Evaluating Forensic Voice Comparison Under Face Mask Conditions Using Pretrained Speaker Embedding Models

Forensic voice comparison (FVC) evaluates whether a known-source recording and a questioned recording are consistent with the same speaker. This study investigates how surgical and FFP2 face masks affect the performance of automatic FVC systems based on pretrained speaker embeddings.

We evaluated five pretrained speaker-embedding models (x-vector, ECAPA-TDNN, ResNet, Wav2Vec2, and WavLM) on recordings obtained under no-mask, surgical-mask, and FFP2-mask conditions within a likelihood-ratio framework. ResNet achieved the strongest overall performance across the evaluated corpora and mask pairings. FFP2-related mismatches tended to produce larger performance changes than surgical-mask mismatches for several models, particularly on MASCFLICHT, although this pattern was not universal.

The lowest C_llr values in the no-mask-no-mask condition were 0.234 ± 0.002 for the Spain corpus and 0.287 ± 0.001 for MASCFLICHT, both obtained with ResNet. Face-mask effects in automatic FVC are condition-dependent rather than uniform.

The results show that robust pretrained embeddings can retain useful speaker information under masked speech, but a mismatch between the conditions of the known-source recording and the questioned recording can reduce discrimination and calibration. These findings support explicit consideration of face-covering conditions when selecting and validating automatic systems for forensic use.