Tracking a varying number of speakers using particle filtering

The extension of particle filtering techniques to the multiple speaker case is difficult as two distinct problems must now be addressed. Firstly, the active speakers must be identified and their locations estimated, requiring the use of multi-dimensional likelihoods, and then each speaker must be correctly associated with his corresponding location. In this paper we propose a multi-speaker tracking algorithm in which the number of active speakers is determined by estimating the profile of the noise-plus-reverberation covariance matrix eigenvalues. The multi-dimensional likelihoods are then decoupled using the Expectation Maximization (EM) algorithm. The tracking accuracy is improved by the inclusion of a pause detection step and estimation of the noise-plus-interference covariance matrix. The results show the benefits of the proposed methods under difficult tracking situations.