Robust speaker verification using self organizing map

This paper proposes a new approach of noise reduction based on the analysis of MFCC feature space using self-organizing map network. Here the U-matrix plot of the feature space is analyzed in presence of white noise at different signal to noise ratio. Based on the observation, boundary neurons separating clusters are identified in the feature space. For each such neuron in the boundary, its 2-D feature vector is extracted from the U-matrix and hit matrix. This collection of feature vectors based on the boundary neurons are eliminated from the original feature space. Thus the new feature space obtained is used to perform the tasks of visualization and speaker verification. Experiments were carried out by combining synthetic white noise with real world data sets.