Using information theoretic vector quantization for GMM based speaker verification

The introduction of Gaussian mixture models in the field of voice recognition systems has established very good results. The process of speaker verification based on Gaussian mixture models is highly expensive in the regard of computational complexity and memory usage perspectives, thus suppressing its adaptability for efficient and low-cost systems. The methods like Expectation Maximization used by GMM to compute the speaker models are highly iterative procedures and contribute significantly to the complexity in the implementation of an efficient system. In this paper we propose the use of Information theoretic vector quantization VQIT for the training of GMM models as a replacement of EM algorithm; we also apply the other vector quantization techniques such as K-means and LBG and compare the performance with the VQIT.