Adapted Gaussian mixture model in likelihood ratio based forensic voice comparison using long term fundamental frequency

In this paper, the Gaussian Mixture Model – Universal Background Model (GMM-UBM) is applied to onedimensional speech data, namely the distribution of long term fundamental frequency (LTF0) in likelihood ratio based forensic voice comparison. A series of experiments were conducted using varying numbers of Gaussians, differing adaptation rates to a UBM, and different lengths of speech samples. The results of the GMM-UBM procedure are compared to two previously proposed procedures for LTF0. All three procedures exhibited unique characteristics in their performances. Thus, there was no consistency in performance in that no one procedure constantly outperformed the others. Index Terms: forensic voice comparison, likelihood ratio, GMM-UBM, long-term F0 distribution