Evaluating a Binaural Model Using Real-Time Assessment of Listening Effort in Human Listeners

A model for predicting speech intelligibility (SI) and listening effort (LE) non-intrusively and in real-time has been developed. The front-end stage models binaural hearing by utilizing interaural parameters and outputs a monaural signal, which is analyzed by a deep learning-based back-end making predictions on SI and LE. The model runs in real-time and can adapt to fast dynamic changes of an acoustic scene due to very frequent predictions in intervals of 23 milliseconds. However, tools for evaluating such predictions with human listeners are missing.This study presents an experiment in which normal-hearing participants assessed perceived LE in real-time using a slider interface with several LE categories via smartphone, while pre-generated dynamic acoustic scenes were played back. The acoustic scenes consisted of a target speaker and a noise interferer in a large reverberant room. The signal-to-noise ratio and the amount of reverberation were varied after random time periods in a rapid fashion in order to induce a change of perceived LE. We analyzed the correlation between human and model data as well as reaction times for tuning the model parameters.