This project implements a convolutional neural network based model to identify a speaker based on a short audio signals from among the known set of speakers enrolled during the model training, with an emphasis on text-independent speaker recognition.
Traditional approaches based on Gaussian mixture model and Universal background model (GMM-UBM) have high success rate but with a higher computational cost during the GMM evaluation phase. We experiment with a deep learning architecture based on convolutional neural network (CNN). Our CNN model is trained and tested against freely available and comprehensive VoxForge (voxforge.org) dataset and provide constant evaluation cost.