Optimization of the number of Gaussian components for speaker identification
R. J. R. Cirigliano, F.G.V. Resende · 2006
In this paper we present an algorithm that searches the optimal number of components for each speaker based on the rate between the maximum and the average distortion in each cell obtained through the k-means algorithm. The speakers are modeled by orthogonal GMM and each of them has a different number of components. The tests showed that the usage of this algorithm decreases the error rate in average 36.2% and also decreases the computational effort of the identification process in average 48% when compared to orthogonal GMM models.