Speaker: V. Vishnu Vardan Varanasi
Date: May 18, 2017
Abstract:
DOA Estimation is a challenging task especially in presence of noise and reverberation. Various applications of acoustic source localization include Distant Automatic Speech Recognition, Music Information Retrieval etc. Spherical Microphone Array(SMA) captures spherical variation of acoustic field with spherical harmonics. A wide range of DOA estimation algorithms are developed in the past few decades. Most of them can be broadly categorized as follows. Subspace based methods such as MUSIC, Steered beamforming methods such as Steered Response Power with Phase Tranform, Time Delay of Arrival based approaches, Learning based approaches with features such as GCC, RTF and Sparsity based methods.
In adverse environments, these methods suffer from inaccurate DOA estimates or high computational complexity. Learning based approaches provide better performance in adverse environments. However, features based on GCC and RTF are high dimensional and hence learning becomes computationally complex. We discuss novel low dimensional spherical harmonic features. An online manifold regularization framework that can utilize the low dimensional nature of these features to compute real time DOA estimates will also be discussed.
