Voice Biometrics Exploring The Boundaries Of Speaker Verification

Recent advancements in computational power, datasets and models have resulted in significant performanceincreases across various speech and speaker recognition models. This work investigates the capabilities ofusing voice as a biometric tool, namely what is referred to as speaker verification. Specifically, two featureextraction models named TitNet and ECAPA from NVIDIA are tested as front-end. Two back-end models,cosine similarity and Euclidean distance are tested. The models are thoroughly tested in terms of their equalerror rate (EER), inference time complexity, language independence, sensitivity to the choice of threshold,impact of the number of enrollments and effects of the duration of the audio files used. A system combinationto reduce the risk of spoofing by replaying recorded audio is presented by incorporating a random wordgenerator with a speech-to-text model and a string similarity checker. The results show EERs as low as 0.2%.Utilizing multiple audio files for the enrollments decreases the EER from e.g. 0.78% with one enrollment filedown to 0.2% with five enrollment files. Further, the results indicate that a trial file with a duration of around7–9 seconds is optimal. Experiments with various thresholds demonstrate the trade-off between convenienceand security. The verification pipeline is shown to be language-independent to a high degree with similarEERs for both English, Swedish, Danish and Norwegian test datasets. Inference time complexity was testedon a GPU and a CPU. The pipeline managed to process up to 14 seconds of data per second on a CPU and4.1 minutes on a GPU. Further, the processing time appears to be linearly increasing as the duration of theaudio file to process increases. Euclidean distance as back-end performed similarly, but slightly worse, to thecosine similarity. In the future, re-training the feature extraction models with more data, especially in otherlanguages, could potentially further increase the performanc