Log-MelSpectrum and Excitation features based Speaker Identification using Deep Learning

In the recent years, deep learning techniques are being used to construct the Speaker Identification systems. Currently, Salient Feature Extraction plays prominent role in the Speaker Identification task. We propose the fusion of Log-MelSpectrum and Excitation Features to develop the text-independent system using simple Dilated Convolution Neural Network(CNN). To hold the huge useful information, the dilated convolution is being used by increasing the receptive area. The proposed system exhibits the overall accuracy of 91.19% under complex speech data. The performance of this end to end system with the Libri speech and 250 samples of each speaker is compared against existing baseline systems. The experiments conveys that the accuracy reaches only 80.7
^%
using existing method and it improves significantly using the proposed framework.