Abstract | ||
---|---|---|
Gaussian distributions are usually parameterized with the ir nat- ural parameters: the mean µ and the covariance �. They can also be re-parameterized as exponential models with canonical parameters P = � 1 and = Pµ. In this paper we con- sider modeling acoustics with mixtures of Gaussians param- eterized with canonical parameters where the parameters ar e constrained to lie in a shared affine subspace. This class of models includes Gaussian models with various constraints on its parameters: diagonal covariances, MLLT models, and the re- cently proposed EMLLT and SPAM models. We describe how to perform maximum likelihood estimation of the subspace and parameters within a fixed subspace. In speech recognition ex - periments, we show that this model improves upon all of the above classes of models with roughly the same number of pa- rameters and with little computational overhead. In partic ular we get 30-40% relative improvement over LDA+MLLT models when using roughly the same number of parameters. |
Year | Venue | Keywords |
---|---|---|
2003 | INTERSPEECH | maximum likelihood estimate,speech recognition,mixture of gaussians,gaussian distribution |
Field | DocType | Citations |
Diagonal,Overhead (computing),Parameterized complexity,Affine space,Subspace topology,Pattern recognition,Computer science,Gaussian,Artificial intelligence,Exponential models,Covariance | Conference | 6 |
PageRank | References | Authors |
0.74 | 8 | 3 |
Name | Order | Citations | PageRank |
---|---|---|---|
Karthik Visweswariah | 1 | 400 | 38.22 |
Scott Axelrod | 2 | 113 | 10.14 |
Ramesh A. Gopinath | 3 | 323 | 42.58 |