Effects of formant settings and channel mismatch on semi-automatic systems in forensic voice comparison
Vincent Hughes, Philip Harrison, Paul Foulkes, John Peter French, Amelia Gully · White Rose Research Online (University of Leeds, The University of Sheffield, University of York) · 2020
This study examines the sensitivity of formant-based semi-automatic speaker recognition systems to feature extraction settings and channel mismatch.A total of 200 systems were tested, varying LPC order and the maximum number of formants tracked across four channels: studio quality, landline telephone, and two GSM mobile phone samples with different bitrates.For each system calibrated log likelihood ratios were computed for 97 speakers using formants extracted from 60 seconds of vowel-only material.System performance was affected markedly by formant settings, with EER ranging from 8% to 37% and Cllr ranging from 0.28 to 0.88 in the high quality condition.However, some individuals are more or less sensitive to such variation, meaning that system performance is entirely dependent on the specific speakers tested.This issue is discussed in the context of the ongoing debate about the validation of methods and the testing of systems under the conditions of the case.