Robust speaker recognition using library of cross‐domain variation compensation transforms

Abstract

Although the state-of-the-art i-vector-based probabilistic linear discriminant analysis systems resulted in promising performances in the National Institute of Standards and Technology speaker recognition evaluations, the impact of domain mismatch when the system development data and the evaluation data are collected from different sources remains a challenging problem. This issue was a focus of the Johns Hopkins University 2013 speaker recognition workshop where a domain adaptation challenge (DAC13) corpus was created to address it. The cross-domain variation compensation (CDVC) approach has been recently proposed to address it when in-domain development data are available. The work reported by the present authors addresses this issue when in-domain development data are unavailable using a library of CDVC transforms. This approach is evaluated on the DAC13 corpus and is shown to be more powerful than nuisance attribute projection-based inter-dataset variability compensation and the whitening library.