5th International Conference on Spoken Language Processing
Speech recognition systems based on hidden Markov models (HMM) favourably apply a linear discriminant analysis transform (LDA) which yields low-dimensional and uncorrelated feature components. However, since the distributions in the HMM states usually are modeled by mixture gaussian densities, the description by second-order moments no longer is correct. For this purpose we introduced a new "extended linear discriminant analysis" transform (ELDA) which starts from conventional LDA. The ELDA transform is derived by use of a gradient descent optimization procedure based on a "minimum classification error" (MCE) principle, which is applied to the original high-dimensional pattern space. The transform matrix, the best fitting prototype of the correct class (i.e. HMM state) and the nearest rival are adapted. We developed a method which additionally updates all prototypes by a separate maximum likelihood (ML) estimation step. This avoids that such means and covariances, which mostly remain unaffected by the MCE procedure, may diverge step by step.
Bibliographic reference. Ruske, Guenther / Faltlhauser, Robert / Pfau, Thilo (1998): "Extended linear discriminant analysis (ELDA) for speech recognition", In ICSLP-1998, paper 0100.