Abstract
There are many situations in which two or more audio programmes may be replayed in the same acoustic space. Such situations can either occur naturally or be deliberately engendered, for example in a system designed to produce personal sound zones. In order to improve the experience of a listener in the presence of an audio interferer, it is desirable to model the relationship between listener experience and physical parameters of the situation. Such a model could be used to optimise audio-on-audio interference situations in a perceptually relevant manner. The first stage of this research involved determination of attributes of the perceptual experience of a listener in an audio-on-audio interference situation. Attributes used in similar research areas were reviewed and an elicitation experiment–combining aspects of individual and group direct elicitation methods preceding a rating experiment–was performed. 'Distraction' was found to be the most relevant attribute for describing the experience of a listener in an audio-on-audio interference situation. The next stage of the research consisted of determining relevant physical parameters and producing a model of distraction. An experiment was performed in which distraction ratings were collected for one hundred randomly created audio-on-audio interference stimuli. Physical parameters were determined following a verbal protocol analysis performed on written descriptions of reasons for the perceived distraction; the resultant categories were used to suggest appropriate features. A linear regression procedure was used to determine the relationship between the features and distraction scores. The selected features were: overall loudness; loudness-based target-to-interferer ratio (TIR); perceptual evaluation of audio source separation (PEASS) toolbox interference-related perceptual score (IPS); level range in the high frequency bands of the interferer; and percentage of temporal windows with low TIR. The resultant model was found to fit well to the training and validation data sets with a root-mean-square error (RMSE) of approximately 10%.