Audio-Only Speaker Identification using Principal Component Analysis based Back-Propagation Learning Neural Network in Noisy Environment

Abstract

This paper introduces text dependent speaker identification system on Principal Component Analysis based Back-Propagation learning neural network which deals with detecting a particular speaker from a known populations under noisy environment. For audio pre-processing, ends point detection, silence parts removal, frame segmentation and windowing techniques have been used and wiener filter has been applied to remove the background noise from the audio speech utterances. To reduce the dimension of the speech features, Principal Component Analysis method has been used. Different types of features extraction techniques such as LPC, LPCC, RCC, MFCC, ï„MFCC and ï„ï„MFCC have compared and feed to Back-Propagation leaning neural network algorithm to develop the training and testing model of speech pattern. To measure the performance of the system, NOIZEUS speech database has been used with various noise addition rates with eight different noisy environments, i.e., airport, babble, car, exhibition hall, restaurant, street, train and train station noise.

Keywords: Audio-only speaker identification, noisy speech utterance, Back-Propagation learning neural network, PCA based dimensionality reduction, noise robust speaker identification