Missing-feature approaches in speech recognition

Approaches that have been conspicuously effective in ameliorating the effects of transient maskers are reviewed. Focus is on two broad classes of missing feature algorithms namely, feature-vector imputation algorithms and classifier-modification algorithms. The mathematics of four major missing feature techniques are reviewed including, the feature-imputation techniques of cluster-based reconstruction and covariance-based reconstruction, and the classifier-modification methods of class-conditional imputation and marginalization. The difficult task of estimating the spectrographic masks is also outlined.