A reliability criterion for time-frequency labeling based on periodicity in an auditory scene

Computational Auditory Scene Analysis (CASA) aims to model our ability to structure our acoustical environment. In a CASA context, this paper deals with a method for time-frequency labeling based on harmonic properties. The method is based on a classical pitch extraction algorithm, termed the « zerocrossing method », which is known to be particularly sensitive to any kind of interference. This work shows that its sensitivity can be turned into an advantage for harmonicity detection in interfering conditions, and provides, according to two estimators which are precisely characterized, a time-frequency representation labeled according to a reliability criterion. A model for speech segregation is subsequently designed and evaluated in different interference paradigms.