Penalized maximum-likelihood estimation, the Baum-Welch algorithm, diagonal balancing of symmetric matrices and applications to training acoustic data (Q1576464)

The authors are concerned with nonparametric density estimation of high-dimensional data. The work is driven by its potential application to train speech data where traditionally only parametric methods have been used. Parametric models typically lead to large-scale optimization problems associated with a desire to maximize the likelihood of the data. In particular, mixture models of Gaussians are used for training acoustic vectors for speech recognition and the parameters of the models are obtained using \(K\)-means clustering and the EM algorithm. The authors consider the possibility of maximizing the penalized likelihood of the data as a means to identify nonparametric density estimators. Various mathematical properties of this point of view are developed, several algorithms for the numeric solution of the optimization problems encountered are proposed and the experience with these methods is reported.