Effect of Noises and GSM Codings on Pre-Trained Speaker Embedding Models in Forensic Voice Comparison

Deep learning has witnessed extensive application within speaker verification (SV) in recent years, particularly in feature extraction or embedding features. This study delves into examining the impact of noise and Global System for Mobile Communications (GSM) coding, with a particular emphasis on metrics relevant to forensic voice comparison, and examining how enrollment methods (averaging embeddings) can mitigate these adverse effects. It further evaluates the influence of practical variables such as speech sample duration and enrollment on verification reliability. It provides critical recommendations for forensic practice, especially when involving GSM coding and different types of noises. In this study, the assessment was based on the ForVoice 120+ Hungarian speech dataset. Samples are compared pairwise with and without enrollment for suspect speakers (known) because, often, forensic voice comparisons don’t have multiple samples of offenders. Speaker embedding features are extracted using four pre-trained techniques: X-vector, Emphasized channel attention, propagation and aggregation in time delay neural network (ECAPA-TDNN), large-scale self-supervised (Wavlm) and framework for self-supervised learning (Wav2vec) pre-trained. The evaluation of speaker verification performance employed the likelihood-ratio framework, utilizing measures such as log-likelihood ratio cost ( ${C} _{\text {llr}}$ ) and equal error rate (EER). Results show that environmental noise and GSM coding significantly degrade speaker verification performance. The best-performing model, ECAPA-TDNN, achieved a minimum EER of 1% and a ${C} _{\text {llr}}$ of 0.048 in clean conditions with speaker enrollment. However, under degraded conditions such as white noise at 10 dB SNR and half-rate GSM coding, performance dropped notably, with EER increasing to 6.4% and ${C} _{\text {llr}}$ rising to 0.241. Despite this, speaker enrollment mitigated some performance loss. These findings highlight the need for forensic practitioners to account for the effects of noise and coding when conducting speaker verification.