A competition-based measurement for HMM speaker verification

Score normalisation with cohort speaker models has been widely used in HMM-based speaker verification. Most of the proposed methods are based on the framework of the hypothesis testing. Based on this framework an overall average of all cohort scores is often used for normalisation, which leads a log likelihood ratio (LLR) for verification. In this paper we use a competition-based criterion to define the measurement. Based on this criterion a measurement is proposed, which reflects the competitiveness of claimed model against cohort models for a given testing utterance. The evaluation is carried out on the YOHO database. A comparison with conventional LLR is given. The results show that with this proposed measurement can improve the speaker verification (SV) performance significantly. The SV equal error rate can be reduced by about 30%.