Robust Heteroscedastic Linear Discriminant Analysis and LCRC Posterior Features in Meeting Data Recognition

摘要

This paper investigates into feature extraction for meeting recognition. Three robust variants of popular HLDA transform are investigated. Influence of adding posterior features to PLP feature stream is studied. The experimental results are obtained on two data-sets: CTS (continuous telephone speech) and meeting data from NIST RT'05 evaluations. Silence-reduced HLDA and LCRC phoneme-state posterior features are found to be suitable for both recognition tasks.