High level forensic voice comparison based on fused long-term fundamental frequency and word n-gram features

Feature robustness is particularly important in forensic applications of speaker recognition, where there are often significant differences in the recording conditions between forensic samples. For this reason, high level features have previously been recommended for use in forensic systems, since they tend to be more robust to the acoustic variability introduced by recording conditions [1]. A drawback of high level features though is their poor performance relative to low-level cepstral features. We suggest, however, it may be possible to improve the performance of high feature systems by combining acoustic and idiolectal information, and this may deliver a better trade-off with respect to robustness, interpretability and discrimination performance. In this paper we evaluate a likelihood ratio-based (LR) forensic voice comparison (FVC) system fusing two high level feature subsystems: word n-grams and long-term fundamental frequency (LTF0). Preliminary experiments demonstrate some promising performance gains. We also examine how the duration of speech data impacts on this proposed system.