Object-based audio capture : separating acoustically-mixed sounds

Abstract

This thesis investigates how a digital system can recognize and isolate individual sound sources, or audio objects, from an environment containing several sounds. The main contribution of this work is the application of object-based audio capture to unconstrained real-world environments. Several potential applications for object-based audio capture are outlined, and current blind source separation and deconvolution (BSSD) algorithms that have been applied to acoustically-mixed sounds are reviewed. An explanation of the acoustics issues in object-based audio capture is provided, including an argument for using overdetermined mixtures to yield better source separation. A thorough discussion of the difficulties imposed by a real-world environment is offered, followed by several experiments which compare how different filter configurations and filter lengths, as well as reverberant environments, all have an impact on the performance of object-based audio capture. A real-world implementation of object-based audio capture in a conference room with two people speaking is also discussed. This thesis concludes with future directions for research in object-based audio capture.