Heteroscedastic Linear Discriminant Analysis (HLDA) - Derivation, Implementation and Integration into Adaptive Training

This work investigates the application of Heteroscedastic Linear Discriminant Analysis (HLDA) in a LVCSR system. After reviewing the equations that are required for defining the transformation and its numerical optimization, two different algorithm for implementation of HLDA are proposed. Each of them has different properties regarding memory requirements and computation time. Both methods are experimentally evaluated on a number of different tasks. Integration of HLDA into an adaptive training procedure has to take into account the properties of the Constrained Maximum Likelihood Linear Regression (CMLLR)transformation which is typically employed for adaptation or normalization. Thus, new algorithms are developed that allow to integrate the estimation of HLDA into an adaptive training procedure in a manner that is both memory- and time-efficient. Finally, experimental results are given that show the effectiveness of the proposed approach