Microphone Identification using Higher-Order Statistics

This paper presents a statistical framework for microphone identification using only digital audio recordings. To accomplish this task, the microphone-induced artifacts are modeled using a nonlinear function and then a statistical tool based on higher-order statistics is used to capture these artifacts. More specifically, polyspectral analysis is used to capture microphone-induced artifacts. Distanceand correlation-based similarity measures are used for automatic microphone identification. More specifically, distance between scale-invariant Hu moments of the bicoherence magnitude spectrum and cross-correlation between bicoherence phase spectra are used. The effectiveness of the proposed framework has been tested with 24 audio recordings captured using eight microphones during three recording sessions. Performance of the proposed scheme is evaluated using ambient noise recordings captured using eight microphones of four different types.