Time-frequency masking based multiple acoustic sources tracking applying Rao-Blackwellised Monte Carlo data association

A Rao-Blackwellised particle filtering approach for tracking multiple simultaneously active and time-varying number of speakers is investigated. A novel measurement extraction method appropriate for the scenario of multiple sources is proposed based on a time-frequency masking technique, in which each source is represented separately by a joint gain-ratio and time-delay histogram. An existing Rao-Blackwellised particle filtering and data association approach is then employed to track the sources. The position of the sources are marginalised by an extended Kalman filter, and it is only the data association that needs to be processed using a particle filter. The tracking capability of the proposed algorithm is demonstrated in different simulated room environments and compared with the tracking performance obtained by applying the observation from the more commonly used generalised cross correlation function.