Multiple linear transforms

摘要: Heteroscedastic discriminant analysis (HDA) has been proposed as a replacement for linear (LDA) in speech recognition systems that use mixtures of diagonal covariance Gaussians to model the data. Typically HDA and LDA involve dimension reduction feature space. A specific version involves no reduction; is popularly known maximum likelihood transform (MLLT) often used on space give significant improvements performance. MLLT approximately diagonalizes class covariances, effect, tries approximate performance full-covariance-system. However, full-covariance system could some cases be much better than using MLLT-based system. We propose method multiple transforms, bridges this gap performance, while maintaining speed efficiency This technique improves system, over what obtained from or MLLT.