A stream-weight and threshold estimation method using adaboost for multi-stream speaker verification

This paper proposes an automatic stream-weight and threshold estimation method for noise-robust speaker verification using multistream HMMs integrating segmental and prosodic information. The proposed method simultaneously optimizes stream-weights and a decision threshold by combining the linear discriminant analysis (LDA) and Adaboost techniques. Experiments were conducted using Japanese connected digit speech contaminated by white noise with various SNRs. In this experiment, a target ratio of false acceptance rate (FAR) and false rejection rate (FRR) was set by 1:1 so as to adjust them to approach an equal error rate (EER). Experimental results show that the proposed method effectively estimates stream-weights and thresholds so that FARs and FRRs are adjusted to EERs in most of the SNR conditions.