Method for generating realistic sound stimuli with given characteristics by controlled combination of audio recordings.

Audio recordings are often used to improve ecological validity of stimuli for laboratory research on effects of noise. In this paper a method is proposed for composing realistic environmental sound stimuli with (1) specified overall spectrum and (2) specified statistical distribution of sound event durations and semantic categories. The combination is addressed as a mixed integer linear programming problem. Objective measurements, for eight stimuli and a moderate-size database, validate the method. The mean error in octave bands exposure level is 2.6 dB and the statistical distribution of sound event durations and semantic categories is perfectly matched.