Reference Sample Size and the Computation of Numerical Likelihood Ratios Using Articulation Rate

This paper explores the effects of variability in the amount of reference data used in quantifying the strength of speech evidence using numerical likelihood ratios (LRs). Monte Carlo simulations (MCS) are performed to generate synthetic data from a sample of existing raw local articulation rate (AR) data. LRs are computed as the number of reference speakers (up to 1000), and the number of tokens per reference speaker (up to 200) is systematically increased. The distributions of same-speaker and different-speaker LRs and system performance (log LR cost (Cllr) and equal error rate (EER)) are assessed as a function of the size of the reference data. Results reveal that LRs based on AR are relatively robust to small reference samples, but that system calibration plays an important role in determining the sensitivity of the LRs to sample size