Conference_programme: 21.1 - Source localization and acoustic array processing

Lecture: Speaker Localization and Tracking Using Recursive and Distributed Expectation-Maximization (EM) Algorithms

Author(s): Gannot Sharon, Schwartz Ofer, Dorfan Yuval

Summary:
We address the problem of acoustic localizing and tracking of unknown number of concurrent speakers in noisy and reverberant enclosures, using spatially distributed microphone array. The localization task is formulated as a maximum likelihood (ML) parameter estimation problem, using a Mixture of Gaussians (MoG) model, with the centroid of each Gaussian corresponding to a candidate position in the enclosure on a predefined grid. The larger is the weight of a Gaussian, the higher is the probability that a speaker is active in the corresponding candidate position. The ML parameter estimation problem is solved by utilizing the expectation-maximization (EM) procedure.\nFor the tracking scenario, we propose to adapt a recursive variant of the EM procedure (REM), proposed by Titterington, which utilizes a Newton-based recursion for the maximization. In this work, we also extend Titterington's method to deal with constrained maximization, encountered in the problem at hand. \nWe then discuss the problem of distributed versions of the localization methodology, suitable for wireless acoustic sensor networks (WASNs) with known microphone positions. WASNs are characterized by low computational resources in each node and by a limited connectivity between the nodes. We propose a novel bi-directional tree-based distributed estimation–maximization (DEM) algorithm that circumvents the inherent network limitations. We further propose a recursive distributed EM (RDEM) variant that is better suited to online applications. \nFinally, we show that an improved acoustic front-end can further increase the robustness of the proposed methods to high reverberation. \nThe applicability of the proposed methods to the localization and tracking problems is demonstrated using both simulated data and actual recordings from our acoustic lab.

Corresponding author

Name: Prof Sharon Gannot

e-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.

Country: Israel