The Cancellation Principle in Acoustic Scene Analysis

A. de Cheveigné, "Cancellation principle", last revised: 1 Dec 2003 2 dB for a two-microphone delay-and-add beamformer). Second, jammer cancellation works well in situations where SNR is unfavorable (for which segregation is most needed). In that case estimating jammer structure is easy, whereas estimating target structure is hard. Cancellation has two weaknesses. The first is that the jammer may be imperfectly structured or predictable, and thus incompletely suppressed. This results in “crosstalk”. If crosstalk is severe the target may not be observable within certain temporal intervals, and/or spectral bands. The second weakness is that cancellation may “damage” the target. Cancellation requires techniques to deal with incomplete target observations, and to compensate the deleterious effects of target distortion. These two weaknesses are distinct. Jammer “structure” takes many forms. One or several jammers may be predictable, or periodic, or there may be multiple sensors. These basic structures may be extended to include amplitude variation, frequency modulation, moving sources, etc. Every bit of exploitable jammer structure opens a window through which the target can be “glimpsed”. The focus here is mainly on artificial systems (typically automatic speech recognition, ASR), but understanding how the auditory system handles the task is also a goal, in itself and as a source of ideas for better algorithms. Conversely, algorithms serve as models to guide our investigation of natural processes.