A Likelihood ratio-based forensic voice comparison in microphone vs. mobile mismatched conditions using Japanese /ai/

Abstract

This paper describes a likelihood ratio-based forensic voice comparison experiment in microphone versus mobile channel mismatched conditions using parametric representations of formant trajectories. Cubic polynomial coefficients of /ai/ from non-contemporaneous recordings of 30 Japanese male speakers are used to derive multivariate likelihood ratios. The results are evaluated separately for a matched and mismatched group to determine the effect of the mismatch on system performance. A calibrated cross-validated log-likelihood ratio cost (Cllr) of 0.93 is achieved for the F-pattern of /ai/ representing an 18% reduction in system validity relative to the matched group. Separate testing involving only F1 and F2 features evinces a smaller (10%) reduction; suggesting F3 may be more impacted by channel differences. Spectral analysis of F3 indicates this stems from formant tracking errors due to weak signal energy in transmission. As such, F3 in /ai/ should be excluded from analysis where it is poorly preserved. Given the relatively small percentage reductions in validity, it is concluded that /ai/ may be reasonably robust to the mismatch. However, poor performance in optimal conditions (Cllr = 0.77) suggests it may not be a particularly useful parameter in the first place. Limitations to the current study are also discussed.