My current research concerns the sampling effect on system performance, speaker individuality and score distributions in forensic voice comparison. The main issues I am trying to address are,
Given that test, training and reference speakers are all from the relevant population, how sampling will effect system performance when different test, training and reference speakers are used.
What makes speaker a 'good speaker' ? i.e. speaker always give good performance in forensic speech comparison.
Is there any correlation between score distributions and system performance?