The effect of sample size on the performance of likelihood ratio-based forensic voice comparison

This is a preliminary investigation into the effect of sample size on a likelihood ratio based forensic voice comparison system. We investigated how the number of vowel tokens used in the testing data affects LR estimation, using Japanese filler expressions elicited from spontaneous speech. In forensic voice comparison casework, we cannot control how many tokens of vowels or segments are available: sometimes only a few comparable tokens for each vowel can be found. We currently do not know if the LR produced using two tokens of a vowel is as reliable as one produced using ten tokens. We found that the two components of quality of LR Cllr_min and Cllr_cal responded very differently to additional tokens. Cllr_cal consistently deteriorated as tokens were added to the testing data, whereas Cllr_min improved rapidly. The improvement was particularly notable up to six tokens.