Music and Noise Fingerprinting and Reference Cancellation Applied to Forensic Audio Enhancement

In this paper, we present a novel application, of audio fingerprinting and reference cancellation, to forensic audio enhancement. In audio surveillance recordings, it is common to find the speech of interest masked by the speech of other non-target speakers in the room or obscured by interfering music or television noise, as well as other noises such as banging, clanging or slamming, in the acoustic environment. These noises can drown the speech of interest and make it difficult to understand easily or be accurately transcribed. When two recordings of a particular acoustic event are available, two-channel adaptive filtering or reference cancellation is a highly effective tool for enhancement. However, the two recordings have to be painstakingly manually aligned, and this is difficult to do by eye or ear. We present an approach using landmark-based acoustic fingerprinting to identify, automatically align, and subtract reference sounds and bring the speech of interest to the forefront.