No abstract

Speaker detection, localization and tracking are required in systems that involve e.g. hands-free speech acquisition, or blind source separation. Localization can be done in the (TF) domain, where location features extracted using microphone arrays are used to cluster the TF bins corresponding to the same source. The TF clustering approaches provide an alternative to the Bayesian tracking approaches that are based on Kalman and particle filters. In this work, we propose a maximum-likelihood approach where detection, localization, and tracking are achieved by online clustering of narrowband position estimates, while incorporating the speech presence probability at each TF bin in a unified manner.