Prof. Dr. Armin Kohlrausch

The ability to segregate and identify sound sources in an auditory "scene" comes naturally to human listeners. Computationally, however, this has proven to be a difficult task. Basically, two approaches can be distinguished: An approach based on sound source properties (i.e. the statistical independence of the corresponding acoustic signals), often summarized under the term blind signal separation (BSS), and an approach trying to mimic human processing in such conditions, called Computational Auditory Scene Analysis (CASA).

As for the human processing of auditory scenes, the cues derived from spatial properties of a sound are often considered to be dominant. This view does, however, neglect the strong role of monaural cues. e.g. temporal onset synchrony or co-modulations across frequency and the role of harmonic relations between spectral components belonging to the same source.

In my presentation I want to give an overview about present-day thinking about the relative role of the various perceptual cues for human sound source separation (and support this overview by acoustic examples), in particular the relative contributions of monaural and binaural cues. Furthermore, I will give examples how this knowledge can be incorporated in algorithms that try to solve the sound source separation problem.