Forensic Acoustic Applications of Siamese Neural Networks for determining Lineal Relationships between Normal and Spectrally Disguised Voices

In forensic applications, it is sometimes required to determine whether two speech samples, one in normal voice and other in a disguised form, are from the same speaker. In this study a Deep Learning based Siamese Neural Network is trained for determining whether normal and electronically disguised versions have same lineage. The Siamese Neural Network model was trained on speech samples from the Ravdess database and their electronically disguised versions. The disguised samples were generated using various mathematical functions to change the spectral content in the Fourier domain. The trained Siamese Neural Network attained an accuracy of 94.58 on trained data and 94% on test data.