Compensation of speech length effects in forensic likelihood ratio estimation

This work shows how variability of utterance lengths has an impact on automatic speaker recognition systems performance due to score misalignment. Thus, when automatic systems are used in real forensic cases, some misalignments appear due to the fact that utterances are of different lengths. That creates practical problems: the user is forced to either use many reference populations with different training lengths or else discard part of the unquestioned material to reduce the misalignment. In this work the above mentioned issues are analysed and improvements to the LR calculation algorithms are proposed to address them. Finally, results of the new improved algorithms are shown and compared to the results of the current algorithms