Two underlying factors that cause potential error-prone speakers in speaker verification (SV) systems are examined, namely the inter-variance which indicates the individuality of the speaker, and the intra-variance which is measure how much the speech varies over time. It is shown that SV errors can be predicted from either inter or intra-variance but more accurately from a product of the two.
An error-prone speaker detection scheme based on a speaker identification (SI) system is proposed which could be useful at the enrollment stage to improve system performance.