Codec-induced mismatch, speech duration, and speaker-dependent effect in a DNN-based forensic speaker recognition system

Abstract

The performance of forensic automatic speaker recognition (FASR) systems under the combined effects of codec-induced mismatch and different questioned-speech (QS) durations remains insufficiently understood, particularly at the individual speaker level. This study investigates the combined impact of representative lossy speech codecs (G.711 A-law, AMR-NB, and Opus) and QS duration on a DNN (ResNet34-MHA)-based FASR system. System-level performance was generally good across all codec conditions, especially with >30s speech. Codec effects vary, however, with low-bitrate AMR-NB causing the greatest degradation in performance and Opus having minimal impact. Crucially, individual-level analyses reveal substantial speaker-dependent variability that is obscured by system-level metrics: some speakers exhibit persistently poor or unstable performance, and speakers with poorer overall performance are more sensitive to codec-induced mismatch. These findings demonstrate that promising system-level performance does not guarantee good performance for all speakers and highlights the need for individual-level analysis when testing DNN-based FASR systems.