Online unsupervised overlapping speaker detection using enhanced classification history-based features

Overlapping speaker localization approaches generally require a binary detector which performs the source/noise classification of the location estimates. This is mainly due to the unknown time-varying number of sources, and to the presence of noise and reverberation. In this paper, we firstly introduce an online implementation of a previously developed offline multiple speaker detector. This classifier is then extended to include new detection features. More precisely, the proposed approach uses the classified location estimates as labelled data to train new classification models for different potential features. The resulting models are then integrated into the online classifier to improve the classification performance. In particular, this paper investigates three different classification history-based models, namely, the location, the kurtosis and the probabilistic steered response power features. Experiments conducted on the AV16.3 corpus show the effectiveness of the proposed approach.