A new cohort normalization using local acoustic information for speaker verification

This paper describes a new cohort normalization method for HMM based speaker verification. In the proposed method, cohort models are synthesized based on the similarity of local acoustic features between speakers. The similarity can be determined using acoustic information lying in model components such as phonemes, states, and the Gaussian distributions of HMMs. With the method, the synthesized models can provide an effective normalizing score for various observed measurements because the difference between the individual reference model and the synthesized cohort models is statistically reduced through fine evaluation of acoustic similarity in model structure level. In the experiments using telephone speech of 100 speakers, it was found that high verification performance can be achieved by the proposed method: the equal error rate (EER) was drastically reduced from 1.20% (obtained by the conventional speaker-selection based cohort normalization) to 0.30% (obtained by the proposed method on distribution-based selection) in a closed test. Furthermore, EER was also reduced from 1.40% to 0.70% in open test (reference speaker: 25, impostor: 75), when the other speakers than the reference speaker were used as impostors.