Integrating Fast-ICA and MFCC Methods for Voice Recognition with Noise Splitter

Abstract

We developed a voice recognition system with noise splitter using Fast Independent Component Analysis (Fast-ICA) and Mel Frequency Cepstrum Coefficients (MFCC) methods. The developed system receives input signals which consist of mixtures of human speeches with environmental noise. Various sounds are played as background noise during voice recording process. The Fast-ICA method is used to distinguish the input signals from the noise (instrumental sounds). After the splitting process of the input signals, only the signals with human speech that are processed to the feature extraction and recognition stages. The MFCC method is then applied as the feature extractor and the Euclidean distance is used for recognition. The performance of the proposed voice recognition system is evaluated using the FTI Untar Pattern Recognition Lab voice database. The accuracy for noise splitting process is promising with 90.91%.