ISCA Archive Odyssey 2024
ISCA Archive Odyssey 2024

Forensic speaker recognition with BA-LR: calibration and evaluation on a forensically realistic database

Imen Ben-Amor, Jean-François Bonastre, David van der Vloed

The Likelihood Ratio (LR) is fundamental in presenting forensic speaker recognition (FSR) results. Despite its theoretical benefits, conventional LR estimation lacks transparency, impeding courtroom reliability assessment. In response, the Binary-Attribute-based Likelihood Ratio (BA-LR) framework models speech extracts based on the presence or absence of a set of speaker-specific attributes. It estimates the LR as a function of attribute-based LRs. Previous works demonstrated BA-LR's three levels of interpretability: explicit computation of attribute-based LRs, explicit contribution of these LRs to the final LR and phonetic description of the attributes, promising a fully transparent FSR solution. This work adds an examination of LR calibration using a forensically realistic database. Logistic regression is used for calibration purposes, as well as for a regularized fusion of attribute-Log LRs. Results highlight robustness and generalization ability of BA-LR, particularly in forensics.