The individual and the system:Assessing the stability of the output of a semi-automatic forensic voice comparison system

Semi-automatic systems based on traditional linguistic-phonetic features are increasingly being used for forensic voice comparison (FVC) casework. In this paper, we examine the stability of the output of a semi-automatic system, based on the long-term formant distributions (LTFDs) of F1, F2, and F3, as the channel quality of the input recordings decreases. Cross-validated, calibrated GMM-UBM log likelihood-ratios (LLRs) were computed for 97 Standard Southern British English speakers under four conditions. In each condition the same speech material was used, but the technical properties of the recordings changed (high quality studio recording, landline telephone recording, high bit-rate GSM mobile telephone recording and low bit-rate GSM mobile telephone recording). Equal error rate (EER) and the log LR cost function (Cllr) were compared across conditions. System validity was found to decrease with poorer technical quality, with the largest differences in EER (21.66%) and Cllr (0.46) found between the studio and the low bit-rate GSM conditions. However, importantly, performance for individual speakers was affected differently by channel quality. Speakers that produced stronger evidence overall were found to be more variable. Mean F3 was also found to be a predictor of LLR variability, however no effects were found based on speakers’ voice quality profiles.