Robust audio forensic software for recovering speech signals drowned in loud music

Audio evidence, when accepted by the court, can decide the final verdict in a trial. In order to be evaluated, these materials must be authenticated, but also the intelligibility of the message must be undoubtable. Two main categories of multimedia forensics solve these problems: content authentication and noise reduction. The application presented in this paper is part of the latter category. In order to conceal a conversation, the first action that comes into mind is also the easiest one: turn loud a nearby audio source. Since the most available audio sources play musical materials, if a microphone was placed in the room, it would record the speech signal heavily masked by music. A classical adaptive filtering method could be applied to recover the speech only if the speakers and the musical source remain perfectly still or, in other words, the acoustic environment does not change in time. This ideal situation is not to be found very often in real situations. This paper presents a method for recovering speech signals masked by loud music that is robust to acoustic environment variations. The method is thoroughly described, tested, and compared with a solution based on the recursive least-squares (RLS) adaptive algorithm using a variable forgetting factor.