Department: Electrical & Computer Engineering
Faculty Advisor(s): Bhaskar Rao

Primary Student
Name: Alireza Masnadi-Shirazi
Email: amasnadi@ucsd.edu
Phone: 858-232-4168
Grad Year: 2012

Methods based on frequency-domain independent component analysis (ICA) in junction with state coherence transform (SCT) have been shown to be robust for extracting source location information like time difference of arrival (TDOA) and direction of Arrival (DOA) in highly reverberant environments and in the presence of spatial aliasing due to large microphone spacing. Also, By exploiting the frequency sparsity of the sources, such methods have proven to be effective when the number of simultaneous sources is larger than the number of microphones. Assuming that the number of sources is known and fixed in time, some methods exist that track the location information for each speaker by incorporating a separate tracker for each source. However, in many real world problems the number of concurrent speakers is unknown and varies with time as new speakers can appear and existing speakers can disappear or undergo silence periods. In order to deal with this challenging scenario of unknown time-varying number of speakers, we propose the use of the probability hypothesis density (PHD) filter to post-process the measurements obtained from ICA/SCT as such measurements can contain clutter and missed detections due to reverberation and short pauses. PHD filter is a Bayesian method based on random finite sets (RFS), where the multi-target states and the number of targets are integrated to form a set-valued variable, making it capable of performing the mutli-source tracking while automatically estimating the number of concurrent sources. A Gaussian mixture implementation of the PHD filter (GM-PHD) is utilized that solves the data association problem intrinsically, hence providing distinct tracks. The tracking and separation capabilities of the proposed method is demonstrated using simulations of multiple sources in reverberant environments.

« Back to Posters or Search Results