The effect of speaker sampling in likelihood ratio based forensic voice comparison

Within the field of forensic voice comparison (FVC), there is growing pressure for experts to demonstrate the validity and reliability of the conclusions they reach in casework. One benefit of a fully data-driven approach that utilises databases of speakers to compute numerical likelihood ratios (LRs) is that it is possible to estimate validity and reliability empirically. However, little is known about the stability of LR output as a function of the specific speakers sampled for use in the training, test and reference data sets. The present study addresses this issue using two large sets of formant data: Cantonese sentence final particle /a/ and British English filled pauses UM. Experiments were replicated 100 times varying the 1) training test and reference speakers, 2) training speakers only, 3) test speakers only, and 4) reference speakers only. The results show that varying the speakers in all three sets has the greatest effect on system stability for both the Cantonese and English variables, with the C-llr varying from 0.60 to 0.97 for /a/ and 0.32 to 1.33 for UM. However, this variability is primarily due to the effects of uncertainty in the test set. Varying only the training speakers has the least effect on system stability for /a/ (C-llr range: 0.76 to 0.88), while varying reference speakers has the smallest effect for UM (C-llr range: 0.40 to 0.54). The results indicate that in LR-based FVC it is important to assess the stability of the system as a function of the samples of speakers used (C(llr )range) rather than just reporting a single C(llr )value based on one configuration of speakers in each set. The study contributes to the general debate on reporting uncertainty in LR computation.