A comparison of procedures for the calculation of forensic likelihood ratios from acoustic-phonetic data: Multivariate kernel density (MVKD) versus Gaussian mixture model-universal background model (GMM-UBM)

Two procedures for the calculation of forensic likelihood ratios were tested on the same set of acoustic-phonetic data. One procedure was a multivariate kernel density procedure (MVKD) which is common in acoustic-phonetic forensic voice comparison, and the other was a Gaussian mixture model-universal background model (GMM-UBM) which is common in automatic forensic voice comparison. The data were coefficient values from discrete cosine transforms fitted to second-formant trajectories of /a/, /e/, /o/, /a/, and // tokens produced by 27 male speakers of Australian English. Scores were calculated separately for each phoneme and then fused using logistic regression. The performance of the fused GMM-UBM system was much better than that of the fused MVKD system, both in terms of accuracy (as measured using the log-likelihood-ratio cost, C_llr) and precision (as measured using an empirical estimate of the 95% credible interval for the likelihood ratios from the different-speaker comparisons).