Robust bootstrapping of speaker models for unsupervised speaker indexing

The conventional bootstrapping approaches of speaker models in unsupervised speaker indexing tasks are very sensitive to the bootstrapping segment duration. If the duration is insufficient to build speaker model, such as in telephone conversations and meetings scenario, serious problems will arise. We therefore propose a robust bootstrapping framework, which employs Multi-EigenSpace modeling technique based on Regression Class (RC-MES) to build speaker models with sparse data, and a short-segment clustering to prevent the too short segments from influencing bootstrapping. For a real discussion archive with a total duration of 8 hours, we demonstrate the significant robustness of the proposed method, which not only improves the speaker change detection performance but also outperforms the conventional bootstrapping methods, even if the average bootstrapping segment duration is less than 5 seconds.