Forensic Automatic Speaker Recognition: Coping with Impersonated Speech

Forensic Automatic Speaker Recognition (FASR) systems are essential for authenticating individuals based on speech characteristics in forensic contexts. The efficacy of the system is investigated through an experiment featuring five celebrities to evaluate the accuracy of identifying whether a test speech belongs to a genuine speaker or an impersonator. A database consisting of recordings from a selected number of celebrities and impersonated speech recordings from professional mimicry artists (who are good at mimicking these celebrities) are used for the experiment described in this paper. Mel-spectrograms of both original and impersonated speech are used to train a CNN model, and the performance of the model is evaluated using performance metrics such as accuracy, recall, and precision. The system gives an overall accuracy of 84.93% while evaluating the CNN models corresponding to 5 different celebrities.