Selective fusion for speaker verification in surveillance

This paper presents an improved speaker verification technique that is especially appropriate for surveillance scenarios. The main idea is a meta-learning scheme aimed at improving fusion of low- and high-level speech information. While some existing systems fuse several classifier outputs, the proposed method uses a selective fusion scheme that takes into account conveying channel, speaking style and speaker stress as estimated on the test utterance. Moreover, we show that simultaneously employing multi-resolution versions of regular classifiers boosts fusion performance. The proposed selective fusion method aided by multi-resolution classifiers decreases error rate by 30% over ordinary fusion.