Machine learning applied to speaker verification of fake voice recordings

For forensic application, speaker verification consists of evaluating whether the voice of a suspect matches the evidence audio recording. In this project, we propose a solution based on machine learning for speaker verification of audios with fake intonation. The input of the system corresponds to indirect characteristics of the audio recordings, and the classifier is a neural network, in which the hyperparameters are adjusted using cross validation. The performance results are: OA (Overall Accuracy) of 88.2%, P (Precision) of 84.5%, R (Recall) of 90%, F1 of 87.2% and AUC (Area under the Curve) of 93.8%.