Kurdish speaker identification based on one dimensional convolutional neural network

Voice is one of the vital biometrics in human identification and/or verification area. In this paper, two different models are proposed for speaker identification which are a ۱D convolutional neural network (CNN) and feature based model. In the feature based model, three global spectral based features including Mel Frequency Cepstral Coefficient (MFCC), Linear Prediction Code (LPC) and Local Binary pattern (LBP) are fed to an SVM and k-NN classifiers. Results show that MFCC is the best feature among the others. Consequently, local MFCC features is extracted from the framed signal and used to both the proposed models. The result shows that the local based MFCC improved the accuracy of the CNN based model.

Keywords:

Convolutional neural network , identification , Machine learing