R-norm: improving inter-speaker variability modelling at the score level via regression score normalisation

This paper presents a new method of score post-processing which utilises previously hidden relationships among client models and test probes that are found within the scores produced by an automatic speaker recognition system. We suggest the name r-Norm (for Regression Normalisation) for the method, which can be viewed as both a score normalisation process and as a novel and improved modelling technique of inter-speaker variability. The key component of the method lies in learning a regression model between development data scores and an ‘ideal’ score matrix, which can either be derived from clean data or created synthetically. To generate scores for experimental validation of the proposed idea we perform a classic GMM-UBM experiment employing mel-cepstral features on the 1sp-female task of the NIST 2003 SRE corpus. Comparisons of the r-Norm results are made with standard score postprocessing/normalisation methods t-Norm and z-Norm. The rNorm method is shown to perform very strongly, improving the EER from 18.5% to 7.01%, significantly outperforming both z-Norm and t-Norm in this case. The baseline system performance was deemed acceptable for the aims of this experiment, which were focused on evaluating and comparing the performance of the proposed r-Norm idea.