Modeling inverse covariance matrices by basis expansion

摘要: This paper proposes a new covariance modeling technique for Gaussian Mixture Models. Specifically the inverse (precision) matrix of each is expanded in rank-1 basis i.e., Σ j −1 = P k 1 D λ ja T, ∈ ℝd. A generalized EM algorithm proposed to obtain maximum likelihood parameter estimates set {a T} and expansion coefficients {λ j}. model, called Extended Maximum Likelihood Linear Transform (EMLLT) extremely flexible: by varying number elements from d d(d + 1)/2 one gradually moves (MLLT) model full-covariance model. Experimental results on two speech recognition tasks show that EMLLT can give relative gains up 35% word error rate over standard diagonal