Projects

Past Projects

Current computational models of sound source identification fall far short of the human capacity for identification. Now, a recent advance in sparse signal encoding suggests a means of significantly improving the performance of these models.

Computational models can play an important role in helping to understand our ability to identify everyday objects and events from sound (Bregman, 1990). The traditional approach to modeling has relied on the extraction of structured features and Gestalt schema for identifying sound sources in a mixture (Ellis, 1996; Martin, 1999). These models require high information rates and much prior knowledge of signals for their performance, yet they still fall short of the human capacity for identification.

Now, a recent significant advance in sparse signal encoding suggests a means of improving the performance of these models. Compressed sensing, CS, (Donoho, 2006), replaces the extraction of knowledge-based features with the projection of signals onto a small set of incoherent basis functions. The result for sparse signals is accurate identification with few samples and little prior information about signals.

Our goal in this project is to determine whether CS can be included as an early stage of encoding in traditional models to substantially reduce the information rate required by these models to approach the identification performance of human listeners.

[Work Supported by NIDCD Grant #R01 DC06875]