Modelling binaural detection of speech stimuli in complex reverberant environments

In listening experiments that assess spatial hearing (e.g. localization) in the presence of background noise it is important to ensure audibility of the target stimulus. Target audibility could be easily controlled using knowledge about the masked thresholds in the respective paradigm. However, if a large number of acoustic conditions as well as target sounds are considered, individual psychoacoutic threshold measurements are not feasible and alternative methods need to be applied. In this study a binaural auditory model is applied to predict masked thresholds for wide-band non-stationary stimuli. This model closely follows the approach of Hant and Alwan (Speech Communication, 2003, Vol. 40, pp. 291-313) [7]. In the first stage of the model an auditory front-end generates an internal representation of the stimuli in both ears. The auditory front-end includes head-related transfer functions, auditory bandpass filtering, squaring, temporal integration, logarithmic compression and additive internal noise. In the second stage, a decision device combines d′ information across time, frequency and ears and provides an estimate of the masked threshold. Three different methods of combining information across ears were compared: A better-ear approach, a cross-ear glimpsing approach and an approach using binaural integration. The model was verified using psychoacoustic masked threshold data of a detection paradigm that considered a target word presented from 15 different locations in a reverberant multi-talker background. The model predictions were in very good agreement with the measured masked thresholds. It is concluded that this model is suitable to control audibility of a target stimulus in a complex reverberant environment. Moreover, the model was used to systematically analyze the effect of head-shadow, signal spectrum, auditory sensitivity and room reverberation on target audibility.