This study examines the impact of deepfake voice generation on the performance of speaker verification systems, with a particular focus on forensic voice comparison. Three speaker embedding models—X-vector, ECAPA-TDNN, and WavLM—are evaluated for their robustness against deepfake audio. The “In the Wild” dataset is used for fine-tuning and evaluation, and an Australian English forensic dataset is employed for calibrating likelihood-ratio scores. Experimental results show that the pre-trained ECAPA-TDNN model delivers the best overall performance, achieving the lowest equal error rate (8.0%) and superior calibration metrics. Fine-tuning on a dataset comprising both real and deepfake samples improves performance in same-speaker scenarios but reduces accuracy and calibration in all-speaker conditions. These findings underscore the importance of carefully integrating deepfake audio during model training to balance spoof detection and generalization in forensic applications.