Speaker verification with combined threshold, identification front-end, and UBM

This paper presents a novel approach to improve accuracy performance of a speaker verification system through combination or cascading three different verification methods using an identification "front-end", a universal background model, and an individual matching score threshold. The performance of a speaker verification system can be determined in terms of false rejection rate and false acceptance rate using a standard benchmark speech corpus, which represents fixed common populations in testing voice and claimed identities. By further assuming uniform distributions, it can show analytically that the false acceptance rate of a standalone system either using the threshold or the universal background model can be significantly reduced when combined with the identification \´\´front-end\´\´. Experiments have provided clear evidence, and even more gains to combine all three methods together. The results show 60% reduction in the false acceptance rate for combining with the identification "front-end" alone, and 80% reduction for combining all three methods without adding penalty in the false rejection rate.