A Binaural Front End for Speech Intelligibility Models: Application to the Hearing-Aid Speech Perception Index (HASPI)

Binaural hearing can improve the intelligibility of a speech source spatially separated from competing sound sources when compared to co-located conditions. Monaural speech intelligibility models cannot predict this spatial release from masking. A binaural front end is proposed that can be combined with monaural models to do so. From the noisy speech signals at the two ears, it produces binaurally-enhanced monaural signals that can be evaluated by monaural models. A stationary and a time-dependent version of the front end were tested here with the monaural Hearing Aid Speech Perception Index (HASPI) that compares the envelope modulations of the noisy speech to those of the clean speech. The model predictions were compared to the intelligibility scores of three datasets collected with normal-hearing listeners via headphones measurements in anechoic conditions. A stationary speech-shaped noise (SSN) was tested at 10 azimuths in dataset 1. In dataset 2, an SSN or a non-stationary noise were tested at three azimuths, with or without ideal binary mask processing. In dataset 3, the competing sounds were obtained by mixing the signals from an SSN co-located with the target speech, a diffuse noise coming from all directions, and a spatially-separated SSN. The stationary-front-end predictions are very accurate in the conditions with a spatially-separated SSN, while monaural HASPI predictions at the ear with the better signal-to-noise ratio under-estimate intelligibility. The front-end predictions are slightly less accurate for non-stationary noise but under-estimate intelligibility with diffuse noise. The time-dependent version of the front end systematically over-estimates intelligibility at low signal-to-noise ratios.