EUROSPEECH 2003 - INTERSPEECH 2003
The Philips speech recognition system uses mixtures of Laplacian densities with diagonal deviations to model acoustic feature vectors. Such an approach neglects the correlations between different feature components that typically exist in the acoustic vectors. This paper extends the conventional Laplacian approach to model the between-feature interdependencies explicitly. These extensions either lead to a full deviation matrix model or to an integrated feature space transformation similar to the semi-tied covariances for Gaussian densities. Both methods can be efficiently implemented by exploiting a strong tying of the feature transformations and the deviation matrices, respectively. The novel approach is evaluated on two different digit string recognition tasks.
Bibliographic reference. Neukirchen, Christoph (2003): "Semi-tied full deviation matrices for laplacian density models", In EUROSPEECH-2003, 2609-2612.