A speaker verification backend with robust performance across conditions

In this paper, we address the problem of speaker verification in conditions unseen or unknown during development. A standard method for consists extracting embeddings with a deep neural network and processing them through backend composed probabilistic linear discriminant analysis (PLDA) global logistic regression score calibration. This is known to result systems that work poorly on different from those used train calibration model. We propose modify backend, introducing an adaptive calibrator uses duration other automatically extracted side-information adapt inputs. The trained discriminatively optimize binary cross-entropy. When number diverse datasets are labeled only respect speaker, proposed consistently and, some cases, dramatically improves calibration, compared PLDA approach, held-out datasets, which markedly training data. Discrimination performance also improved. show joint essential -- same benefits cannot be achieved when freezing fine-tuning calibrator. To our knowledge, results paper first evidence literature it possible develop system robust out-of-the-box large variety conditions.