Using interaural mistuning for modelling binaural processing inaccuracies in human speech recognition

A non-intrusive binaural model that predicts speech intelligibility in real-time would be a useful tool in research and applications. One way to model binaural release from masking is to use a front-end using Equalization-Cancelation (EC) processing. Non-intrusiveness and real-time capability require the calculation of a binaurally speech enhanced signal, which can then be further processed for speech intelligibility prediction. A further requirement is the modelling of inaccuracies in the human binaural processing. As far as we know, none of the existing EC front-ends fulfils all of these requirements. One way to consider human inaccuracies is using Monte-Carlo simulations which, however, do not produce a defined signal. Therefore, this contribution presents an alternative realization by replacing them by mistuning the interaural equalization parameters. For the evaluation of this modification, the Speech Intelligibility Index (SII) was used as back-end to compare the model predictions with previous studies. The results indicate that the non-intrusive model with mistuning works well without reverberation. However, it underestimates the binaural benefit in rooms with reverberation. A further analysis showed that a stronger weighting of signal parts with more positive SNR improves the prediction accuracy in reverberation.