University of Chicago - Department of Statistics
In this talk I will describe a new approach to automated recognition of bioacoustic vocalizations, based on Decision trees and labeled graphs.
We develop a procedure that automatically captures properties of the gross shape of the spectrum that are almost invariant within each unit of vocalization, and that discriminate among different units of vocalizations.
Mimicking human experts, our procedure looks at the spectrograms associated to the signals, seeking for relations among the frequency energies over time, that are, statistically speaking, rare events in the space of all signals, but highly probable events within determined units. These relations over time can be thought of as labeled graphs, where vertices correspond to local cues, such as the energy at certain frequency band, and edges, to temporal relations between local cues.
Applications of this methodology to the recognition of segmented spoken digits over the telephone line, and to continuous bird song recognition, will also be described.
Part of this work has been done in collaboration with Y. Amit.