Acoustic environment identification by Kullback-Leibler divergence

This paper presents a forensic methodology that determines, from among a set of recording places, the probable place where allegedly a disputed digital audio recording was made. The methodology considers that digital audio recordings are noisy signals that have two involved noise components. One component is the multiplicative noise, which is an internal feature on the audio recording that is related to the recording device. The other component is the additive noise, which is an external feature on the audio recording that can be related to the recording place. Therefore, the proposed methodology estimates a likelihood rate that helps to decide which recording place is more plausible to be associated with a disputed audio recording. This likelihood rate is defined as the probability of a finding, supposing that a specific proposition is true, divided by the probability of a finding if an alternative proposition is true. Such probabilities are calculated by performing a statistical comparison through the Kullback-Leibler divergence [1], between the probability distribution function of the additive noise associated to the disputed recording and the probability distribution function of the additive noises associated to a set of audio recordings made on the possible recording places. Then, in order to determine the recording place, the analyst requires a list of possible places where the recording could have been carried out; in these places some reference recordings will be made. In this work, the additive noise is estimated by the Geometric Approach to Spectral Subtraction (GA-SS) filter [2], applied to the noisy audio recording.