scenarios. However, the STOI measure lacks the ability to predict binaural advantage (i.e. the advantage of listening with two ears in conditions with spatial separation between target and masker), which is important in many applications. We therefore propose a binaural version of the STOI measure, based on extending the STOI measure with a modified version of the equalization cancellation (EC) model. The binaural STOI measure is shown to retain many of the favorable properties of the STOI measure in diotic conditions. On top of this, the measure can predict the advantage gained from spatial separation between a talker and a point source masker in an anechoic environment. Lastly, as an example of an application, we show how the measure can predict the outcome of a listening experiment, comparing the intelligibility of speech processed by different hearing aids. In this case, the binaural STOI measure is able to predict the relative performance of both normal hearing and hearing impaired listeners quite accurately.