Abstract—The ever-increasing size of datasets in speaker recognition systems is the primary reason why challenges arise with regard to accuracy and computational complexity. Inadequate speaker-specific information on vocal tracts may lead to poor modeling and can adversely affect the performance under large-scale data conditions. In this work, we have developed a speaker recognition system, based on the excitation source information by blending pitch and pitch strength vectors. We investigate various approaches to improve the performance from two directions. First, we investigate various dimensionality reduction techniques during the feature extraction phase such as Multi-linear Principle Component Analysis (MPCA), Principle Factor Analysis (PFA), and Maximum Likelihood Factor Analysis (MLFA). We have evaluated the performance of different large-scale Support Vector Machine (SVM) algorithms as a function of training time to attain convergence. Combinations of the proposed dimensionality reduction methods and SVM algorithms have successfully produced an effective recognition system. We have demonstrated the performances of our approaches by conducting experiments on standard large-scale data bases, and then, compared these with the existing state-of-the-art recognition systems. The experimental results have shown that these approaches significantly improve the performance under large-scale data conditions in comparison with the conventional procedures.
 

Index Terms—maximum likelihood factor analysis, large scale-SVM, pitched, dimensionality reduction, support vector machine

Cite: P. R. K. Rao and Y. S. Rao, "Dimensionality Reduction Techniques and SVM Algorithms for Large Population Speaker Identification," International Journal of Signal Processing Systems, Vol. 4, No. 2, pp. 86-96, April 2016. doi: 10.12720/ijsps.4.2.86-96