2.5. Experimental procedure

The 50-s long speech segments were used to generate auditory scenes comprising a male and a female simultaneously speaking in anechoic or reverberant rooms. The two concurrent speech streams were normalized to have similar root-mean square values. The speech stimuli were delivered to the subjects via ER-2 insert earphones (Etymotic Research). The speech mixtures were presented binaurally to the listeners, with the two speech streams lateralized at respectively −60° and +60° along the azimuth direction and a source-receiver distance of 2.4 m. This was achieved using non-individualized head-related impulse responses that were simulated using the room acoustic modeling software, Odeon (version 13.02). Each subject undertook sixty trials in which they were presented the 50 s-long speech mixtures. Before each trial, the subjects were cued to listen selectively to one speech stream and ignore the other. After each trial, the subjects were asked a comprehension question related to the content of the attended speech stream. The position of the target streams as well as the gender of the target speaker were randomized across trials. Moreover, the type of acoustic room condition (either anechoic, mildly reverberant or highly reverberant) were pseudo-randomized over trials. In the analysis, data recorded from all acoustic conditions were pooled together. The reasons for doing this were twofold. Firstly, it provides sufficient data for the stimulus-response analysis. This is particularly important as insufficient data in worst case can lead to poorer model estimates (Mirkovic et al., 2016). Secondly, by using this approach we get a better idea of how well the models will generalize to different experimental conditions. This is an important practical aspect, as it gives a better estimate of how well a classifier will perform in different listening conditions (rather than just focusing on training on anechoic data and evaluating on anechoic data).