Voice Comparison and Rhythm: Behavioral Differences between Target and Non-target Comparisons

Abstract

It is common to see voice recordings being presented as a forensic trace in court. Generally, a forensic expert is asked to analyze both suspect and criminal's voice samples in order to indicate whether the evidence supports the prosecution (samespeaker) or defence (different-speakers) hypotheses. This process is known as Forensic Voice Comparison (FVC). Since the emergence of the DNA typing model, the likelihood-ratio (LR) framework has become the golden standard in forensic sciences. The LR not only supports one of the hypotheses but also quantifies the strength of its support. However, the LR accepts some practical limitations due to its estimation process itself. It is particularly true when Automatic Speaker Recognition (ASpR) systems are considered as they are outputting a score in all situations regardless of the case specific conditions. Indeed, several factors are not taken into account by the estimation process like the quality and quantity of information in both voice recordings, their phonological content or also the speakers intrinsic characteristics. In our recent study, we showed the importance of the phonemic content and we highlighted interesting differences between inter-speakers effects and intra-speaker's ones. In this article, we wish to take our previous analysis a step farther and investigate the impact of rhythm variation separately on target and non-target trials.

Multi-task Knowledge Distillation with Rhythm Features for Speaker Verification

“…Compared with x-vec, x-vec+asoftmax achieves a 4.6% EER reduction at the cost of target comparison accuracy. At the meantime, it is worth noting that x-vec+asoftmax+rhythm have significantly improved the target comparison, and it does not affect the accuracy of the non-target comparison, which is consistent with the conclusion in [19]. Besides, multi-task knowledge distillation makes the student-64-TS network have better performance in both target and non-target comparisons, with a 7.1% EER reduction.…”

On the limits of automatic speaker verification: Explaining degraded recognizer scores through acoustic changes resulting from voice disguise

“…Second, change in the short-term spectral envelope (as used by the ASV systems) impacts formants, and we therefore expect them to form reasonable predictors of the LLR score. Concerning prosody, there is a vast body of literature ranging from frame-level F0 characterization (Mary and Yegnanarayana, 2008), stylized (Shriberg et al, 2005;Adami, 2007) and polynomially modeled F0 contours (Dehak et al, 2007), along with energy and timing-or rhythm-related features (Dellwo et al, 2012;Ajili et al, 2018). A number of studies have addressed the impact of such parameters in a forensic context (Leemann et al, 2014;Moez et al, 2016).…”