Validation of ECAPA-TDNN System for Forensic Speaker Recognition

This article presents a technical report on the validation of a Forensic Automatic Speaker Recognition (FASR) system utilizing the Emphasized Channel Attention, Propagation and Aggregation in Time Delay Neural Network (ECAPA-TDNN) model under conditions that simulate real forensic voice comparison cases. The study outlines the evaluation process, which incorporates various normalization strategies applied to embeddings and scores, assessing the system's performance in terms of discriminating power, accuracy, and precision metrics. The findings indicate that the ECAPA-TDNN model can effectively serve as a foundational component of a FASR system, demonstrating improved performance compared to previous models within the specified operational conditions. The report also discusses the importance of empirical validation of FASR systems and highlights the forensic_eval_01 evaluation campaign, which aimed to test the validity and reliability of such systems across multiple independent laboratories. The article provides detailed insights into the methodologies employed and the implications for forensic voice comparison practices.