Binary and Multiclass Calibration in Speaker and Language Recognition
Explore a comprehensive lecture on binary and multiclass calibration techniques in speaker and language recognition systems. Delve into the importance of well-calibrated probabilistic outputs for automatic pattern classifiers, with a focus on applications in speaker and language recognition technologies. Examine the derivation and re-interpretation of cross-entropy as an objective function for classifier training, and understand its relationship to expected cost in Bayes decision-making. Learn about evaluation methodologies, including criteria for measuring calibration quality, and gain insights into optimizing classifier performance across various applications. Benefit from the expertise of Niko Brummer, a renowned researcher in the field, as he shares his knowledge on probabilistic modeling, generative and discriminative recognizers, and evaluation techniques for classifiers producing well-calibrated class likelihoods.