Collecting population statistics: The discriminant power of clicks

Abstract

This research gathers population statistics on clicks for use in likelihood ratios (LRs). As reported in Gold and French (2011), clicks have been analyzed by 57% of experts in forensic speaker comparison cases and 18% of experts find them to be useful speaker discriminants. Eight minutes of speech from 100 male speakers of Southern Standard British English were analyzed from the DyVis Database, using categorical annotations of clicks (Wright, 2007). The distribution of click use in subjects is highly skewed with a large majority not clicking. However, the distribution of clicks is highly variable with non-clickers ranging from 25–44% of the population depending on the length of the speech sample. The same 100 speakers were also analyzed for click use when speaking with two additional interlocutors. Again the results are highly variable, which suggests the intra- and inter-speaker instability of clicks, the lack of overall robustness, and the accommodation of clicks in speech. This study serves as a beginning point in incorporating previously unreported population statistics into LRs, and specifically examining the potential of including higher order and paralinguistic features in a Bayesian framework. [Research funded by the European Community's Seventh Framework Program (FP7/2007-2013) under grant agreement #238803].