Sequential UBM adaptation for speaker verification

GMM-UBM-based speaker verification heavily relies on a well trained UBM. In practice, it is not often easy to obtain an UBM that fully matches acoustic channels in operation. To solve this problem, we propose a novel sequential MAP adaptation approach: by being sequentially updated with data from new enrollments, the UBM learns and converges to the working channel. Our experiments are conducted on a time-varying speech database, with two channel-mismatched UBMs as the initial model. The results confirm that the sequential UBM adaptation provides significant performance improvement, leading to a relative EER reduction of 6.3% and 14.8% for the two mismatched UBMs, respectively.